Juancc-ctic/TerritoryExpert_assessment

Territory Expert Assessment Queries Dataset

This dataset was built for evaluating the multimodal Large Language Model (LLM) within the agricultural domain presented in the publication at the bottom of the page.

Data Structure

The dataset contains the following columns visible in the viewer table above:

• QueryID (string): The unique identifier for each evaluation setup file.
• LU (string): Land Use classification code of the parcel.
• Query (string): The natural language prompt.
• Parcel (image): Rendered imagery of the specific parcel corresponding to the query.
• Response (string): The obtained response.
• Context (string): The retrieved context used to elaborte the answer.

Evaluation methodology

This section details the systematic approach used to validate the tool's performance and the contribution of its multimodal architecture. The experimental design is divided into two main phases, totaling 143 evaluation queries.

1. Plot Selection and Heterogeneity

A total of 15 plots were selected, categorized into four distinct Land Use (LU) types. The distribution is weighted according to the operational utility of the tool for different professional profiles:

Land Use	Count
Fruit Tree (FR)	7
Pasture (PA)	5
Unproductive (UN)	2
Forest (FO)	1

To ensure a bias-free assessment, the plots were selected to represent a high degree of heterogeneity in their physical characteristics, including varying slopes (0.5% to 21.6%), altitudes, irregular shapes, and diverse vegetative health (NDVI ranges).

2. Phase I: Broad Performance Assessment

This phase evaluates the system’s overall capability across diverse scenarios. A total of 71 queries were designed and distributed as follows:

• Productive Plots (FR & PA): 5 queries per plot, focusing on irrigation, fertilization, pest management, and investment ROI.
• Non-Productive Plots (UN & FO): 4 queries per plot, consisting of 2 Functional Queries and 2 Negative Testing Querie (scenarios where the system must identify unsuitability)

3. Phase II: Ablation Study (Architecture Validation)

To quantify the impact of each context module, an ablation study was conducted using a factorial design. The architecture features three binary modules (Orthophoto, Satellite, and RAG), resulting in $2^3 = 8$ possible configurations.

• Configurations: 8 modular combinations + 1 Zero-Shot baseline (LLM without parcel context) = 9 total configurations.
• Sample Subset: 1 representative plot per Land Use category.
• Query Volume: 2 specific queries per plot across all 9 configurations -> $4 \text{ Land Uses} \times 2 \text{ queries} \times 9 \text{ configurations} = \mathbf{72 \text{ queries}}$

Configurations

Modules	R	O	S	PD	Active tool	Code
Baseline	0	0	0	1	1	0
B + SAT	0	0	1	1	1	1
B + ORTHO	0	1	0	1	1	2
B + ORTHO + SAT	0	1	1	1	1	3
B + RAG	1	0	0	1	1	4
B + RAG + SAT	1	0	1	1	1	5
B + RAG + ORTHO	1	1	0	1	1	6
Complete	1	1	1	1	1	7
ZERO SHOT	0	0	0	0	1	ZS

4. Summary of Evaluation Volume

The combined results provide a comprehensive dataset for both absolute performance measurement and incremental value analysis:

Evaluation Phase	Methodology	Queries
Broad Assessment	Sequential	71
Ablation Study	Side-by-Side	72
Total		143

J. Cañada, R. Alonso, J. Molleda and F. Díez, "A Multimodal Conversational Assistant for the Characterization of Agricultural Plots from Geospatial Open Data," in 2025 4th International Conference on Geographic Information and Remote Sensing Technology (GIRST), 2025, pp. 127-132. doi: 10.1109/GIRST67753.2025.11382141

@inproceedings{canada2025multimodal,
  author={Cañada, J. and Alonso, R. and Molleda, J. and Díez, F.},
  booktitle={2025 4th International Conference on Geographic Information and Remote Sensing Technology (GIRST)}, 
  title={A Multimodal Conversational Assistant for the Characterization of Agricultural Plots from Geospatial Open Data}, 
  year={2025},
  pages={127-132},
  doi={10.1109/GIRST67753.2025.11382141}
}