page_content
stringlengths 1
1.5k
| metadata
dict |
|---|---|
Optimal trap cropping investments to maximize agricultural yield Matthew H Holden School of Mathematics and Physics, The University of Queensland, St Lucia, 4072, Australia Centre for Biodiversity and Conservation Science, The University of Queensland, St Lucia, 4072, Australia Abstract Trap cropping is a pest management strategy where a grower plants an attractive “trap crop” alongside the primary crop to divert pests away from it. We propose a simple framework for optimizing the proportion of a grower’s field or greenhouse allocated to a main crop and a trap crop to maximize agricultural yield. We implement this framework using a model of pest movement governed by trap crop attractiveness, the potential yield threatened by pests, and functional relationships between yield loss and pest density drawn from the literature. Focusing on a simple case in which pests move freely across the field and are attracted to traps solely by their relative attractiveness, we find that allocating 5–20 percent of the landscape to trap plants is typically required to maximize yield and achieve effective pest control in the absence of pesticides. For highly attractive trap plants, growers can devote less space because they are more effective; less attractive plants are ineffective even in large numbers. Intermediate attractiveness warrants the greatest investment in trap cropping. Our framework offers a transparent and tractable approach for exploring trade-offs in pest management and can be
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
Intermediate attractiveness warrants the greatest investment in trap cropping. Our framework offers a transparent and tractable approach for exploring trade-offs in pest management and can be extended to incorporate more complex pest behaviors, crop spatial configurations, and economic considerations. Keywords: ecological pest management, integrated pest management, inter-cropping, companion planting, organic agriculture, trap crop Recommendations for Resource Managers • Planting a highly attractive “trap crop” that lures pests away from the main crop in an agricultural field can control pests, improve yield, and lower reliance on chemical pesticides. • Because trap plants occupy precious space that could otherwise be used to grow harvestable crops, it is important to optimize their share of the field. • Our results suggest that allocating 5–20% of the field to trap crops may be required to maximize yield, depending on pest pressure and trap plant attractiveness. 1 Introduction Pests remain one of the most significant threats to agricultural productivity worldwide, reducing yields and increasing the need for chemical interventions . However, widespread pesticide use has led to well-documented issues, including the evolution of pest resistance , harm to non-target species and pollinators , and environmental contamination . These concerns have motivated a growing interest in ecologically based pest management approaches that reduce reliance on chemical controls while
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
and pollinators , and environmental contamination . These concerns have motivated a growing interest in ecologically based pest management approaches that reduce reliance on chemical controls while maintaining productivity . Among these, trap cropping has emerged as a promising, environmentally friendly tactic for manipulating pest behavior to protect the main crop from pests . 1 arXiv:2508.05896v1 [q-bio.PE] 7 Aug 2025 Trap cropping is a strategy where a more attractive plant species or variety is deliberately planted to draw pests away from the main crop, thereby reducing damage and improving overall yield . The effectiveness of trap cropping has been demonstrated in a variety of agricultural systems, including cotton, brassicas, cucurbits, and legumes . However, despite its widespread use, practical guidance on how much land to allocate to the trap crop is lacking. Many studies remain empirical, focusing on evaluating the attractiveness of candidate trap plants for specific crop–pest pairs or small-scale field trials, with limited generalization to broader agroecological conditions . Moreover, quantitative theory linking system parameters to optimal spatial allocation remains underdeveloped. Mathematical and simulation models have played a critical role in advancing our understanding of trap cropping and related ecological interventions such as intercropping and companion planting. Early models focused on insect pest dispersal and behavioral responses to attractants or
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
of trap cropping and related ecological interventions such as intercropping and companion planting. Early models focused on insect pest dispersal and behavioral responses to attractants or traps embedded within cropping systems . Since then, a range of modeling approaches have been used to explore how pest movement, crop attractiveness, and spatial layout interact to influence trap crop performance . Simulation-based studies have shown that trap effectiveness depends not only on the strength of attraction but also on spatial configuration and pest mobility. For example, models incorporating biological control agents alongside trap crops reveal strong context dependence, with outcomes shaped by both natural enemy dynamics and landscape structure . Individual-based models (IBMs) have been used to account for fine-scale behavioral heterogeneity and movement biases in response to trap placement , while recent work has examined how plant dispersion and patch shape influence trapping efficiency . A broader synthesis of movement modeling in cropping systems underscores the importance of aligning spatial designs with pest dispersal mechanisms and highlights trade-offs between model tractability and ecological realism . While these studies provide rich insight into pest–crop dynamics and demonstrate the potential of spatially targeted interventions, they do not offer general analytic solutions for the optimal proportion of land to devote to trap cropping. Our work aims to fill this
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
demonstrate the potential of spatially targeted interventions, they do not offer general analytic solutions for the optimal proportion of land to devote to trap cropping. Our work aims to fill this gap by proposing a stylized but tractable framework that captures key trade-offs and yields explicit recommendations for growers. We start by developing a simple mathematical framework to determine the optimal proportion of a grower’s field that should be allocated to a trap crop to maximize total yield. Our framework is intentionally stylized and analytically tractable. Analytic solutions are especially valuable because they (i) improve transparency by making a direct link between assumptions and outcomes easy to interpret and interrogate, (ii) their solutions serve as baselines for validating more complex future simulation models, and (iii) they allow for general insights to be computed and communicated efficiently. The model is built around two central parameters: the proportion of yield that would be lost to pests in the absence of trap cropping and the relative attractiveness of the trap crop, which determines the likelihood that pests are diverted away from the main crop. Both parameters can be estimated from experimental data, expert elicitation, or synthesis studies on pest behavior and damage functions . To illustrate the utility of the framework, we analyze a simplified scenario in which pests move freely across the field and select plants purely based on relative
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
behavior and damage functions . To illustrate the utility of the framework, we analyze a simplified scenario in which pests move freely across the field and select plants purely based on relative attractiveness. This assumption corresponds to highly mobile pests, such as adult moths that lay their eggs in batches over several days or weeks across multiple plants . The assumption of high mobility relative to crop area is especially plausible in smaller greenhouse environments, since pests can rapidly traverse the entire planting area . Under these conditions, we derive explicit expressions for expected yield and characterize the optimal fraction of land to devote to trap cropping. Using plausible parameter values from the literature, we find that between 5–20% of the field must typically be allocated to trap crops for effective pest control in the absence of pesticides. Interestingly, we show that the relationship between trap crop attractiveness and optimal area is non-monotonic: highly attractive traps require relatively little area; traps with low attractiveness are ineffective even in large numbers; and intermediate levels of attractiveness require the largest investment. Our framework contributes to the growing body of theory guiding sustainable pest management and can be extended to incorporate more complex dynamics such as pest reproduction, spatial heterogeneity, and economic costs. 2 2 Framework In this work, we consider a scenario with N pests in a field damaging
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
extended to incorporate more complex dynamics such as pest reproduction, spatial heterogeneity, and economic costs. 2 2 Framework In this work, we consider a scenario with N pests in a field damaging the main crop harvested by the grower. To reduce this damage, the grower may plant a trap crop to divert pests, potentially increasing main crop yield. A fundamental question in implementing trap cropping is how much land to allocate to trap plants and, consequently, how much to assign to the main crop. Trap plants are typically not harvested and therefore do not directly contribute to yield. In fact, they take up space that would otherwise be used for planting productive main crop plants. However, trap crops can indirectly increase the grower’s yield by reducing pest density on the main crop. This creates a tradeoff: growers are reluctant to allocate space to non-producing plants, but desire lower pest densities. Assume that the yield of a single main crop plant depends on the pest density on that plant. In other words, a plant’s yield is y(ρ), where ρ is the number of pests on that plant. Let n be the total number of plants that can fit in the field. Let nc be the number of cash (main crop) plants; then n −nc is the number of trap plants. Since trap plants are not harvested, and assuming all cash and trap plants are otherwise equal, the total yield across the whole field is given by: Y (nc) = y (ρ(nc)) · nc (1) This equation says that total yield is just the average yield of a
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
all cash and trap plants are otherwise equal, the total yield across the whole field is given by: Y (nc) = y (ρ(nc)) · nc (1) This equation says that total yield is just the average yield of a cash plant, given the average pest density on cash plants, multiplied by the number of cash plants. As the grower increases the number of cash plants, the multiplication by nc on the right has an increasing effect, but y(ρ(nc)) is more complicated. If trap plants are effective at reducing pest densities on cash plants, then this term, yield per plant, can decrease when the grower adds cash plants through increases in pest density, ρ. However, note that y is not guaranteed to decrease with nc (even though it decreases in ρ) because if trap plants are not effective at trapping or retaining pests, then pests are diluted across an increased number of cash plants. With N pests in the field, the average pest density is ρ(nc) = x(nc)N nc , (2) where x(nc) is the proportion of insects on the cash crop as a function of the number of cash crops. To maximize total yield, one can differentiate equation (1), which results in general insight into the optimal solution. It says the optimal number of cash plants should satisfy, y(ρ(nc)) = −dy dnc · nc, (3) provided some conditions on the functions in equations (1) and (2) hold (see section 7). Equation (3) implies that at the optimal number of cash plants, the yield per plant equals the marginal loss in yield caused by adding another cash plant, summed
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
(1) and (2) hold (see section 7). Equation (3) implies that at the optimal number of cash plants, the yield per plant equals the marginal loss in yield caused by adding another cash plant, summed over all plants. Roughly, if you added a cash plant, the yield of that cash plant should equal the reduced yield summed over all plants due to the increased pest density. If the yield of adding another cash plant is greater than the yield loss it would cause due to pests, it would be optimal to increase the number of cash plants in the landscape. Conversely, if the yield from that plant is not greater than the loss it causes, it is optimal to add more trap plants (i.e., decrease the number of cash plants in the landscape). While flexible, the framework requires specifying two key relationships: y(ρ), and x(nc). In the following sections, we demonstrate how empirically supported functions from the literature can be used in this framework to achieve practical recommendations for growers deciding how many trap plants to deploy in their agricultural field or greenhouse. 3 Yield–Pest relationships The relationship between yield and pest density has been studied both theoretically and empirically in the literature . Many candidate relationships have been proposed. However, they all share some common properties. First, there is some assumed maximum yield per plant in the absence 3 0 2 4 6 8 10 0 20 40 60 80 100 Pest density per plant Yield per plant Linear Reciprocal Exponential Figure 1:
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
common properties. First, there is some assumed maximum yield per plant in the absence 3 0 2 4 6 8 10 0 20 40 60 80 100 Pest density per plant Yield per plant Linear Reciprocal Exponential Figure 1: Yield curves, y(ρ), as a function of pest density per plant ρ, with N = 1,000, β = 0.8, k = 100, and n = 10. of pests, call this maximum plant yield k, in other words, y(0) = k. Yield is also assumed to decrease monotonically with pest density, dy/dρ < 0 for all ρ > 0. Common functional forms for y(ρ) include linear declines, exponential and reciprocal curves (which drop sharply at low densities and saturate at high densities), and S-shaped curves, where yield remains high at low pest densities, declines rapidly around a threshold, and then saturates again . For our application, we also consider the maximum proportion of yield lost in the absence of trap plants. That is, we parameterize the yield functions such that y(N/n) = (1−β)k, where β is the proportion of yield lost per plant due to N total pests in a field of n plants in the absence of pest management. For example, if β = 1, N pests will wipe out the entire grower’s yield in the absence of trap crops, whereas if β = 0.5, half of the grower’s yield is lost. Linear yield loss If we assume the yield of a single plant decreases linearly with pest density from a maximum yield of k to (1 −β)k when pest density is N/n, we obtain the yield function, y(ρ) = k 1 −βn N ρ . (4) Reciprocal yield Assuming yield declines reciprocally
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
with pest density from a maximum yield of k to (1 −β)k when pest density is N/n, we obtain the yield function, y(ρ) = k 1 −βn N ρ . (4) Reciprocal yield Assuming yield declines reciprocally with pest density, i.e., y(ρ) = α/(1 + γρ), and parameterizing such that y(0) = k and y(N/n) = (1 −β)k, we obtain: y(ρ) = k(1 −β)N (1 −β)N + βnρ. (5) Exponential yield Similarly, exponential yield, parameterized in the same way, can be expressed as y(ρ) = k(1 −β)nρ/N. (6) 4 4 Pest distribution and movement For simplicity, we assume the pest population is held constant. However, we still require a method to allocate pests between trap and cash plants, since yield depends on both the number of cash plants and the pest load they experience. We describe the pest distribution using a single variable xt, the proportion of the pest population on the main crop (cash plants) at time t. Therefore, 1−xt is the proportion of the pest population on the trap crop (trap plants) at time t. Assume the movement dynamics of the pests are described by a simple difference equation. xt+1 = f(xt), (7) If f has a stable equilibrium that is reached quickly, this equilibrium can be used in equation (2). Below, we describe a simple example model first proposed by Holden et al. and then use it to compute the optimal allocation of land to cash and trap plants. Consider the case where pests leave plants at each time step with probability l. Pests that leave a cash plant settle on another cash plant with
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
optimal allocation of land to cash and trap plants. Consider the case where pests leave plants at each time step with probability l. Pests that leave a cash plant settle on another cash plant with probability s, while those leaving a trap plant settle on a cash plant with probability σ. Under these assumptions, the proportion of the pest population on the main crop changes through time iteratively according to the difference equation (7) with, f(x) = (1 −l)x + lsx + lσ(1 −x). (8) The first term accounts for pests that remain on the main crop, the second for those that left and returned, and the third for pests that moved from the trap crop to the main crop. The equilibrium of this model is x∗= σ 1 + σ −s, (9) which is approximately the proportion of the pest population on the main crop in the long run. Because equations (7) and (8) define a linear difference equation, the equilibrium is globally asymptotically stable. Therefore, the equilibrium is globally asymptotically stable. There is even an analytic expression for xt, and it can be used to show that solutions, no matter the initial conditions, approach the equilibrium exponentially quickly. Therefore, it suffices to look at equilibrium pest density as a proxy for cumulative pest load as long as crop damage is consistently related to pest density. So far, our model is simple. We have assumed very little about the spatial structure of the field. The only critical assumption is that all trap plants are treated as
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
related to pest density. So far, our model is simple. We have assumed very little about the spatial structure of the field. The only critical assumption is that all trap plants are treated as identical, and all cash plants are treated identically as well. This assumption implies that pests leaving any trap plant have the same probability of settling on a cash plant, regardless of their starting location. The same applies to pests leaving cash plants. However, this assumption is likely to hold approximately in many systems. If trap plants are planted in rows, for example, from the eyes of a pest on a trap plant, locally, the world looks the same. Note we have yet to introduce any specific model for movement. A specific movement model is needed to determine the settlement probabilities s and σ. To specify s and σ, we assume a common pool dispersal model, in which pests can access all plants each time step. Let nT be the number of trap plants in the field, and nC be the number of cash plants. Also, assume a trap plant is a times more attractive than a single cash plant. Since pests view all plants at each time step, their prior location does not affect their settlement decision, and therefore, s = nC anT + nC = σ. (10) In this movement model, the equilibrium in equation (9) simplifies to x∗= s, providing a simple formula relating the proportion of pests on the cash crop directly to the number of cash crops, nC. Substituting our expression for x∗for x(nC) then allows us to
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
to x∗= s, providing a simple formula relating the proportion of pests on the cash crop directly to the number of cash crops, nC. Substituting our expression for x∗for x(nC) then allows us to proceed using the framework to determine the optimal number of cash plants (and hence trap plants) to allocate to the landscape. It is important to note that in such a simplified, stylized model, we are assuming that pests leave trap plants and cash plants with the same probability, l, and trap plants only accumulate more pests than cash plants via their superior attraction once an insect has initiated movement. There is some support for this in the literature for whiteflies as pests of crops in greenhouses . 5 5 Optimal number of cash/trap plants to maximize grower yield Given the movement model described in equations (7) – (10) with the simplest yield–pest density relationship, linear yield loss as pest density increases, the optimal number of cash plants in the field to maximize total yield can be derived analytically as n∗ C = a −√βa a −1 n. (11) To understand this expression, let us first consider the case where β = 1. Recall that β = 1 means the total number of pests threatening the farm is so great that yield would be zero in the absence of trap crops. In such a case, the optimal number of trap plants, n −n∗ C, makes up a high proportion of the landscape. For example, if a trap plant were four times as attractive as a cash plant, the grower would have to devote a third of the
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
number of trap plants, n −n∗ C, makes up a high proportion of the landscape. For example, if a trap plant were four times as attractive as a cash plant, the grower would have to devote a third of the landscape to trap plants to maximize yield. Even if the trap crop is 100 times more attractive than the cash crop, the grower would still need to dedicate 9/99 proportion of the landscape to trap plants (nearly 10 percent). These numbers are quite large as they mean the grower is giving up 10% of their field to plants that aren’t producing anything they can harvest. As we decrease β, this conclusion becomes less severe, but still a substantial proportion of the landscape needs to be devoted to trap plants. For example, even if 40 percent of the grower’s yield is threatened by pests (a typical yield loss due to higher pest densities in organic agriculture), and the trap plant is four times as attractive as the cash plant, then nine percent of the landscape should be devoted to trap plants. If the trap plant is 100 times more attractive, then the optimal allocation of the landscape to trap plants is five percent. Under nonlinear yield loss functions, such as exponential and reciprocal yield loss, the effects are similar. For reciprocal yield, the optimal number of cash plants is n∗ C = " a(1 −β) + β − p β2 + βa(1 −β) (1 −β)(a −1) # n, (12) if the quantity is in [0, n], or n otherwise. For exponential yield, the optimal number of cash plants is n∗ C = " 2a −ln(1 −β) − p ln(1
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
n∗ C = " a(1 −β) + β − p β2 + βa(1 −β) (1 −β)(a −1) # n, (12) if the quantity is in [0, n], or n otherwise. For exponential yield, the optimal number of cash plants is n∗ C = " 2a −ln(1 −β) − p ln(1 −β)(ln(1 −β) −4a) 2(a −1) # n. (13) These expressions are more complex than for the linear yield relationship. However, note a key commonality across all three expressions: the optimal number of cash plants in equations (11) – (13) is directly proportional to n. This means the bracketed term in each equation can be interpreted as the optimal proportion of the landscape allocated to the cash plant. To visualize how these expressions differ, in the next section, we compare them for all possible parameter combinations. 6 Illustrative results We start by parameterizing the model using reported quantities available in the literature to form a plausible baseline. For example, it has been shown in choice experiments that 98% of greenhouse whiteflies, Trialeurodes vaporariorum, a common agricultural pest, choose to settle on an eggplant trap plant versus a poinsettia cash plant . This implies that eggplant is 49 times more attractive than poinsettia when the pest has decided to move. It is commonly suggested that without the use of pesticides, a roughly 40 percent reduction in yield is typical in agricultural landscapes . We will use these two parameter values as an illustrative baseline, but will also explore the whole parameter space. If a = 49 and β = 0.4, we find that roughly 91.5 –
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
in agricultural landscapes . We will use these two parameter values as an illustrative baseline, but will also explore the whole parameter space. If a = 49 and β = 0.4, we find that roughly 91.5 – 92.9 percent of a grower’s greenhouse or field should be allocated to cash plants to maximise yield (Figure 2). The nonlinear yield–pest relationships lead to more space being required for trap cropping: 8.5% for reciprocal yield, equation (12), and 7.8% for exponential yield, equation (13), compared to the case when yield declines linearly with pest density (7.1% trap plants). 6 This occurs because the yield loss per pest is steeper at low pest densities under these relationships, so even small increases in pest density across many cash plants have a large negative impact on total yield. 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 Cash plant proportion Proportion of a pest−free landscape's yield Linear Reciprocal Exponential Figure 2: Proportion of the total pest-free field’s yield (i.e., a field planted only with cash plants) achieved in a system with pests threatening up to 40% of the yield (β = 0.4), shown as a function of the proportion of the field devoted to trap plants. Curves correspond to a linear (blue solid), reciprocal (purple dotted), and exponential (red dashed) yield–pest relationship, assuming the trap plant is 49 times more attractive than the cash plant (a = 49). Note that when there are no trap plants, the grower loses 40% of their yield (right endpoint), with
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
relationship, assuming the trap plant is 49 times more attractive than the cash plant (a = 49). Note that when there are no trap plants, the grower loses 40% of their yield (right endpoint), with steep gains from adding only a few trap plants (moving left). The optimal proportion of the field to devote to trap plants is 7-8.5%, depending on the yield-pest relationships. Specifically, the optimal cash plant (trap plant) proportion is 92.9% (7.1%), 92.2% (7.8%), and 91.5% (8.5%) for the linear, exponential, and reciprocal yield–pest density relationships, respectively. In the case where there are all trap plants and no cash plants, every additional cash plant adds approximately an additional yield of y(ρ). In other words, in a landscape of no cash plants, adding a cash plant generates a pest-free plant’s worth of yield, because all the pests are on the many trap plants. This explains the linear increasing relationship between total yield and the proportion of the landscape with cash plants 7 for low cash plant proportions (see bottom left of Figure 2). Once the proportion of cash plants begins to exceed 80 percent, then the incremental yield of an additional cash plant starts to be reduced by the associated increased pest density across all cash plants in the system, with an optimal cash plant proportion of 91.5–92.9 percent. At approximately 95 percent of the landscape devoted to cash plants, each additional cash plant severely reduces yield compared to an equivalent
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
an optimal cash plant proportion of 91.5–92.9 percent. At approximately 95 percent of the landscape devoted to cash plants, each additional cash plant severely reduces yield compared to an equivalent investment in trap plants due to pest damage (see the rapid drop off in the right of Figure 2). However, a = 49 represents an extremely attractive plant. The question remains: What happens when the trap plant is less attractive? In Figure 3a, we see that for a = 2, and β = 0.4, trap cropping is entirely ineffective at increasing yield. This is because in such cases, the pest is only weakly drawn towards trap plants, and therefore, the reduction in pest pressure on cash plants is minimal; the lost yield from sacrificing land to unharvestable trap crops outweighs the small gains from pest control. 8 0.4 0.5 0.6 0.7 0.8 0.9 Linear Reciprocal Exponential a) a=2 0.4 0.5 0.6 0.7 0.8 0.9 b) a=5 0.5 0.6 0.7 0.8 0.9 1.0 0.4 0.5 0.6 0.7 0.8 0.9 c) a=25 Cash plant proportion Proportion of a pest−free, 100% cash plant, landscape's yield Figure 3: Grower yield as a function of the proportion of cash plants for trap cropping systems with less attractive trap plants than the baseline, where trap plants are (a) twice, (b) five times, and (c) 25 times more attractive than the cash plant. For a trap plant that is only twice as attractive as a cash plant, trap cropping is ineffective, and the grower should plant only cash plants, accepting the yield loss caused by pests. For more attractive trap
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
plant that is only twice as attractive as a cash plant, trap cropping is ineffective, and the grower should plant only cash plants, accepting the yield loss caused by pests. For more attractive trap plants, 5 and 25 times more attractive, approximately 7–10% of the field should be allocated to trap plants. As attractiveness is increased further to five, meaning a pest is five times more likely to settle on a trap plant than a cash plant, it is optimal to devote roughly ten percent of the land to the trap crop (Figure 3b). However, differences in prevented yield loss are less sensitive to the number of trap plants than for more attractive trap plants (compare Figure 3b to Figure 2 and Figure 3c). Interestingly, from figures 2 and 3, the optimal number of trap plants is bigger for intermediate attractiveness values of 5 and 25 than for 49. This is because intermediate attractiveness is sufficient to divert pests but not highly efficient, so more area is needed to sufficiently reduce pest pressure and maximize yield. This suggests, counterintuitively, that growers using intermediately attractive trap plants may need to invest 9 more land in trap cropping than those using highly attractive varieties. The figures also show that the nonlinear yield-pest relationships (red and purple) can increase the optimal number of trap crops or decrease it compared to the linear pest-density model, depending on the attractiveness parameter (the order of the colored circles, representing
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
can increase the optimal number of trap crops or decrease it compared to the linear pest-density model, depending on the attractiveness parameter (the order of the colored circles, representing optimal yield, changes when comparing Figure 3b and to Figure 3c). These reversals reflect that, at intermediate attractiveness levels, nonlinear damage responses can either amplify or buffer the effects of pest distribution. This depends on which part of the damage curve is most influential. To investigate this further, we plotted the optimal proportion of cash plants in equations (11) – (13) versus attractiveness, for the baseline yield at risk of 40 percent (Figure 4a). When trap plants are between one and 2.6 times more attractive than the main crop, the whole field should be cash plants, as trap cropping is not effective enough to make up for lost yield. From an attractiveness of 3.6 to eight, the optimal proportion of cash plants declines rapidly to a minimum of 88 percent of the landscape under all yield-pest density relationships. However, the decline is slightly more rapid for the linear case, where the 88 percent minimum occurs at an attractiveness of eight, compared to the exponential and reciprocal cases, where the minimum is achieved at higher attractiveness values of 10 and 12.5, respectively. After achieving the minimum, for larger attractiveness values, the optimal proportion of the landscape devoted to cash plants gradually rises again. This is because, as trap crops
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
12.5, respectively. After achieving the minimum, for larger attractiveness values, the optimal proportion of the landscape devoted to cash plants gradually rises again. This is because, as trap crops become highly effective, less land allocation is required to achieve adequate pest control. For even nearly perfect attractiveness, e.g., 100 times more attractive than the cash crop, the grower still needs to devote five, six, and seven percent of the landscape to the trap crop, for the linear, exponential, and reciprocal relationships, respectively. 0 20 40 60 80 100 0.6 0.7 0.8 0.9 1.0 a) Trap plant attractiveness Linear Reciprocal Exponential 0.0 0.2 0.4 0.6 0.8 1.0 b) Proportion of yield at risk Optimal cash plant proportion Figure 4: The optimal proportion of the landscape to allocate to the main crop (cash plants) as a function of (a) trap plant attractiveness and (b) the proportion of yield at risk from pests. For intermediately attractive trap plants (5–20 times as attractive as the cash plant), more than 10% of the landscape should be sacrificed to trap plants, as they provide sufficient efficacy to achieve pest reduction, but are not effective enough to work at low densities. For low attraction (a < 4), trap cropping is ineffective. For highly attractive trap plants, their efficacy is high enough that a smaller proportion of the landscape can achieve the necessary reduction in pest densities to maximize yield. Optimal trap cropping is even more sensitive to the
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
their efficacy is high enough that a smaller proportion of the landscape can achieve the necessary reduction in pest densities to maximize yield. Optimal trap cropping is even more sensitive to the maximum yield at risk due to pests, β, than the attraction parameter. We plotted the optimal proportion of cash plants in equations (11) equation (13) versus yield at risk, β, for the baseline trap plant attractiveness of 49 percent (Figure 4b). In general, the optimal proportion of the landscape dedicated to the cash crop declines monotonically as there is more yield 10 at risk due to high pest densities (see decreasing trend in Figure 4b). This is particularly more severe for the more nonlinear pest yield relationships. For the reciprocal yield relationship, this is particularly severe. If 100% of the yield is at risk, the cash crop should make up roughly 60% of the landscape to maximize yield. For high, but more typical, yield at risk, such as 60 and 80 percent, optimal cash crop allocations correspond to 87 and 80 percent, respectively. For linear and exponential relationships, the trend is similar but less severe. The relationship is strongest for the reciprocal yield curve because even low pest density causes a steep drop-off in yield. Therefore, more trap plants are required to achieve optimal yield. The exponential is an intermediate case between this severely nonlinear yield–pest relationship and the more gradual, linear, pest–yield relationship. However, even in the
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
to achieve optimal yield. The exponential is an intermediate case between this severely nonlinear yield–pest relationship and the more gradual, linear, pest–yield relationship. However, even in the linear case, if 100 percent of the yield is at risk, still 88 percent of the field needs to be devoted to cash plants to maximize yield. To see the differences between the yield achieved at all cash crop proportions for different yields at risk, see Figure 5. 0.2 0.4 0.6 0.8 1.0 Linear Reciprocal Exponential a) β=0.2 0.2 0.4 0.6 0.8 1.0 b) β=0.6 0.5 0.6 0.7 0.8 0.9 1.0 0.2 0.4 0.6 0.8 1.0 c) β=0.8 Cash plant proportion Proportion of a pest−free landscape's yield Figure 5: Grower yield as a function of the proportion of cash plants for trap cropping systems with different levels of yield at risk compared to the baseline in Figure 2: when (a) 10%, (b) 30%, and (c) 60% of maximum yield is at risk due to damage from pests. 11 In the previous plots, we have fixed one parameter to be at the baseline value while varying the other. A complete sensitivity analysis of the optimal cash crop allocation across all possible combinations of attractiveness and yields at risk is displayed as a heatmap in Figure 6. Across all yield-pest density relationships, yields at risk, and attractiveness, the vast majority of the parameter space corresponds to the case when 8095 percent of the landscape should be dedicated to the cash crop to maximize agricultural yield. Low optimal cash plant proportions of
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
majority of the parameter space corresponds to the case when 8095 percent of the landscape should be dedicated to the cash crop to maximize agricultural yield. Low optimal cash plant proportions of less than five percent only occur when less than 30 percent of the grower’s yield is at risk (left second darkest region). Trap cropping should be avoided only when trap plants are both unattractive and low yields are at risk (bottom-left dark regions in all panels) 0.0 0.2 0.4 0.6 0.8 20 40 60 80 100 Linear 0.0 0.2 0.4 0.6 0.8 20 40 60 80 100 Exponential 0.0 0.2 0.4 0.6 0.8 20 40 60 80 100 Reciprocal 0.5 0.75 0.8 0.85 0.95 0.95 1 Optimal proportion (nc * n) Attractiveness Proportion of yield at risk, β Figure 6: Heatmap of the optimal proportion of the landscape to allocate to a cash crop across a grid of yields-at-risk, β, and trap attractiveness, a. Darker regions represent greater investment in cash crops (low trap crop area). The white and lightest regions indicate 25–50% of the landscape is allocated to cash crops, implying substantial trap cropping. In most of the parameter space, optimal trap cropping investments require 5–20% of land area, as the white and next lightest regions, along with the darkest region, are small. The figure also demonstrates the result that intermediate attraction warrants the greatest trap cropping investment, compared to highly attractive and unattractive trap crops, is general across parameterizations. That is for all yield-pest dentistry
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
attraction warrants the greatest trap cropping investment, compared to highly attractive and unattractive trap crops, is general across parameterizations. That is for all yield-pest dentistry relationships (all panels) and parameter values for yields at risk (horizontal axis), as long as the trap crop is attractive enough to warrant some investment, the cash plant allocation that maximizes yield is highest for trap plants that are only a few times more attractive than the cash plant. To see this, examine vertical transects through each figure. Note there is a dark region on each side of the transect and a lighter region for intermediate attraction values for a wide range of yields at risk. 7 Generalities: existence of a single optimal cash plant proportion In sections 3 – 6, we focused on a specific model of pest movement and crop yield to demonstrate the framework and improve intuition. This required specifying several functional relationships. The question remains whether such results are possible across a wide range of functional forms and models. Below, we present general conditions on these functions, such that they guarantee a unique optimal number of cash plants (a sufficient condition). Theorem 1. There is a unique optimal number of cash plants, n∗ C ∈(0, n], satisfying equation (3), that maximizes total yield, Y (nC), if the following three properties hold: 1. The function x is positive, continuously differentiable, and strictly increasing on (0, n). 2. The function
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
(3), that maximizes total yield, Y (nC), if the following three properties hold: 1. The function x is positive, continuously differentiable, and strictly increasing on (0, n). 2. The function E, defined by E(nC) = nC x(nC) dx dnC −1, (14) is positive and strictly increasing with respect to nC on nC ∈(0, n). 12 3. The function y is continuously differentiable, and y′/y is monotonic in ρ. Below, we provide a biological explanation of the theorem. However, see the appendix for a rigorous proof. The intuition behind condition 1 is simple: As you increase the number of cash plants (and therefore decrease the number of trap plants), it requires that the density of pests per plant increases. In other words, trap plants have to reduce pest density on the cash plants as you add more trap plants to the landscape. This is almost surely satisfied for any trap cropping system; otherwise, the trap plant is ineffective and would not be considered by the grower. Note that equation (14) in condition 2 is just the elasticity of pest density on a single cash plant with respect to the number of cash plants. This is because nC ρ dρ dnC = nC x(nC) dx dnC −1 := E(nC). (15) Elasticity, in economics, is the proportional rate of change of one quantity with respect to a proportional change in the other quantity. For example, if the elasticity in equation (15) is two, then, approximately, a one percent increase in cash plants will cause a two percent increase in the proportion of pests on a cash plant.
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
For example, if the elasticity in equation (15) is two, then, approximately, a one percent increase in cash plants will cause a two percent increase in the proportion of pests on a cash plant. Therefore, condition 2 says that the elasticity of pest density on a single plant with respect to the number of cash plants present should be an increasing function. To provide intuition for condition 2, remember that total yield across the field is the product of yield per cash plant times the number of cash plants, as seen in equation (1). Consider the case where there are only 100 cash plants. In this case, a one percent boost in cash plants only adds a single plant, increasing yield by the yield from only that one plant, roughly y(ρ). Now, in the case where there are 1,000 cash plants, increasing the number of plants by one percent increases yield by roughly ten times y(ρ). Because yield scales with the number of cash plants, any yield loss per plant resulting from additional cash plants must be steep enough to offset the benefits of increasing plant numbers Since we have already assumed that the yield of a plant decreases as you increase the density of the pests (see section 3), condition 3 of the theorem just restricts the behavior of this decrease. It requires that either every additional individual pest added to a plant decreases yield more than the one before, or each pest decreases yield less than the one before, but it never switches back and forth depending on how many
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
additional individual pest added to a plant decreases yield more than the one before, or each pest decreases yield less than the one before, but it never switches back and forth depending on how many pests are on the plant. Theorem 1 means that for many mathematical models of pest movement and yield-pest relationships, one can set the derivative of Y to zero and solve for the unique optimal number of cash and trap plants, either analytically or numerically, without needing any global optimization algorithms. It is easy to show that the model of pest movement in equations (7) – (9) and all proposed yield-pest density relationships in equations (4) – (6) satisfy the conditions in Theorem 1. However, a yield-pest density curve with an inflection point will fail to satisfy condition 3, and therefore, multiple local optima will need to be checked in such a model. 8 Discussion Our results show that trap cropping can substantially improve agricultural yield, but only under specific conditions. Using a simple yet flexible framework, we derived analytic expressions for the optimal number of cash and trap plants, demonstrating how optimal strategies depend critically on two parameters: the proportion of yield at risk due to pests (β), and the relative attractiveness of the trap crop compared to the cash crop (a). When trap crops are only marginally more attractive than the cash crop (e.g., a = 2), trap cropping fails to improve yield, even under extreme pest pressure. In contrast,
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
to the cash crop (a). When trap crops are only marginally more attractive than the cash crop (e.g., a = 2), trap cropping fails to improve yield, even under extreme pest pressure. In contrast, when trap crops are highly attractive (e.g., a = 49), yield gains can be substantial even with relatively small land allocations to trap plants. Our theoretical results provide general conditions under which a unique optimal solution for the number of cash plants exists. In particular, the condition that the elasticity of pest density on a cash plant with respect to the number of cash plants is increasing (Condition 2 in Theorem 1) plays a central role. This elasticity captures how pest distribution responds to changes in landscape composition and connects directly to the efficiency of trap cropping. These results complement other theoretical work on pest suppression 13 strategies that use elasticity-based reasoning, in diverse fields from agricultural economics to invasive species management . Our findings have clear practical relevance. First, they demonstrate that highly attractive trap crops can offer strong control with minimal land sacrifice. However, even the most attractive trap crops require 5–10% of the landscape to be devoted to non-harvestable plants when pest pressure is severe. This finding is particularly important for organic and low-input systems, where pesticides are avoided and trap cropping is one of few viable pest management options . Second, the analysis
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
is severe. This finding is particularly important for organic and low-input systems, where pesticides are avoided and trap cropping is one of few viable pest management options . Second, the analysis emphasizes that blanket recommendations about trap crop allocation are unlikely to be effective. Instead, effective implementation requires system-specific knowledge: the value of β and a must be estimated for each crop–pest combination. Our formulas provide a valuable tool for translating this information into actionable trap cropping recommendations. As with any modeling approach, our framework makes simplifying assumptions. We assume a fixed pest population (N constant), homogeneous mixing of pests across the landscape, and identical cash and trap plants aside from attractiveness. These assumptions ignore important ecological complexities, such as spatial heterogeneity in plant placement within a field, pest dispersal behaviors, and pest reproduction . Furthermore, we consider an agricultural field or greenhouse in isolation, ignoring its place within a landscape of natural areas, roads, and other agricultural fields, where optimal strategies may depend on neighboring pest management strategies . Additionally, our model also treats trap plant effectiveness solely in terms of attractiveness, without considering pest arrestment, retention, or mortality on trap plants. In systems where trap plants kill or immobilize pests (e.g., through glandular trichomes or nectar-feeding
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
attractiveness, without considering pest arrestment, retention, or mortality on trap plants. In systems where trap plants kill or immobilize pests (e.g., through glandular trichomes or nectar-feeding traps), trap cropping will likely be more effective, and our model’s recommendations may not hold. Nonetheless, our framework is both general and flexible. We chose to demonstrate it with a deliberately simple and analytically tractable model, designed to clarify the fundamental trade-offs in trap cropping and to serve as a theoretical foundation for future extensions. By distilling the problem to its essential elements, the model provides general insights that remain easily interpretable and transferable. Specifically, the fact that analytic solutions are achieved and only depend on two parameters allowed us to explore the whole parameter space, presenting a complete analysis. This means future researchers adapting the model to incorporate more complex mechanisms have a concrete baseline that they can robustly compare the results to. Several avenues merit further exploration. Spatially explicit models could help examine how the arrangement of trap and cash plants influences pest dynamics, building on work in habitat fragmentation and landscape ecology . Stochastic models could incorporate variable pest dispersal patterns or plant preferences. Further, while our framework was designed for trap cropping, other systems where the grower sacrifices space via companion planting to
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
variable pest dispersal patterns or plant preferences. Further, while our framework was designed for trap cropping, other systems where the grower sacrifices space via companion planting to increase yield in the rest of the field may also work for our framework. For example, a modified version of the model could be used to optimize the amount of area devoted to companion plants that attract beneficial organisms such as pollinators, which, unlike pests, enhance rather than reduce crop yield . In this context, the spatial trade-off between yield-producing and service-providing plants remains, but the ecological mechanism is reversed. Such applications may help inform broader ecological management strategies in agriculture. Trap cropping remains a promising pest management strategy, especially for growers seeking to minimize chemical inputs. However, its success depends critically on the amount of space the grower allocates to trap plants. Our framework offers a tractable approach for identifying optimal land allocations. As such, it provides a foundation for both future theoretical exploration and practical decision-making in sustainable pest management.
|
{
"Authors": "Matthew H Holden",
"Published": "2025-08-07",
"Summary": "Trap cropping is a pest management strategy where a grower plants an\nattractive \"trap crop\" alongside the primary crop to divert pests away from it.\nWe propose a simple framework for optimizing the proportion of a grower's field\nor greenhouse allocated to a main crop and a trap crop to maximize agricultural\nyield. We implement this framework using a model of pest movement governed by\ntrap crop attractiveness, the potential yield threatened by pests, and\nfunctional relationships between yield loss and pest density drawn from the\nliterature. Focusing on a simple case in which pests move freely across the\nfield and are attracted to traps solely by their relative attractiveness, we\nfind that allocating 5-20 percent of the landscape to trap plants is typically\nrequired to maximize yield and achieve effective pest control in the absence of\npesticides. For highly attractive trap plants, growers can devote less space\nbecause they are more effective; less attractive plants are ineffective even in\nlarge numbers. Intermediate attractiveness warrants the greatest investment in\ntrap cropping. Our framework offers a transparent and tractable approach for\nexploring trade-offs in pest management and can be extended to incorporate more\ncomplex pest behaviors, crop spatial configurations, and economic\nconsiderations.",
"Title": "Optimal trap cropping investments to maximize agricultural yield",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
GPT-4 AS EVALUATOR: EVALUATING LARGE LANGUAGE MODELS ON PEST MANAGEMENT IN AGRICULTURE Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang* ABSTRACT In the rapidly evolving field of artificial intelligence (AI), the application of large language models (LLMs) in agriculture, particularly in pest management, remains nascent. We aimed to prove the feasibility by evaluating the content of the pest management advice generated by LLMs, including the Generative Pre-trained Transformer (GPT) series from OpenAI and the FLAN series from Google. Considering the context-specific properties of agricultural advice, automatically measuring or quantifying the quality of text generated by LLMs becomes a significant challenge. We proposed an innovative approach, using GPT-4 as an evaluator, to score the generated content on Coherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and Exhaustiveness. Additionally, we integrated an expert system based on crop threshold data as a baseline to obtain scores for Factual Accuracy on whether pests found in crop fields should take management action. Each model’s score was weighted by percentage to obtain a final score. The results showed that GPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories. Furthermore, the use of instruction-based prompting containing domain-specific knowledge proved the feasibility of LLMs as an effective tool in agriculture, with an accuracy rate of 72%, demonstrating
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
the use of instruction-based prompting containing domain-specific knowledge proved the feasibility of LLMs as an effective tool in agriculture, with an accuracy rate of 72%, demonstrating LLMs’ effectiveness in providing pest management suggestions. Index Terms— Large Language Model, Prompt Engineering, Large Language Model Evaluation, Agriculture, Pest Management 1. INTRODUCTION Language models (LMs), as computer algorithms or systems, are capable of understanding and generating human language, contributing a core component of the field of natural language processing (NLP) [1]. Language models are trained on a vast corpus of text data [2], enabling the model to capture word order or contextual associations, which allows the model to predict the next word or a sequence of words based on a particular probability distribution given an input [2, 3, 4]. LLMs are sophisticated LMs with a considerably larger scale, encompassing billions or hundreds of billions of parameters, and are typically founded upon deep learning methodologies [1]. In contrast to standard LMs, LLMs necessitate massive data for training, thereby enabling LLMs with a broad expanse of knowledge and generalization capabilities. LLMs exhibit enhanced adaptability to a diverse range of tasks and domains [5, 6]. Large, pre-trained language models (PLMs) like BERT (Bidirectional Encoder Representations from Transformers) and GPT have significantly altered the NLP landscape, delivering state-of-the-art results across
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
pre-trained language models (PLMs) like BERT (Bidirectional Encoder Representations from Transformers) and GPT have significantly altered the NLP landscape, delivering state-of-the-art results across various tasks [7]. Traditional NLP methods require handcrafted features and task-specific training, whereas PLMs use a generic latent feature representation learned from extensive training on a wide range of texts adapted for specific NLP tasks [7]. LLMs such as GPT-3.5 and GPT-4 have demonstrated remarkable capabilities as general-purpose computational tools, conditioned by natural language instructions. The efficacy of these models in task performance is substantially contingent upon the quality of prompts used to guide them. Notably, most effective prompts are crafted manually by humans [8]. Prompt engineering emerges as a pivotal area within AI, dedicated to optimising prompts to proficiently direct AI models, particularly those grounded in machine learning and NLP. The emerging research domain includes the design, refinement, and implementation of prompts or instructions that steer the output of LLMs, facilitating the completion of diverse tasks. Generally, LLMs are pre-trained on a massive corpus of unlabeled data to capture a broad understanding of language and knowledge. Followed by small fine-tuning, LLMs are adapted to task-specific datasets to particular applications of interest [9]. Consequently, identifying appropriate evaluation metrics for LLMs across diverse
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
Followed by small fine-tuning, LLMs are adapted to task-specific datasets to particular applications of interest [9]. Consequently, identifying appropriate evaluation metrics for LLMs across diverse domains has emerged as a novel and significant research theme. Due to the efficiency in understanding and generating human language, LLMs have been applied across various domains, including finance, medicine and education. However, their adoption in agriculture has been limited, constrained by the field’s specialized nature and the paucity of research exploring their potential in this area. The main contributions of our paper can be summarized as follows: 1 arXiv:2403.11858v1 [cs.CL] 18 Mar 2024 1. Feasibility Study of LLMs for Pest Management Advice Generation in Agriculture: We demonstrate the viability of LLMs in the agricultural pest management domain. 2. Innovative Evaluation Methodology: We introduce a novel approach using GPT-4 for multi-dimensional assessment of generated pest management suggestions. 3. Effective Application of Instruction-Based Prompting Techniques: Our findings highlight a 72% accuracy in LLMdriven pest management decisions through instruction-based prompting that incorporates domain-specific knowledge. 4. Nuanced Differences Between GPT-3.5 and GPT-4: Our research uncovers subtle differences between GPT-3.5 and GPT-4 in decision-making on pest management, emphasizing the importance of model selection in agricultural contexts. 2. RELATED WORK 2.1.
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
Our research uncovers subtle differences between GPT-3.5 and GPT-4 in decision-making on pest management, emphasizing the importance of model selection in agricultural contexts. 2. RELATED WORK 2.1. Application of LLMs “FinBERT” is a LM tailored in financial domain, a variant of the BERT model where lies in the specialized pre-training on financial texts, enabling the adaptability to handle the distinctive language and expressions prevalent in the financial sector. “FinBERT” has been applied for financial text mining [10], financial sentiment analysis [11], and financial communications [12]. However, the specialization of “FinBERT” is limit on effectiveness in domains outside of finance as the model’s performance is highly dependent on the quality and representativeness of the financial corpus used for training [10, 11, 12]. Beyond “FinBERT”, Xiao-Yang et al. [13] have introduced “FinGPT”, a novel model based on the transformer architecture, aimed at enhancing the applicability of LLMs in the financial domain. “FinGPT” addresses the limitations in data acquisition and processing faced by traditional financial LLMs by automating the collection of real-time financial data from the Internet. In evaluating LLMs in educational domains, Kung et al. [14] demonstrated that ChatGPT could achieve scores at or near the passing threshold for all three components of the United States Medical Licensing Exam without specific training or reinforcement, underscoring the potential of LLMs to
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
achieve scores at or near the passing threshold for all three components of the United States Medical Licensing Exam without specific training or reinforcement, underscoring the potential of LLMs to support medical education and possibly influence clinical decision-making processes. Similarly, Thirunavukarasu et al. [15] discussed the use of LLMs in healthcare, which covered development and applications in clinics. The review guides clinicians on using LLM technology for patient and practitioner benefits. In agriculture, Dr Som [16] explored the potential applications of OpenAI’s LLM, ChatGPT. Specifically, the paper discusses using ChatGPT across various agricultural tasks, including crop forecasting, soil analysis, crop disease and pest identification. Dr Som highlights that ChatGPT exhibits professional competence in analyzing agricultural data to generate accurate and timely reports, alerts, and insights, facilitate informed decision-making, and enhance customer service. However, it is noted that ChatGPT’s predictions’ accuracy relies heavily on input data quality. Inaccurate, biased, or incomplete data can significantly impact the model’s outputs. Moreover, AI systems like ChatGPT can assist decision-making but are not a substitute for human intuition and experience in complex agricultural environments [16]. Besides, Silva et al. [17] evaluate the capability of LLMs, including GPT-4, GPT-3.5, and Llama2, in responding to agriculturally-related queries. The queries were
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
agricultural environments [16]. Besides, Silva et al. [17] evaluate the capability of LLMs, including GPT-4, GPT-3.5, and Llama2, in responding to agriculturally-related queries. The queries were sourced from agricultural examinations and datasets from the United States, Brazil, and India. The study assessed the accuracy of answers produced by LLMs, the effectiveness of retrieval-augmented generation (RAG) and ensemble refinement (ER) techniques, and the comparative performance against human respondents. Silva et al. [17] discovered that in various tasks, GPT-4 performed better than GPT-3.5 and Llama2, achieving an impressive 93% accuracy rate in the certified crop adviser (CCA) exam. Additionally, in the study by Jiajun et al. [18], the application of LLMs, particularly GPT-4, in agriculture for pest and disease diagnosis is explored. Jiajun [18] introduces a novel approach that combines the deep logical reasoning capabilities of GPT-4 with the visual comprehension abilities of the You Only Look Once (YOLO) network. The paper evaluates the YOLO-PC, a new lightweight variant of YOLO, using metrics such as accuracy rate (94.5%) and reasoning accuracy (90% for agricultural diagnostic reports), assessing the quality of model-generated text in correlation with the recognized information [18]. 2.2. Prompt & Prompt Engineering Prompts are a mechanism for interaction with large language models (LLMs) to accomplish specific tasks [19]. Prompts act as essentially instructions
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
information [18]. 2.2. Prompt & Prompt Engineering Prompts are a mechanism for interaction with large language models (LLMs) to accomplish specific tasks [19]. Prompts act as essentially instructions directed towards LLMs, comprise the input provided by users and guide the model to generate answers for the response [20]. The nature of the inputs are vary, encompassing explanations, queries, or any other form of input, contingent upon the intended application of the model [19]. In contrast to traditional supervised learning, where models are trained to predict output from input using a probability distribution, prompt-based learning operates on LLMs that directly model textual probabilities. Prompt-based learning involves modifying the original input into a text string prompt with unfilled slots using templates. Subsequently, prompts are populated using the probabilistic capabilities of the LLM to generate the final 2 string [21]. Essentially, prompt engineering represents a practice of engaging effectively with AI systems to optimise the utility [22]. In addition, prompt engineering has been applied in various domains such as medical [22, 23, 24], generative art [25], multilingual legal judgment prediction [26], and the extraction of accurate materials data [27]. As delineated in the “Prompt Engineering Guide [28]”, constructing an effective prompt can involve integrating four elements or a combination: Instruction, Context, Input Data and Output Indicator. Instruction
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
in the “Prompt Engineering Guide [28]”, constructing an effective prompt can involve integrating four elements or a combination: Instruction, Context, Input Data and Output Indicator. Instruction refers to a specific task or directive to guide the model to perform a designated operation. Context encompasses providing supplementary information or background, instrumental in steering the model towards more accurate responses. Input Data pertains to the specific question or input content the model solicits to respond to. Lastly, the Output Indicator concerns the desired type or format of the model’s output. The iterative development process also outlined four prompt guidelines [29]: • Be clear and specific: The prompts should be unambiguous and detailed enough to guide the model precisely towards the intended task or output. • Analyze why the result does not give the desired output: If the output from the prompt does not meet expectations, it is crucial to analyze the reasons behind the discrepancy. • Refine the idea and the prompt: Based on the analysis, adjustments should be made to both the underlying idea and the wording or structure of the prompt to improve results. • Repeat: The process is not linear but cyclical, after refining, the new prompt is tested, and the cycle of analysis and refinement continues until the desired outcome is achieved. 3. EXPERIMENT DESIGN 3.1. Experiment Models This section provides an overview of the two LLMs evaluated in the experiment: Section
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
and refinement continues until the desired outcome is achieved. 3. EXPERIMENT DESIGN 3.1. Experiment Models This section provides an overview of the two LLMs evaluated in the experiment: Section 3.1.1 covers the GPT series from OpenAI, specifically GPT-3.5 and GPT-4, while Section 3.1.2 describes the FLAN-T5 model developed by Google. 3.1.1. GPT The transformer architecture, proposed by Vaswani et al. in the paper “Attention is All You Need” [30], became the cornerstone for the GPT [31, 32]. The OpenAI GPT model, introduced in the paper ”Improving Language Understanding by Generative PreTraining” [33], undergoes pre-training through language modelling on a substantial dataset to capture long-range dependencies within the text. Due to the GPT model’s advanced capability to understand and generate human-like text [34], it becomes an ideal choice for exploring complex agriculture tasks and serves as the experimental model. Specifically, the GPT-3.5 and GPT-4 models were used in the experiments. GPT-3.5 and GPT-4 are successive generations of artificial intelligence language models developed by OpenAI. The GPT-3.5 model is proficient in understanding and generating natural language or code and has been optimized specifically for chat-based interactions through the Chat Completion API. However, it remains applicable to non-chat tasks. The GPT-4 model, as a large multi-modal model, exhibits a broader comprehension of general knowledge and reasoning capabilities, enabling it to
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
API. However, it remains applicable to non-chat tasks. The GPT-4 model, as a large multi-modal model, exhibits a broader comprehension of general knowledge and reasoning capabilities, enabling it to solve complex problems with greater accuracy compared to GPT-3.5 and its predecessors [35]. 3.1.2. FLAN-T5 The T5 model significantly advances natural language processing through its novel unified framework. T5 converts all language problems into a text-to-text format, facilitating extensive exploration of transfer learning techniques. Employing a combination of supervised and self-supervised training methods, including a novel use of corrupted tokens for pre-training, T5 sets new benchmarks across a range of NLP tasks by leveraging its encoder-decoder architecture and the extensive “Colossal Clean Crawled Corpus” [36]. FLAN-T5 is an evolution of the original T5 model, which was fine-tuned on over a thousand additional tasks and expanded language coverage. FLAN-T5 significantly enhances performance and versatility, even in few-shot scenarios, achieving state-of-the-art results on various benchmarks [37]. The FLAN-T5 model is available in various sizes, including Small, Base, Large, XL, and XXL, with the XXL version being the largest, encompassing 11 billion parameters. Unlike the GPT model, FLAN models require downloading the checkpoints locally to generate the response. Given the considerations for computational speed and memory constraints, the FLAN-T5XL variant is selected as
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
model, FLAN models require downloading the checkpoints locally to generate the response. Given the considerations for computational speed and memory constraints, the FLAN-T5XL variant is selected as a more practical option for experimental use containing 3 billion parameters. The pre-trained model 3 “google/flan-t5-xl” weights and configuration are loaded using the transformers library. The weights and configurations are based on the previously saved checkpoint containing all the model parameters. 3.2. Baselines This section elucidates the methodology for generating labelled samples used to construct pest scenarios based on the expert system. To assess the ability of LLMs to determine whether specific pest scenarios necessitate action, a baseline of labelled samples is essential. Section 3.2.1 delineates the composition of the expert system, including four data files, while Section 3.2.2 elaborates on the process of generating labelled samples from the expert system’s data for the construction of pest scenarios. 3.2.1. Expert System As the baseline for this experiment, an Expert System is used to evaluate the Factual Accuracy of three different Large Language Models (LLMs) below on whether pests found in the crop fields necessitate management actions. The Expert System comprises four datasets extracted from the AHDB’s Encyclopaedia of Pests and Natural Enemies in Field Crops [38]. These datasets include two in structured data ‘JSON’ format: ‘pest to affected crop.json’ and
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
four datasets extracted from the AHDB’s Encyclopaedia of Pests and Natural Enemies in Field Crops [38]. These datasets include two in structured data ‘JSON’ format: ‘pest to affected crop.json’ and ‘pest to threshold.json’ and two in ‘XLSX’ format ‘thresholds database.xlsx’ and ‘pest to management.xlsx’. • File ‘pest to affected crop.json’ summarises various pests and the crops. It lists different types of pests where each pest is associated with one or more crops. • File ‘pest to threshold.json’ provides information on the thresholds for pests, specifying when action should be taken to manage them. Each entry includes the pest name and a threshold description, which details the criteria for deciding when to take action, such as the temperature, location, plant stages, pest density levels and the extent of crop damage. • File ‘thresholds database.xlsx’ features a first column listing the names of pests, with other columns containing threshold information extracted from the file ‘pest to threshold.json’. The threshold includes pest density metrics such as ‘per square meter’, ‘per plant’, ‘leaf area eaten’, ‘per trap’, ‘of petioles damaged’, ‘of plants are infested’, ‘of tillers infested’, ‘per pheromone trap’, ‘per ear’, ‘per trap on two consecutive occasions’, ‘per yellow sticky trap’, and ‘per gram of soil’. • File ‘pest to management.xlsx’ has two columns, the first listing the names of pests and the second detailing management suggestions for Non-chemical control
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
sticky trap’, and ‘per gram of soil’. • File ‘pest to management.xlsx’ has two columns, the first listing the names of pests and the second detailing management suggestions for Non-chemical control solutions that meet the criteria for affected plants and thresholds achieved. Notably, the Expert System is designed to output Non-chemical control solutions only when all specified conditions are met, defined as action is necessitated. As the benchmark for Factual Accuracy, the Expert System is not engaged in evaluating the accuracy, being designated as unequivocally 100% accurate. Only the outputs, Non-chemical control solutions, from expert systems are subject to evaluation by GPT-4, focusing on Coherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and Exhaustiveness. 3.2.2. Generation of Input Samples from Expert System Figure 1 shows the process for generating labelled pest samples. Files ‘pest to affected crop.json’, ‘thresholds database.xlsx’, and ‘pest to management.xlsx’ are used for the generation of labelled pest samples, serving as inputs for constructing prompts for LLMs. Although the data across these files are indexed by pest name, variations exist in the pests included due to the differing extraction methods employed from the AHDB database [38]. By querying pests of the same species, 25 types of pests, along with their affected crops, thresholds and Non-chemical control solutions, have been extracted. The ‘generate densities’ function provides a
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
pests of the same species, 25 types of pests, along with their affected crops, thresholds and Non-chemical control solutions, have been extracted. The ‘generate densities’ function provides a mechanism for determining ‘true’ and ‘false’ density values. By iterating through a list of density-related columns in file ‘thresholds database.xlsx’, the function searches for non-null entries that signify recorded density thresholds. When encountering a valid density value, the function performs a series of operations to cleanse and standardize the data, including removing percentage symbols or relational operators. Subsequently, the numerical density value is manipulated to generate a series of ‘true’ densities, inflating the original value by adding a random integer ranging from 1 to 10 to simulate density conditions exceeding the threshold for pest management action. Conversely, ‘false’ densities are generated by subtracting a random integer from the original value, ensuring the resultant value does not fall below zero. These reduced values represent conditions below the pest management threshold, indicating no action is needed. The generated ‘true’ and ‘false’ density values are then appended with the original measurement metric (e.g., ‘per plant’, ‘per square meter’) and a percentage symbol if the original value was expressed as a percentage. 4 Data Extraction and Indexing thresholds_database.xlsx Generate True (exceeding threshold) and False (below threshold) density values by
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
symbol if the original value was expressed as a percentage. 4 Data Extraction and Indexing thresholds_database.xlsx Generate True (exceeding threshold) and False (below threshold) density values by manipulating the original density values with random integers. Density Values Generation thresholds_database.xlsx pest_to_management.xlsx pest_to_affected_crop.json Query and extract pests by indexing pest names, obtaining 25 pests with their affected crops, thresholds, and non-chemical control solutions. 25 Pests Affected Crops Generation pest_to_affected_crop.json True Densities False Densities True Crops False Crops Samples Generation Pest Name: … True Crops: … False Crops: … True Densities: … False Densities: … Control Solutions: … Random Temperature Random Location 1. Generate Positive Samples (1) by pairing True crops with True density values. 2. Generate Negative Samples (0) by either pairing True crops with False densities or False crops with any density values. 3. Augment samples with random temperature and latitude parameters. 4. Randomly select one positive (labelled 1) and one negative (labelled 0) sample for each pest, resulting to a total of 50 samples for experiment. Positive Samples (1): True Crops + True Densities Negative Samples (0): True Crops + False Densities False Crops + True Densities False Crops + False Densities Positive Sample Pest Name: … Crop: … Density: … Temperature: … Location: … Label: 1 Negative Sample Pest: … Crop: … Density: … Temperature: …
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
Crops + True Densities False Crops + False Densities Positive Sample Pest Name: … Crop: … Density: … Temperature: … Location: … Label: 1 Negative Sample Pest: … Crop: … Density: … Temperature: … Location: … Label: 0 Generate True (crops affected by the pest) and False (crops unaffected by the pest) crops by classifying the data file pest_to_affected_crop.json. Fig. 1: Generation of Input Samples from Expert System. This image outlines a process for generating labelled pest samples, detailing steps from data extraction and indexing of pests, through generating true and false density values and crops, to the creation of positive and negative samples for experiment. The core of the sample generation process is to create the combinations that represent an action that is needed (labelled as ‘1’) and not needed (labelled as ‘0’) for various pest conditions. This bifurcation is achieved by deliberately pairing crops and pest density values under varying conditions. Positive samples are formulated by coupling ‘true’ crops (crops affected by the pest) with ‘true’ density values. Conversely, negative samples emerge from the strategies: pairing ‘true’ crops with ‘false’ densities and ‘false’ crops (crops unaffected by the pest) with either ‘true’ or ‘false’ densities. These combinations are augmented with randomly generated temperature and latitude location parameters to diversify the dataset further. Considering computational resources and experimental costs, also ensuring an even
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
are augmented with randomly generated temperature and latitude location parameters to diversify the dataset further. Considering computational resources and experimental costs, also ensuring an even distribution of positive and negative samples, one positive sample (labelled as ‘1’) and one negative sample (labelled as ‘0’) are randomly selected for each of the 25 pest types. Eventually, a total of 50 samples are generated for experimentation. The samples are indexed in pest names, with other columns containing crops, pest density, temperature and location. Among these, crops and pest density determine the label as 1 or 0, whereas temperature and location only enrich the scene and do not affect the label. 3.3. Experiment Prompting This section lists the prompts constructed using different techniques in the experiment. Four prompt techniques: zero-shot prompting described in Section 3.3.1, few-shot prompting in Section 3.3.2, instruction-based prompting in Section 3.3.3, and self-consistency prompting in Section 3.3.4, incorporate samples of pest scenarios generated in Section 3.2.2 into prompts. These prompts serve as inputs for LLMs to generate responses, which are then evaluated. 3.3.1. Zero-shot Prompting Zero-shot prompting refers to providing instructions or requests to LLMs without needing prior examples or contextual information. Zero-shot prompting necessitates the ability of the model to comprehend and respond to tasks or queries not directly encountered before
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
LLMs without needing prior examples or contextual information. Zero-shot prompting necessitates the ability of the model to comprehend and respond to tasks or queries not directly encountered before [39]. The model relies on the extensive knowledge and understanding acquired during the training phase for zero-shot prompting. For instance, when posed with a question that has not previously been addressed, the model can still understand and attempt to provide an answer [40]. 5 For zero-shot prompting, 50 input samples are iteratively filled into the following prompt template via a loop: I discovered {Pest} in my {Crop}, with a density of {Density}. The temperature was {Temperature}, and the location was at {Location}. Could you please provide some control and management suggestions? This prompt is then input into a GPT or FLAN model. 3.3.2. Few-shot Prompting In contrast to zero-shot prompting, few-shot prompting provides relevant examples to guide the model to understand and execute a task. Few-shot prompting can be employed to facilitate in-context learning, where demonstrations in the prompt guide the model towards enhanced performance [41]. According to Min et al. [42], in the context of few-shot learning, both the label space and the distribution of the input text defined by the demonstrations are crucial for performance, irrespective of the accuracy of individual labels. Additionally, the format of the demonstrations, including random labels, significantly influences
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
by the demonstrations are crucial for performance, irrespective of the accuracy of individual labels. Additionally, the format of the demonstrations, including random labels, significantly influences effectiveness, which is better than not using any labels. The core of the few-shot learning approach is encapsulated within a create_prompt function. The function filters 50 input samples to select only samples with a label of ‘1’ and a pest different from the current input pest. It randomly selects three samples and constructs a few-shot prompt containing questions and answers. Each question is formulated from the pest, crop, density, temperature, and location of the selected input samples, same as the zero-shot prompting template, followed by the respective Non-chemical control solutions from file ‘pest to management.xlsx’. Finally, the prompt adds a new question using the current input sample without providing an answer. The template of the few-shot prompt is shown below: Question: I discovered {Pest 1} in my {Crop 1}, with a density of {Density 1}. The temperature was {Temperature 1}, and the location was at {Location 1}. Could you please provide some control and management suggestions? Answer: {Non-chemical control solutions for Pest 1} Question: I discovered {Pest 2} in my {Crop 2}, with a density of {Density 2}. The temperature was {Temperature 2}, and the location was at {Location 2}. Could you please provide some control and management suggestions? Answer: {Non-chemical
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
2}, with a density of {Density 2}. The temperature was {Temperature 2}, and the location was at {Location 2}. Could you please provide some control and management suggestions? Answer: {Non-chemical control solutions for Pest 2} Question: I discovered {Pest 3} in my {Crop 3}, with a density of {Density 3}. The temperature was {Temperature 3}, and the location was at {Location 3}. Could you please provide some control and management suggestions? Answer: {Non-chemical control solutions for Pest 3} Question: I discovered {Pest} in my {Crop}, with a density of {Density}. The temperature was {Temperature}, and the location was at {Location}. Could you please provide some control and management suggestions? Answer: 3.3.3. Instruction-based Prompting As mentioned in Section 2.2, constructing an effective prompt can involve any of the four elements: Instruction, Context, Input Data and Output Indicator [28]. Giray [43] discussed the importance of understanding the prompt component and its role in facilitating effective communication with the model. Through prompt design with these four elements, Giray [43] found one can guide model behaviour and improve response quality, ensuring output is precise and meaningful. The template of the instruction-based prompt is: Instruction: Generate comprehensive and sustainable pest management suggestions based on the given crop, pest type and density, and environmental conditions, including temperature and location. Context: Pest management in
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
comprehensive and sustainable pest management suggestions based on the given crop, pest type and density, and environmental conditions, including temperature and location. Context: Pest management in agriculture requires balancing control measures with environmental sustainability. Different crops and pests respond to varied strategies, and local environmental conditions significantly influence the effectiveness of these strategies. Input Data: For example: For pest: {Pest} The affected crops are: {Affected Crops} The threshold is: {Threshold} The non-chemical control solution could be: {Non-chemical control solutions} Output Indicator: Question: I discovered {Pest} in my {Crop}, with a density of {Density}. The temperature was {Temperature}, and the location was at {Location}. Could you please provide some control and management suggestions? Please first determine whether management measures are needed, then output your own control solution in about 200 words. The Instruction defines the pest management task for the model and guides the model to focus on the data in the input question. The context explains why pest management in agriculture is essential, helping the model better understand the broader implications of pest management and the necessity of tailoring suggestions to specific scenarios. The structured example 6 systematically introduces the input data, incorporating placeholders for designated variables. Precisely, the {Pest} variable corresponds to the pest
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
to specific scenarios. The structured example 6 systematically introduces the input data, incorporating placeholders for designated variables. Precisely, the {Pest} variable corresponds to the pest identified in the inquiry, while the {Affected Crops} are derived from the ‘pest to affected crop.json’ file. Similarly, the {Threshold} values are extracted from the ‘thresholds database.xlsx’ file, and the {Non-chemical control solutions} are obtained from the ‘pest to management.xlsx’ file. All input data variables are dynamically populated based on the specific pest mentioned in the question. 3.3.4. Self-consistency Prompting Self-consistency is an advanced prompting technique introduced by Wang et al. [44] building upon chain-of-thoughts (CoT). This innovative approach involves generating diverse reasoning paths rather than relying on the most immediately probable path. Self-consistency then deduces the most consistent answer by aggregating across these varied reasoning paths. Self-consistency prompting summarised the responses from the zero-shot, few-shot, and instruction-based prompting and gave a final response. The template of self-consistency prompting is: Given these three responses: Response 1: {Response 1 from zero-shot prompting} Response 2: {Response 2 from few-shot prompting} Response 3: {Response 3 from instruction-based prompting} Create a summary response that combines the best elements of question: I discovered {Pest} in my {Crop}, with a density of {Density}.
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
Response 3: {Response 3 from instruction-based prompting} Create a summary response that combines the best elements of question: I discovered {Pest} in my {Crop}, with a density of {Density}. The temperature was {Temperature}, and the location was at {Location}. Could you please provide some control and management suggestions? 3.4. GPT-4 as Evaluator Twelve combinations emerge when integrating FLAN, GPT-3.5, and GPT-4 models with four prompting methodologies. Each combination is subjected to fifty input pest samples characterized by varying density and environmental conditions, generating respective responses. These responses are then evaluated by GPT-4 regarding the accuracy and the linguistic quality of the generated pest management suggestions. The prompt guiding the GPT to serve as an evaluator for determining the necessity of action in responses and assessing the linguistic quality of these responses draws inspiration from the article “G-EVAL: NLG Evaluation using GPT-4 with Better Human Alignment”, where the article introduces the G-EVAL framework, designed for evaluating the quality of text generated by Natural Language Generation (NLG) systems [45]. For accuracy evaluation, the prompt begins with the ‘Evaluation Guide’, which instructs the GPT to assess and decide whether action is required. The prompt followed with the ‘Evaluation Criteria’ to inform the GPT that this is a binary evaluation to assign ‘1’ or ‘0’. Through the Evaluation Steps, the GPT is guided with
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
action is required. The prompt followed with the ‘Evaluation Criteria’ to inform the GPT that this is a binary evaluation to assign ‘1’ or ‘0’. Through the Evaluation Steps, the GPT is guided with the CoT sequence, asked to decide whether the information presented within a response indicates that an action is required (‘1’) or not required (‘0’) in the following prompt and the corresponding response. The template for accuracy evaluation is shown below: Evaluation Guide: You will be provided with a prompt and the corresponding response for pest management. Your task is to evaluate the response based on the criteria below and decide whether action is required based on the response. Please read and understand these instructions carefully. Refer back to this document as needed during your evaluation. Evaluation Criteria: Action Required (1 or 0) This is a binary evaluation to determine if action is needed based on the response provided. Evaluation Steps: 1. Carefully read the pest management suggestion in the response, identifying the main content, pay special attention to the first sentence in the response, as it generally contains the decision of whether to take actions. 2. Analyze the response to see if it states whether action is required or not required to manage the pest. 3. Assign a score based on the evaluation criteria: 0 means no action is needed, 1 means the suggestion requires action. 4. If the response suggests the action is optional, needs further observation or
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
Assign a score based on the evaluation criteria: 0 means no action is needed, 1 means the suggestion requires action. 4. If the response suggests the action is optional, needs further observation or continuous monitoring, leaves room for doubt, lacks clearly direction, contains not be necessary or not immediate control, or if you cannot determine with complete certainty that it indicates for management action, please mark it as 0. Here are the prompt and response you need to evaluate: Prompt: {Prompt} Response: {Response} Please state whether action is required (Answer 0 or 1 ONLY): The linguistic quality evaluation contains six dimensions: Coherence, Logical Consistency, Fluency, Relevance, Comprehensiveness, and Exhaustiveness. The structure of the prompt for linguistic quality evaluation is similar to accuracy evaluation, 7 comprising an Evaluation Guide with instructions, Evaluation Criteria that include scoring standards, and Evaluation Steps based on a CoT approach. Except for some differences in details and descriptors, the principal distinction lies in the judgment required from the GPT. For accuracy evaluation, the GPT is tasked with making a binary decision regarding the necessity of action. In contrast, evaluating linguistic quality required the GPT to assign scores ranging from 1 to 10 for each of the six dimensions. 4. RESULTS Model & Prompting Coherence Consistency Fluency Relevance Comprehensibility Exhaustiveness FLAN zero-shot 2.52 2.52 3.30 2.36 2.76 2.96
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
ranging from 1 to 10 for each of the six dimensions. 4. RESULTS Model & Prompting Coherence Consistency Fluency Relevance Comprehensibility Exhaustiveness FLAN zero-shot 2.52 2.52 3.30 2.36 2.76 2.96 FLAN few-shot 2.68 3.00 3.42 2.44 3.32 3.46 FLAN instruction-based 3.70 3.92 4.84 5.06 5.04 4.36 FLAN self-consistency 2.64 3.22 4.04 1.94 3.92 3.18 GPT-3.5 zero-shot 8.82 8.24 9.90 8.74 9.54 7.54 GPT-3.5 few-shot 8.14 8.24 9.86 9.26 8.36 6.28 GPT-3.5 instruction-based 8.28 8.20 9.60 8.92 9.14 6.92 GPT-3.5 self-consistency 7.98 8.00 9.80 7.70 9.44 7.16 GPT-4 zero-shot 9.14 8.88 10.00 9.86 9.38 8.74 GPT-4 few-shot 8.32 8.46 9.98 9.46 8.92 7.14 GPT-4 instruction-based 8.62 8.76 9.64 9.46 9.32 7.68 GPT-4 self-consistency 8.72 8.90 10.00 9.30 9.88 8.14 Table 1: Linguistic quality of different models and prompting methods evaluated by GPT-4 Model & Prompting TP TN FP FN Accuracy Precision Recall F1 Score Final Score FLAN zero-shot 20 6 19 5 0.52 0.51 0.80 0.62 37.22 FLAN few-shot 10 10 15 15 0.40 0.40 0.40 0.40 34.32 FLAN instruction-based 14 14 11 11 0.56 0.56 0.56 0.56 49.32 FLAN self-consistency 24 1 24 1 0.50 0.50 0.96 0.66 38.94 GPT-3.5 zero-shot 25 4 21 0 0.58 0.54 1.00 0.70 75.98 GPT-3.5 few-shot 17 8 17 8 0.50 0.50 0.68 0.58 70.14 GPT-3.5 instruction-based 24 12 13 1 0.72 0.65 0.96 0.77 79.86 GPT-3.5 self-consistency 25 0 25 0 0.50 0.50 1.00 0.67 70.08 GPT-4 zero-shot 24 4 21 1 0.56 0.53 0.96 0.69 78.40 GPT-4 few-shot 21 7 18 4 0.56 0.54 0.84 0.66 74.68 GPT-4
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
1 0.72 0.65 0.96 0.77 79.86 GPT-3.5 self-consistency 25 0 25 0 0.50 0.50 1.00 0.67 70.08 GPT-4 zero-shot 24 4 21 1 0.56 0.53 0.96 0.69 78.40 GPT-4 few-shot 21 7 18 4 0.56 0.54 0.84 0.66 74.68 GPT-4 instruction-based 24 9 16 1 0.66 0.60 0.96 0.74 79.88 GPT-4 self-consistency 25 0 25 0 0.50 0.50 1.00 0.67 74.94 Table 2: Performance metrics of different models and prompting methods with final scores Tables 1 and 2 respectively present the linguistic quality of different models and prompting methods evaluated by GPT-4, and the performance metrics of different models and prompting methods with the final scores for each model. The linguistic quality evaluation involves scoring the responses based on the generated pest management suggestions across 50 samples, each representing an average derived from these responses. In performance metrics, the TP (True Positives), TN (True Negatives), FP (False Positives), and FN (False Negatives) are used as foundational elements for calculating Accuracy, Precision, and Recall. To calculate the final scores for each “Model & Prompting” combination, we use a weighted average approach based on pre-determined weights for various evaluation metrics. Specifically, the weights for Coherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and Exhaustiveness are each allocated 10%, while Accuracy is assigned a higher weight of 40%. In the computation of the Final Score, the metrics of Coherence, Consistency, Fluency, Relevance,
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
and Exhaustiveness are each allocated 10%, while Accuracy is assigned a higher weight of 40%. In the computation of the Final Score, the metrics of Coherence, Consistency, Fluency, Relevance, Comprehensibility, and Exhaustiveness are evaluated on a scale from 1 to 10. In contrast, Accuracy is averaged, falling in a range from 0 to 1. To harmonize these scores for a unified presentation in a percentage format, the scores for the linguistic quality are multiplied by 10, and the Accuracy score is multiplied by 100, facilitating a standardized evaluation outcome expressed on a 100-point scale. 8 The mathematical formulation for the final score for each model can be expressed as follows: Final Score = 0.1 × (V alCoherence + V alConsistency + V alFluency + V alRelevance + V alComprehensibility + V alExhaustiveness) × 10 + 0.4 × V alAccuracy × 100 (1) Where V alCoherence, V alConsistency, V alFluency, V alRelevance, V alComprehensibility and V alExhaustiveness respectively represent the numerical values scored by GPT-4 for the dimensions of Coherence, Consistency, Fluency, Relevance, Comprehensibility and Exhaustiveness as listed in Table 1, and V alAccuracy denotes the average accuracy value in Table 2. From Table 1, it can be observed that the performance of the different models and their application of different prompting methods on the various dimensions of language quality. Specifically, the FLAN model scores low on each assessed dimension, showing its understanding and
|
{
"Authors": "Shanglong Yang, Zhipeng Yuan, Shunbao Li, Ruoling Peng, Kang Liu, Po Yang",
"Published": "2024-03-18",
"Summary": "In the rapidly evolving field of artificial intelligence (AI), the\napplication of large language models (LLMs) in agriculture, particularly in\npest management, remains nascent. We aimed to prove the feasibility by\nevaluating the content of the pest management advice generated by LLMs,\nincluding the Generative Pre-trained Transformer (GPT) series from OpenAI and\nthe FLAN series from Google. Considering the context-specific properties of\nagricultural advice, automatically measuring or quantifying the quality of text\ngenerated by LLMs becomes a significant challenge. We proposed an innovative\napproach, using GPT-4 as an evaluator, to score the generated content on\nCoherence, Logical Consistency, Fluency, Relevance, Comprehensibility, and\nExhaustiveness. Additionally, we integrated an expert system based on crop\nthreshold data as a baseline to obtain scores for Factual Accuracy on whether\npests found in crop fields should take management action. Each model's score\nwas weighted by percentage to obtain a final score. The results showed that\nGPT-3.4 and GPT-4 outperform the FLAN models in most evaluation categories.\nFurthermore, the use of instruction-based prompting containing domain-specific\nknowledge proved the feasibility of LLMs as an effective tool in agriculture,\nwith an accuracy rate of 72%, demonstrating LLMs' effectiveness in providing\npest management suggestions.",
"Title": "GPT-4 as Evaluator: Evaluating Large Language Models on Pest Management in Agriculture",
"Topic": "Pests & Diseases",
"source": null,
"summary": null,
"title": null
}
|
End of preview. Expand
in Data Studio
README.md exists but content is empty.
- Downloads last month
- 9