aaaifss2022_15.txt

Predicting Wildfire Risk Under Novel 21st-Century Climate Conditions
Matthew Cooper
Sust Global
595 Pacific Ave., Floor 4
San Francisco, California 94133
Abstract
Wildfires are one of the most impactful hazards associ-
ated with climate change, and in a hotter, drier world,
wildfires will be much more common than they have
historically been. However, the exact severity and fre-
quency of future wildfires are difficult to estimate, be-
cause climate change will create novel combinations of
vegetation and fire weather outside what has been his-
torically observed. This provides a challenge for AI-
based approaches to long-term fire risk modeling, as
much future fire risk is outside of the available feature
space provided by the historical record. Here, I give an
overview of this problem that is inherent to many cli-
mate change impacts and propose a restricted model
form that makes monotonic and interpretable predic-
tions in novel fire weather environments. I then show
how my model outperforms other neural networks and
logistic regression models when making predictions on
unseen data from a decade into the future.
Introduction
One way to describe the future effects of climate change is
with the phrase global weirding . The 21st century will be in-
creasingly uncanny, as we will see Caribbean beach weather
in Iceland; deserts that become soggy and green; and an Arc-
tic Ocean that is entirely free of ice, potentially by 2035
(Guarino et al. 2020). Novel assemblages of temperature,
precipitation, land cover, and vegetation will emerge that are
unlike anything in human history, giving rise to hazards un-
precedented in severity and posing major challenges to adap-
tation. Additionally, these weird conditions are a challenge
to any form of modeling that depends on rich training data,
as much of the future will be entirely outside of the feature
space of available observational data.
This is especially true in the case of wildfire, because fire
depends on two things: burnable vegetation and dry enough
conditions to ignite that vegetation. Under stable climate
conditions, weather and vegetation reach an equilibrium,
where the amount of burnable vegetation is proportional to
the amount of rainfall (See Fig. 1). However, under climate
change, we are seeing increasingly novel pairings of pre-
cipitation and vegetation (See Fig. 2). For example, Califor-
Copyright © 2022, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.nia has historically had dry summers and wet winters, lead-
ing to chaparral and spare forest vegetation communities.
However, in the past decade, California had weather condi-
tions more characteristic of a desert climate. This extremely
dry weather, coupled with high levels of vegetation, is what
has caused the unprecedented fire crisis in California (Abat-
zoglou and Williams 2016). A similar situation is occurring
in the Amazon, where tropical rainforest vegetation is expe-
riencing increasingly long dry seasons and is converting into
a tropical savanna, with fire consuming the excess biomass
(Le Roux et al. 2022).
These emerging conditions are causing significant prob-
lems for sectors like the insurance industry, which has
traditionally used historic risk to estimate future risk and
appropriately price premiums. Unable to accurately esti-
mate fire risk under unprecedented conditions, many home
insurance companies are withdrawing from fire-prone ar-
eas, leaving homeowners without coverage (Poizner 2022;
Singh 2022). Given that a typical home mortgage can last
up to 30 years, a period over which climatological and eco-
logical systems will continue to disequilibrate, it is impera-
tive that we develop better methods for estimating fire risk
that can make reasonable predictions outside of the existing
feature space provided by historic data.
Data
For this analysis, I use data on fire occurrence provided glob-
ally and at a 500 meter resolution derived from NASA’s
MODIS satellite program (Giglio et al. 2009). This dataset
goes back to November 2000 and provides a binary indica-
tor of whether a fire was observed at a given pixel at a daily
timestep. From this dataset, I collected 240 million sample
locations on a given day across the terrestrial world, over-
sampling fire occurrence to make up approximately 10% of
the dataset, but otherwise sampling completely at random.
For each sample point, I calculate a daily fire weather in-
dex known as the Keetch-Byram Drought Index, or KBDI
(Brown, Wang, and Feng 2021; Gannon and Steinberg
2021). KBDI is an index updated on a daily time step and is
indicative of the amount of water in the top 203 millimeters
of soil. A KBDI score of 0 corresponds to saturated soil and
very little fire risk, while a KBDI score of 203 indicates that
soil is dry up to 203 millimeters deep and that fire risk is very
high. To calculate historic values of this index, I use daily

Figure 1: Historically, precipitation and biomass have been
in equilibrium.
HIGH Biomass 
Rainfall HIGH LOW LOW 
Figure 2: Under climate change, precipitation and biomass
are decoupled, leading to unprecedented fire severity in Cal-
ifornia and the Amazon.
HIGH Biomass Rainfall HIGH LOW 
LOW 
Amazon 
Wildfires California 
Wildfires 
historic data on temperature and precipitation from the 10
kilometer ERA5-Land reanalysis dataset (Mu ˜noz-Sabater et
al. 2021). Additionally, to better determine the fire risk con-
text I determine the local climate zone for each point using
the Koppen-Geiger methodology (K ¨oppen 2011), as well as
the local land cover type using the 300 meter ESA land cover
dataset (ESA 2017).
For my analysis, I use observed data from November 2000
to October 2011 as my training data ( n= 135,559), and ob-
served data from November 2011 to October 2021 as my
validation data ( n= 123,428). Testing my model on obser-
vations that occurred a decade beyond the end of the train-
ing data can give me an indication of how my model will
perform over the course of the next decade. Additionally, I
subset my analysis to eastern Oregon to constrain the discus-
sion, although I have data processed and prepared for analy-
ses at a global scale.
Finally, for future estimates of fire weather to use a fea-
tures in model inference, I derive KBDI from ensembled
and bias-corrected simulations of temperature and precipita-
tion throughout the 21st century using Global Climate Mod-
els (GCMs) from the 6th Climate Model Intercomparison
Project (CMIP6) (O’Neill et al. 2016).The Problem
To better illustrate the modeling challenge presented by
novel fire conditions, also referred to as domain shift, I show
daily fire weather values (KBDI) in eastern Oregon for peri-
ods where observed KBDI scores were indicative of elevated
fire risk (KBDI >100), typically in the summer (See Fig.
3). Eastern Oregon is an area without significant historic fire
activity but is increasingly threatened by fire. There, KBDI
values are increasing every decade, with the next decade
modeled to have KBDI values at the maximum potential fire
risk. This prevalence of increasingly out-of-sample and un-
precedented fire weather is also associated with heightened
fire risk, something models trained on only historic data will
struggle to capture.
Figure 3: Shifting of fire weather towards unprecedented risk
each decade complicates empirical AI modeling. Histogram
of daily KBDI values in Eastern Oregon, by decade. Values
for 2000-2010 and 2011-2021 are observed, values for 2022-
2032 are taken from an ensemble of bias-corrected climate
models.
I further illustrate this domain shift modeling challenge
by training a simple 3-layer feed-forward neural network to
predict the probability of fire in eastern Oregon as a function
of KBDI using sample data from 2000-2011 and validation
data from 2012-2022. I compare that model against a logis-
tic regression model using the same dataset. I find that the
neural network under-estimated fire risk at high KBDI lev-
els, while the logistic regression, due to its implicit mono-
tonicity, better captured the trend of increasing fire risk with
increasing KBDI levels (See Fig. 4).
While these test datasets illustrate the nature of the prob-
lem, both models used here were quite simple. In addition
to fire weather, fire risk is heavily determined by other con-
textual factors, including biomass, land cover, long-term cli-
mate conditions, and elevation. I therefore construct more
complex models based on 24 features derived from my sam-
ple dataset, one-hot encoding for land cover type and climate
zone, as well as including terms for latitude and longitude,
allowing the models to learn location-specific fire risk rela-
tionships. Additionally, I fit a hierarchical logistic regression
using the same features as the multivariate neural network.
Overall, I find that multivariate models perform better
than univariate models based only on KBDI when evaluated
on a held out test dataset from the next decade (See Table
1). Additionally, I find that logistic regression models out-
perform neural networks on the test data, because they make

Figure 4: Observed probability of fire by KBDI value, in the
training and testing datasets. Additionally, I show the predic-
tions of a simple feed-forward neural network and a logis-
tic regression. Note that the neural network under-estimates
out-of-sample future fire risk.
predictions that are monotonic. This suggests that the neural
networks struggle to capture extreme behavior.
New Architecture
Because simple neural networks struggle to capture fire ex-
tremes under novel data domains, I propose a new neural
network architecture, based on two premises. The first is
that the relationship between KBDI and fire probability is
monotonic, and as ongoing climate change leads to condi-
tions drier than any previously observed in many locations,
it will be necessary to use models that can extrapolate mono-
tonically, such as logistic regression models. Secondly, the
parameterization of the weather-fire relationship is complex
and context dependent, with a large number of influenc-
ing variables that interact nonlinearly, requiring models like
neural networks that can handle such estimation problems.
Drawing from both of these premises, I have implemented
a neural network architecture that uses a large number of
features describing the geographic context to estimate the
parameters of a logistic model that describes the KBDI-fire
relationship in that context. In this case, I use features for
the spatial location, local land cover type, and historic cli-
mate zones indicative of prevailing vegetation communities;
however, this architecture could be extended to incorporate
other important features, such as topography, proximity to
human settlements, or aboveground biomass. This approach
has the advantage of drawing on complex interactions within
the geophysical environment that influence the relationship
between fire and weather conditions, while still being con-
strained to make predictions in line with my strong prior as-
sumption that the relationship between dryness and fire risk
is monotonic.
The model feeds a large number of features in four dense
hidden layers that condense from 32 to 8 nodes with a ReLU
activation function. The model then diverges into two sepa-
rate hidden layers, each of which converges into a single-
parameter output, which are treated as the two parameters
in a logistic regression ( 0and1). The model’s loss func-
tion is therefore the performance of those two parameters in
a logistic regression using observed KBDI, evaluated with
binary cross-entropy (See Fig. 5).Figure 5: Diagrammatic representation of fire neural net-
work used to estimate logistic regression parameters.
Linear 
Predictor 
β1 β0y
Binarized 
Cross- 
Entropy (       , ) β0 β1Linear 
Pred. y +24 Input Features 
X 
8-Node Dense 8-Node Dense 16-Node Dense 32-Node Dense 
8-Node Dense 8-Node Dense 
Loss Function 
Model R2MSE
Univariate NN 0.0091 0.0442
Logistic Regression 0.0139 0.0440
Multivariate NN 0.0156 0.0439
Hierarchical Logistic Regression 0.0166 0.0438
NN-Estimated Logistic Regression 0.0202 0.0436
Table 1: Model performance by R2and mean squared error
(MSE).
I fit a model with this architecture using the same fea-
tures as the aforementioned multivariate neural network and
find that it improves performance on R2by 22%. This archi-
tecture is able to draw on the advantages of using gradient
descent to explore complex relationships among features,
while still making predictions that are interpretable and ex-
trapolate well outside of the observed range of fire weather
values.
Conclusion
While there would be many benefits of using this method-
ology, it would have the drawback of requiring a very large
dataset, as is typical of neural network based approaches.
This would evolve the state of the art of predicting wildfires
by focusing specifically on making predictions outside of
the feature space available for training. Having better long-

term fire predictions would help state agencies and govern-
ments to eliminate risks, as they currently rely on projec-
tions that are more near-term, focusing on weekly to sea-
sonal timescales.
Neural networks provide a number of advantages and can
explore a hyper-dimensional and complex feature space ef-
ficiently. However, they are brittle outside of their training
space. In such situations where it is necessary to make pre-
dictions in the absence of available training data, predictions
must be guided by theory and model behavior must be in-
terpretable. I therefore developed an architecture that flex-
ibly draws on complex environmental variables while still
making predictions that are aligned with my theoretical prior
that drier weather leads to increased fire risk. I find that this
model performs better than other approaches when used to
make predictions a decade into the future. Given the theoret-
ical support of this approach, it is likely to be especially use-
ful for making estimates at even longer timescales of up to
two or three decades. This approach has relevance for mod-
eling many of the novel risks posed by climate change.
References
Abatzoglou, J. T., and Williams, A. P. 2016. Impact of an-
thropogenic climate change on wildfire across western US
forests. Proc. Natl. Acad. Sci. U.S.A. 113(42):11770–11775.
Brown, E. K.; Wang, J.; and Feng, Y . 2021. US wildfire
potential: a historical view and future projection using high-
resolution climate data. Environ. Res. Lett. 16(3):034060.
ESA. 2017. Land cover cci product user guide. Technical
report.
Gannon, C. S., and Steinberg, N. C. 2021. A global as-
sessment of wildfire potential under climate change utilizing
keetch-byram drought index and land cover classifications.
Environmental Research Communications 3(3):035002.
Giglio, L.; Loboda, T.; Roy, D. P.; Quayle, B.; and Justice,
C. O. 2009. An active-fire based burned area mapping
algorithm for the MODIS sensor. Remote Sens. Environ.
113(2):408–420.
Guarino, M.-V .; Sime, L. C.; Schr ¨oeder, D.; Malmierca-
Vallet, I.; Rosenblum, E.; Ringer, M.; Ridley, J.; Feltham,
D.; Bitz, C.; Steig, E. J.; et al. 2020. Sea-ice-free arctic
during the last interglacial supports fast future loss. Nature
Climate Change 10(10):928–932.
K¨oppen, W. 2011. The thermal zones of the earth accord-
ing to the duration of hot, moderate and cold periods and to
the impact of heat on the organic world. Meteorologische
Zeitschrift 20(3):351–360.
Le Roux, R.; Wagner, F.; Blanc, L.; Betbeder, J.; Gond, V .;
Dessard, H.; Funatzu, B.; Bourgoin, C.; Cornu, G.; Herault,
B.; Montfort, F.; Sist, P.; Begue, A.; Dubreuil, V .; Laurent,
F.; Messner, F.; Hasan, A. F.; and Arvor, D. 2022. How
wildfires increase sensitivity of Amazon forests to droughts.
Environ. Res. Lett. 17(4):044031.
Mu˜noz-Sabater, J.; Dutra, E.; Agust ´ı-Panareda, A.; Al-
bergel, C.; Arduini, G.; Balsamo, G.; Boussetta, S.; Choulga,
M.; Harrigan, S.; Hersbach, H.; Martens, B.; Miralles, D. G.;Piles, M.; Rodr ´ıguez-Fern ´andez, N. J.; Zsoter, E.; Buon-
tempo, C.; and Th ´epaut, J.-N. 2021. ERA5-Land: a state-of-
the-art global reanalysis dataset for land applications. Earth
Syst. Sci. Data 13(9):4349–4383.
O’Neill, B. C.; Tebaldi, C.; Van Vuuren, D. P.; Eyring, V .;
Friedlingstein, P.; Hurtt, G.; Knutti, R.; Kriegler, E.; Lamar-
que, J.-F.; Lowe, J.; et al. 2016. The scenario model inter-
comparison project (scenariomip) for cmip6. Geoscientific
Model Development 9(9):3461–3482.
Poizner, S. 2022. Op-Ed: Wildfires never threatened my
home. But my insurer said they do — and dumped me. Los
Angeles Times .
Singh, A. G. 2022. The need to modernize california wild-
fire insurance regulation with climate science. Journal of
Science Policy and Governance 20(1).