=Paper=
{{Paper
|id=Vol-2884/paper_121
|storemode=property
|title=Uncertainty Aware Wildfire Management
|pdfUrl=https://ceur-ws.org/Vol-2884/paper_121.pdf
|volume=Vol-2884
|authors=Tina Diao,Samriddhi Singla,Ayan Mukhopadhyay,Ahmed Eldawy,Ross Shachter,Mykel Kochenderfer
}}
==Uncertainty Aware Wildfire Management==
https://ceur-ws.org/Vol-2884/paper_121.pdf
Uncertainty Aware Wildfire Management
Tina Diao,1 Samriddhi Singla,2 Ayan Mukhopadhyay,1
Ahmed Eldawy,2 Ross Shachter, 1 Mykel Kochenderfer 1
1
Stanford University,
2
University of California, Riverside
1
{tdiao, ayanmukh, shachter, mykel}@stanford.edu, 2 {ssing068, eldawy}@ucr.edu
Abstract particularly susceptible to wildfires. Known to be notori-
ously unpredictable, sudden changes in wind directions or
Recent wildfires in the United States have caused many lives
and billions of dollars, destroying countless structures and weather can change the way fires spread (Garza 2020). Fire-
forests. Fighting wildfires is extremely complex. It is difficult fighters need to allocate limited resources in dynamic and
to observe the true state of fires due to smoke and risk asso- uncertain environments to intervene and stop the spread of
ciated with ground surveillance. There are limited resources fire.
to be deployed over a massive area and the spread of the The problem of combating wildfires is an example of dy-
fire is challenging to predict. This paper proposes a decision- namic resource allocation. Such an approach to combat natu-
theoretic approach to combat wildfires. We model the re- ral calamities like wildfires, floods, and earthquakes are dif-
source allocation problem as a partially observable Markov ficult for several reasons. The dynamics of the events are
decision process. We also present a data-driven model that
lets us simulate how fires spread as a function of relevant
unknown and complicated to model in closed-form. There
covariates. A major problem in using data-driven models to are multiple covariates that affect the spread of fire like
combat wildfires is the lack of comprehensive data sources the vegetation, fuel, altitude, and wind, and the exact re-
that relate fires with relevant covariates. We present an algo- lationship between each covariate and fire spread is uncer-
rithmic approach based on large-scale raster and vector anal- tain. Resource allocation is extremely challenging because
ysis that can be used to create such a dataset. Our data with of the difficulty of making predictions about the spread of
over 2 million data points is the first open-source dataset that fire and expected damage. Designing principled approaches
combines existing fire databases with covariates extracted to deploy resources is important to mitigate the effects of
from satellite imagery. Through experiments using real-world wildfires and emergencies in general (Mukhopadhyay et al.
wildfire data, we demonstrate that our forecasting model can 2020).
accurately model the spread of wildfires. Finally, we use sim-
ulations to demonstrate that our response strategy can signif- Events like fires affect large areas, and it is difficult to
icantly reduce response times compared to baseline methods. correlate large-scale data from various sources to analyze
and study fires. To the best of our knowledge, there exists
no comprehensive data source that combines fire occurrence
Introduction with geospatial features, fuel levels, and weather to allow the
In 2018, a large wildfire (named Camp Fire) in California re- research community to develop approaches to combat wild-
sulted in the loss of 88 lives, displaced countless more, and fires. Finally, the problem of deploying resources to manage
destroyed more than 18,500 structures. The estimated cost wildfires is full of uncertainties. As fires spread, it becomes
of the destruction was a staggering $15 billion. As we write increasingly difficult to observe the true state on the ground
this manuscript at the start of the wildfire season in the state, due to the presence of smoke. As a result, first responders
more than a million acres have already burned in California only have imperfect information to allocate their resources.
this year alone due to more than 7,000 wildfires. Wildfires Contributions: We make the following contributions in
have destroyed many towns and structures across the state. this paper. 1) We present a decision-theoretic approach to
At one point in August 2020, the entire northern half of the dynamically allocate resources in uncertain environments to
state had been instructed to prepare for evacuation (State of intervene against wildfires. Specifically, we model the re-
California 2020). Crucially, the time of the year that is re- sponse problem as a partially observable Markov decision
ferred to as the “wildfire season” in the state has only just process (POMDP) and present an approach to find opti-
begun. mal actions for suppression for a given state of the prob-
Fighting wildfires is difficult. Rapid urbanization and the lem. Our approach accommodates the constraint that the
effects of climate change make urban and suburban areas true state of the fire is difficult to observe in practice. 2)
AAAI Fall 2020 Symposium on AI for Social Good. Instead of focusing on physics-based models, we present a
Copyright c 2020 for this paper by its authors. Use permitted un- data-driven approach that can simulate the spread of wild-
der Creative Commons License Attribution 4.0 International (CC fires. We extract relevant covariates such as fuel levels, veg-
BY 4.0). etation type, height of canopy, and elevation from satellite
imagery to drive the simulation. 3) The calculation of accu- is divided into T time steps. We assume access to histor-
rate zonal statistics is a bottleneck for large scale geospa- ical data of fire incidents D, which is a vector of tuples
tial data analysis (Singla and Eldawy 2018). We demon- {(t1 , `1 , u1 , w1 ), (t2 , `2 , u2 , w2 ), . . . , (tn , `n , un , wn )},
strate how extremely large-scale geospatial data pertaining where each incident di ∈ D is identified by its time of
to wildfires can be combined with other covariates through occurrence ti , location `i (mapping to a cell in G), intensity
large-scale distributed raster and vector analysis. Crucially, R
of fire observed ui ∈ + , and a vector of spatio-temporal
we release the dataset open-source for the research com- R
features wi ∈ m . The features w capture potential deter-
munity. 4) Through experiments using real-world wildfire minants of fire such as weather and the type of vegetation in
data from California, we demonstrate that our forecasting a cell.
model can accurately model the spread of wildfires and our Observing the true state of the world is almost always
response strategy results in significant improvement in sup- impossible when wildfires occur. Gathering real-time in-
pression efforts compared to baseline methods that do not formation by visiting the affected areas by land is nat-
consider potential fire spread. urally difficult. Therefore, information must be gathered
through air surveillance obstructed by smoke. Consequently,
Prior Work the true dynamics of wildfires are only partially observ-
able. We model the fire suppression problem more real-
The dynamics of fire spread are usually modeled us-
istically as a partially observable Markov decision pro-
ing physics-based models. Popular fire spread models in-
cess (POMDP). A POMDP can be defined by the tuple
clude BehavePlus (Andrews 1986) and Farsite (Finney
{S, A, O, Z, T, R, γ} (Kochenderfer 2015). We define each
1998). They are based on mathematically modeled surface
component of the POMDP formulation below:
fire spread as a function of heat flux and fuel availabil-
ity (Rothermel 1972). Such models are widely used by first 1. States: S is a finite set of states. The state at time t is de-
responders and fire fighters to forecast the spread of fires. A noted by st = {Xt , Ft }, where Xt = {Xt1 , Xt2 , . . . , Xtk }
relatively modern approach is to predict the rate of spread by and Ft = {Ft1 , Ft2 , . . . , Ftk } denote the status of the fire
integrating real-time information about weather from sen- and fuel level in each of the cells in G, respectively. We
sors (Altintas et al. 2015). Data-driven modeling has also consider the status of the fire X as a binary variable such
been used to model fire spread. Supervised machine learn- that
ing techniques have been applied to uncover strong associ-
1 if uit ≥
i
ations of factors to wildfire sizes and frequency using dif- xt =
ferent data sources. For example, Joseph et al. (2019) in- 0 otherwise
vestigated weather conditions and geographic characteris-
tics of extreme fire patterns in the contiguous United States
R
where uit ∈ + denotes the measured intensity of the fire
in cell gi ∈ G at time step t. We consider that the fuel level
and Ghorbanzadeh et al. (2019) examined wildfire suscepti- F is a discrete variable, such that F ∈ {0, 1, . . . , m}. The
bility using geographic data in northern Iran. exogenous parameters and m can be estimated from data
Response to wildfires traditionally uses simulation-based or through domain knowledge.
approaches to select locations of intervention that maximize
the expected utility of suppression efforts. Petrovic, Alder- 2. Actions: A is a finite set of actions. The actions denote the
son, and Carlson (2012) model wildfire dynamics and ex- different permutations of cell indices that fire suppression
amine the trade-off between multiple competing suppres- efforts can be applied to, up to a maximum number of
sion efforts to compute an optimal strategy for fire re- cells specified as a resource constraint.
sponses. Stochastic simulation and multi-agent coordination 3. State Transitions: T defines conditional transition proba-
has also been explored to combat wildfires (Fried, Gilless, bilities, with T (s0 | s, a) denoting the transition probabil-
and Spero 2006; Martin-fernÁndez, Martı́nez-Falero, and ity from state s to s0 when action a is taken. The transition
Pérez-González 2002). model T includes the following three components:
Griffith et al. (2017) explore how suppression efforts can
be optimized by solving a mixed-integer linear program and (a) Burning: If cell gi ∈ G is burning at time step t, we
by using Monte Carlo approaches to find optimal actions in a assume that the fuel level decreases by one unit in the
Markov decision process (MDP). Our approach to optimize next time step. Therefore, Ft+1i
= Fti − 1 if xit ≥ .
suppression improves upon prior work (Griffith et al. 2017) (b) Action effectiveness: At any time step t, if an action,
to address the uncertainty in state information. We also inte- i.e. fire suppression effort(s) is applied to a cell gi ∈ G
grate a data-driven generative model to simulate the spread such that xit ≥ , there is probability q that the effort
of fire to aid decision-making under uncertainty. successfully puts out the fire.
(c) Fire dynamics: We use a generative model to simu-
Problem Description late the spread of fire. We represent spread dynamics
We consider a spatial area divided into a set of spatial cells by the probability distribution f (Xt+1i
| xjt , w), where
G. Let gi ∈ G denote the ith cell. We represent the neigh- gi ∈ Nj . Therefore, given that a specific cell gj ∈ G
bors of a cell gi by Ni , for some definition of neighborhood is on fire at time step t, f denotes the likelihood of its
(for example, neighbors of a cell can be the set of its neighboring cells being on fire at the subsequent time
adjacent cells). Consider that the total time in consideration step t + 1.
We assume that in each time step, fire from a cell can
spread only to its neighboring cells. This assumption aids
computational tractability, but is without loss of gener-
ality because the decision-maker can discretize time fine
enough such that the assumption is realistic.
4. Reward function: R : S ×P A → R is the reward func-
tion, such that R(st , a) = gi ∈G xit U (gi ), where U (gi )
denotes the utility for a cell gi ∈ G to be on fire. Nat-
urally, U varies across the cells. A cell with human oc-
cupants is presumably more valuable and more costly to
burn than a cell composed of forested land. Cells can also
carry unequal ecological utilities (Bradshaw and Lueck
2012). Without loss of generality, we consider three tiers
of damage across cells, representing costs of residences Figure 1: An example of a spatial area discretized in a grid
(red), valuable ecological resource (yellow), and wildland with different costs and the colors correspond to varying
(green) in decreasing values. Figure 1 shows an example costs to burn: red, yellow, and green in decreasing cost.
grid with different types of cells.
5. Observations and Observation Transitions: O represents
Bt−1 Bt Bt+1
the set of observations with Z(o | s, a) denoting the prob-
ability of receiving observation o at state s when action
ot−1 ot
a is taken. We denote specific observations by oit , which
correspond to whether cell gi ∈ G is seen to be burning or
not at time step t. Its transitions Z(o | s, a) are determin- St−1 St St+1
istic, i.e. oit,a = uit if an action has been applied to cell
gi ∈ G. This is because we assume to have “eyes on lo- at−1 at
cation” when an action is applied to a cell. Otherwise, for
cells where action for suppression is not applied (denoted
Rt−1 Rt
by ā in the expression below), a generative representa-
tion is used based on prior work (Julian and Kochenderfer
2019) such that
Figure 2: POMDP (factored) representation of the dynamic
1 if P rt (Xti ) > η in state s decision problem. Shaded ovals reflect what is known to the
i
ot,ā = decision maker.
0 otherwise
where η is an exogenous parameter.
6. Discount factor: γ ∈ [0, 1] denotes the discount factor. We use random forests (Liaw, Wiener et al. 2002) which
involves constructing a large number of decision trees at
In a POMDP, the decision-maker cannot directly observe training time and then aggregating the outputs of the trees.
the state. Instead, they only have access to beliefs that are The aggregation technique is typically using the mode of
generated probabilistically based on the actions taken. In- the outputs for classification and the mean of the outputs for
formation about states can be inferred from the history (h) regression. The central idea behind using random forests is
of observations and actions. It is common to maintain a dis- to average many noisy but (approximately) unbiased mod-
tribution over states given the history; this distribution is els, thereby reducing the variance of the overall forecasting
known as the belief B, such that B(s | h) denotes the prob- model.
ability of being in state s given history h. The goal for the
decision-maker is to find a mapping from belief states to ac- Resource Allocation
tions that maximizes the expected future discounted reward.
The general dynamic decision framework following the
Approach POMDP formulation is shown in Figure 2. Recall that the
sources of uncertainty are the current state of the fire (S)
Modeling Fire Spread and the dynamics of the spread (driven by the set of co-
In order to accurately simulate the spread of fire, we model variates w). At each time step, the available set of actions
the fire dynamics f using a data-driven model. Recall that f (A) is a combination of the fire location(s) to suppress. The
is a probability distribution over a cell being on fire in time decision-maker acts based on a belief distribution of the true
step t + 1, conditional on its neighbors being on fire at the state of the world and receives an observation. Utility (U) re-
previous time step t. The goal of modeling the function f is flects a measure of the expected damage of having a fire in a
to understand the effect of various covariates like wind, veg- specific cell (e.g. residence versus wildland).
etation type, canopy height, altitude, etc. on fire spread. Typ- Due to the large state space and action space (the ac-
ically, covariates in geospatial analysis are heterogeneous. tion space is combinatorial), we use the sampling-based on-
line Monte Carlo tree search (MCTS) (Kochenderfer 2015). Data Processing
To address the computational complexity of POMDPs, ap- Data needed to model the spread of wildfires comes from
proaches based on MCTS typically use a particle filter to varied sources. The temporal and spatial resolutions of such
represent beliefs in the search tree. Specifically, an approach data sources are typically different. The data sources can
that is of relevance to our problem is the partially observable also be of different forms (vector or raster). The vector
Monte Carlo planning algorithm with observation widen- model uses points and line segments to identify spatial lo-
ing (POMCPOW) (Sunberg and Kochenderfer 2017). POM- cations while the raster model uses a set of cells for the
CPOW differs from other online MCTS algorithms in that in same purpose. Combining large-scale vector and raster data
the simulation step, given a state (s), history (h), and depth is known to be a difficult problem (Singla and Eldawy 2018).
(for tree exploration), it weights the belief nodes and ex- We collected fire occurrence data in vector form from the
pands the belief updates gradually as more simulations are Visible Infrared Imaging Radiometer Suite (VIIRS) thermal
added. At each step, a single simulated new state is added to anomalies/active fire database (Schroeder et al. 2014). The
the particle collection, weighted to approximate the belief in spatial resolution of VIIRS data is in the form of pixels rep-
every tree node, which is then used to sample the next new resenting 375 × 375 meter square cells (Schroeder et al.
state. 2014). The latitude and longitude values correspond to the
Although POMCPOW is shown to outperform other algo- center. Evidence of fire was read from the daily fire radia-
rithms (Sunberg and Kochenderfer 2017), an issue with di- tive power (FRP) levels in the VIIRS dataset. The data used
rectly using POMCPOW on our problem is that the observa- to build the feature space was collected in raster form from
tion space is large and complex. This leads to a severe spar- the LANDFIRE project (Ryan and Opperman 2013). The
sity of particles, i.e., the probability of sampling a relevant foundation of the LANDFIRE project is based on satellite
observation is very small. To alleviate this, we modify the imagery. The raster files had a spatial resolution of 30 × 30
routine used to update belief in POMCPOW. Specifically, meter square cells. This included features like canopy base
we replace the weighted particle filter with the standard par- density, canopy cover, and vegetation type. We list features
ticle filter without rejection (Kochenderfer 2015). As shown used and the years from which the data was collected in Ta-
in Algorithm 1, given a current belief (b), action (a), and ble 1.
observation o, |b| samples are generated from the simulator,
weighted, and subsequently resampled by its weights. Be- Name Year(s)
fore the resampling step, all weights of an observation o may Canopy Base Density 2012, 2014, 2016
be 0 due to the large observation space. If all the weights of Canopy Base Height 2012, 2014, 2016
an observation are 0, the probabilities of the sampled states Canopy Cover 2012, 2014, 2016
are normalized to be proportional to the number of states al- Canopy Height 2012, 2014, 2016
ready sampled. This reweighting step makes an approxima- Existing Vegetation Cover 2012, 2014, 2016
tion to importance resampling that seeks to estimate prop- Existing Vegetation Height 2012, 2014, 2016
erties of a target distribution by sampling from a different Existing Vegetation Type 2012, 2014, 2016
distribution. Elevation 2016
Slope 2016
Algorithm 1: UpdateBelief (b, a, o)
Table 1: LANDFIRE Raster Data
Input : Belief b, action a, observation o
Output: Updated belief b0
To reconcile the different spatial resolutions, we divide
1 b0 ← ∅ the state of California into a grid G of 375 × 375 meter
2 for i ← {1, . . . , | b |} do cells. The center of each fire pixel from the vector data can
3 si ← random state in b therefore overlap with exactly one cell in G. To compute
4 s0i ∼ G(si , a) the feature vector associated with each data point, we com-
5 wi ← O(o | s0i , a) pute zonal statistics for the vector data using the raster data.
6 end The method of zonal statistics refers to calculating summary
P|b| statistics using a raster dataset within zones defined by an-
7 if i=1 wi = 0 then
8 wi = len(w)
1
; // reweight states other dataset (typically in vector form).
Traditional systems to compute zonal statistics require the
9 end
data to be converted into the same format, either raster or
10 for i ← {1, . . . , | b |} do
vector. Automated tools can then be used to either vectorize
11 Randomly select k with probability proportional to the raster dataset or rasterize the vector dataset. The first ap-
wk proach converts each pixel in the raster to a point and then
12 Add s0k to b0 tests the point against each polygon in the vector data to find
13 end a match. This approach has a computational complexity of
14 return normalized b
0
O(np log np · c · r), where np is the number of polygons in
the vector data, and c and r are the number of columns and
rows in the raster data respectively. The second approach
Cluster
Vector Data Distribute
Intersections
Files
Intersections
Raster Data
Computation
Input Zonal Statistics
Computation
Figure 3: Architecture for calculating zonal statistics for large-scale raster and vector data
rasterizes the vector data by converting each polygon to a neighboring cell of gi at time step t + 1. To calculate the
raster (mask) layer with the same resolution as the input features for a specific cell gi ∈ G at a specific time step,
raster layer. It then combines the two raster layers to com- we calculated summary statistics (maximum, minimum, me-
pute the desired aggregate function. Most systems that use dian, sum, mode, count, and mean of the feature values) us-
this approach keep the mask layer in memory. If the input ing all raster cells within gi at time step t. To the best of our
raster layer has a very high resolution, the size of the mask knowledge, our data is the first comprehensive open-source
layer can become too large to be kept in memory. This ap- dataset that combines fire occurrence with relevant covari-
proach has a computational complexity of O(np · c · r). ates extracted from satellite imagery. The data is available at
Neither of the two approaches mentioned above scale for https://wildfire-modeling.github.io. We used data from 2012
the high resolution raster and vector data due to the re- to 2017 as our training set and data from 2018 as our test set.
quirement of converting between vector and raster formats. We set the time step for our experiments to a day, based on
Specifically, in our problem, the vector data consist of over 3 the minimum time fidelity of the VIIRS dataset. All exper-
million polygons and each of the raster data sources consists iments were run on an Intel Xeon 2.2 GHz processor with
of over 1 billion entries. To deal with the large-scale geospa- 125 GB of memory.
tial data, we create a fully decentralized approach to com-
pute large-scale zonal statistics based on Singla and Eldawy Fire Spread
(2020). Our approach does not require data to be converted We label a forecast as a true positive prediction when both
from one form to another (vector or raster). Instead, it com- the predicted fire intensity and the recorded fire intensity are
putes an intermediate data structure, called an intersections greater than the pre-specified threshold . We observe that
file between the raster and vector data. The intersections file the random forest regression model is insensitive to the num-
can be computed using only the vector data and the meta- ber of trees used (5, 50, 100, and 500). We also observed
data of the raster data (coordinate reference, resolution, etc.). similar accuracy across training and test sets. We tested sev-
Further, our approach can leverage parallel computation by eral realizations of to examine the robustness of our fore-
using the intersection files. Our approach, with a computa- casting approach on different sizes of fire. Our results show
tional complexity of O(np log np + c · r), is scalable and that while it is relatively difficult to predict spread from ex-
efficient for large raster and vector datasets. Furthermore, as tremely small fires ( = 0.5), our forecasting model achieves
a by-product, our approach makes it easier to find the neigh- high accuracy (> 90%) in predicting spreads from relatively
borhoods of polygons by performing a spatial self-join op- larger fires. We summarize the results in Table 2.
eration using the predicate intersects on all the cells in the
vector data. We show the architecture of our approach in FRP Threshold () Accuracy on Test Set
Figure 3.
0.5 77%
1 81%
Experiments 5 92%
We used fire data from California, USA spanning 2012–
2018 for the prediction modeling. We divided the state into Table 2: Accuracy with 5-tree Random Forest Regression
a set of 375 × 375 meter cells. Our goal was to capture how
fire spreads given an initial occurrence of fire. As a result, we
only considered cells and days that exhibited the possibility Fire Response
of fire spreading from an existing fire. Specifically, each cell Our goal is to create a pipeline for modeling response to
in our data has a fire in its neighborhood. Our data has a total wildfires by utilizing the data-driven model of fire spread
of 2,367,209 data points. Each row in our data represents a to optimize response decisions. However, in this paper,
cell gi at a specific time t, a set of spatial-temporal features we present experimental results based on a simple spread
wit , and the status of fire uti as well as the spatial-temporal model. Specifically, we simulate the spread of fire using
features wjt+1 and the status of fire ut+1 j of gj ∈ Ni , a fixed environmental conditions (for example, we use a fixed
Baseline UAFR 1,800 Baseline UAFR Baseline UAFR
700
Negative Utility
Negative Utility
Negative Utility
3,000
1,600
600 2,500
1,400
q=1 q=0.9 q=0.8 q=1 q=0.9 q=0.8 q=1 q=0.9 q=0.8
(a) (b) (c)
·104
6,000 8,000
Baseline UAFR Baseline UAFR Baseline UAFR
Negative Utility
Negative Utility
Negative Utility
1.4
5,500
7,000
5,000 1.2
4,500 6,000
q=1 q=0.9 q=0.8 q=1 q=0.9 q=0.8 q=1 q=0.9 q=0.8
(d) (e) (f)
Figure 4: Effectiveness measured by negative utility on test data (lower is better) on varying grid sizes: (a) 4 × 4 (b) 6 × 6 (c)
8 × 8 (d) 10 × 10 (e) 12 × 12 and (f) 16 × 16.
wind direction and rate of spread in each simulation). We combat due to difficulties in surveillance and scarcity of
calculate the probability of fire spread following a simpli- resources. In this paper, we built a data-driven forecast-
fied version of the deterministic fire dynamics model in prior ing model by extracting relevant determinants of fire spread
work (Griffith et al. 2017). through satellite imagery. Then, we developed an approach
We conducted experiments on different sized square grids to wildfire suppression that explicitly takes state uncertainty
(42 , 62 , 82 , 102 , 122 and 162 ). We varied the starting states into account. To the best of our knowledge, we were the first
by random initialization of fire maps and cell types. We to create a comprehensive dataset on wildfires that combines
experimented with 6 different initial states. For each ini- historical fire data with relevant covariates. Our dataset, with
tial state, we created 256 spread scenarios by varying wind over two million data points, and our codebase are open-
and rate of spread. For each initialization, a fixed percent- source for the research community to use.
age (10%) of cells was considered to be on fire. The pro- There are some limitations to our current work. While our
portions of red, yellow, and green cells were set to be 20%, forecasting model shows high accuracy, we observe that the
30%, and 50% respectively, consistent with the distribution random forest regression model is insensitive to the number
of land use in California (United States Department of Agri- of decision trees in the ensemble. As a result, simpler meth-
culture 2016). In all experiments, we set the fuel level in ods like classification and regression trees (CART) (Breiman
each cell to 5, i.e., each cell takes 5 days to completely burn et al. 1984) might result in better generalization to unseen
out and deplete its fuel. We use a baseline model consistent data. Second, our approach to suppression needs to be inte-
with fire suppression strategies used in practice. Specifically, grated with the data-driven fire spread model. Finally, while
our baseline fire suppression targets the cell with the max- we simulated wind for our experiments, our open-source
imum utility that shows evidence of fire (through observa- dataset does not contain information about wind. We are
tions). Our open-source code is built using the POMDPs.jl currently incorporating hourly wind data from the National
framework (Egorov et al. 2017) and is available at https: Oceanic and Atmospheric Administration (NOAA)1 with
//github.com/wildfire-modeling/response model. our fire data to develop a more comprehensive dataset.
Figure 4 shows the mean negative utility averaged across
start states and spread scenarios. We also vary the action ef-
fectiveness q since in practice, suppression efforts do not Acknowledgements
always completely puts out wildfires. We see that our ap-
proach (referred to as “Uncertainty Aware Fire Response” We would like to acknowledge the Department of Manage-
or UAFR) consistently outperforms the baseline method in ment Science & Engineering at Stanford University, USDA
all scenarios with more significant improvements in larger NIFA AFRI (Grant 2019-67022-29696), and the Center of
grids and increased action effectiveness. Automotive Research at Stanford (CARS) for funding this
research.
Discussion
Wildfires have caused massive damage in the last decade. 1
https://www.climate.gov/maps-data/dataset/wind-roses-
They are particularly challenging for first responders to charts-and-tabular-data
References Mukhopadhyay, A.; Pettet, G.; Vazirizade, S.; Lu, D.;
Altintas, I.; Block, J.; De Callafon, R.; Crawl, D.; Cowart, Baroud, H.; Jaimes, A.; Vorobeychik, Y.; Kochenderfer, M.;
C.; Gupta, A.; Nguyen, M.; Braun, H.-W.; Schulze, J.; Goll- and Dubey, A. 2020. A Review of Emergency Incident Pre-
ner, M.; et al. 2015. Towards an integrated cyberinfrastruc- diction, Resource Allocation and Dispatch Models.
ture for scalable data-driven monitoring, dynamic prediction Petrovic, N.; Alderson, D. L.; and Carlson, J. M. 2012. Dy-
and resilience of wildfires. Procedia Computer Science 51: namic resource allocation in disaster response: Tradeoffs in
1633–1642. wildfire suppression. PloS one 7(4): e33285.
Andrews, P. L. 1986. BEHAVE: fire behavior prediction and Rothermel, R. C. 1972. A mathematical model for predicting
fuel modeling system: BURN subsystem, Part 1, volume 194. fire spread in wildland fuels, volume 115. Intermountain
US Department of Agriculture, Forest Service, Intermoun- Forest & Range Experiment Station, Forest Service.
tain Research Station.
Ryan, K. C.; and Opperman, T. S. 2013. LANDFIRE–A
Bradshaw, K.; and Lueck, D. 2012. Wildfire policy: Law and national vegetation/fuels data base for use in fuels treatment,
Economics Perspectives. Routledge. restoration, and suppression planning. Forest Ecology and
Breiman, L.; Friedman, J.; Stone, C. J.; and Olshen, R. A. Management 294: 208–216.
1984. Classification and regression trees. CRC Press. Schroeder, W.; Oliva, P.; Giglio, L.; and Csiszar, I. A. 2014.
Egorov, M.; Sunberg, Z. N.; Balaban, E.; Wheeler, T. A.; The New VIIRS 375 m active fire detection data product: Al-
Gupta, J. K.; and Kochenderfer, M. J. 2017. POMDPs.jl: gorithm description and initial assessment. Remote Sensing
A Framework for Sequential Decision Making under Uncer- of Environment 143: 85–96.
tainty. Journal of Machine Learning Research 18(26): 1–5. Singla, S.; and Eldawy, A. 2018. Distributed zonal statistics
URL http://jmlr.org/papers/v18/16-300.html. of big raster and vector data. In Proceedings of the 26th
Finney, M. A. 1998. FARSITE, Fire Area Simulator–model ACM SIGSPATIAL International Conference on Advances
development and evaluation. 4. US Department of Agricul- in Geographic Information Systems, 536–539.
ture, Forest Service, Rocky Mountain Research Station. Singla, S.; and Eldawy, A. 2020. Raptor Zonal Statistics
Fried, J. S.; Gilless, J. K.; and Spero, J. 2006. Analysing : Fully Distributed Zonal Statistics of Big Raster + Vector
initial attack on wildland fires using stochastic simulation. Data. In Proceedings of 2020 IEEE International Confer-
International Journal of Wildland Fire 15(1): 137–146. ence on Big Data (To Appear). IEEE.
Garza, A. 2020. AI Is Helping Fight Wildfires Before State of California. 2020. 2020 Incident Archive. https:
They Start. https://time.com/5497251/wildfires-artificial- //www.fire.ca.gov/incidents/2020/.
intelligence/. Sunberg, Z.; and Kochenderfer, M. 2017. Online algorithms
Ghorbanzadeh, O.; Valizadeh Kamran, K.; Blaschke, T.; for POMDPs with continuous state, action, and observation
Aryal, J.; Naboureh, A.; Einali, J.; and Bian, J. 2019. Spa- spaces. arXiv preprint arXiv:1709.06196 .
tial prediction of wildfire susceptibility using field survey United States Department of Agriculture. 2016. FIA State
gps data and machine learning approaches. Fire 2(3): 43. Stats. https://www.fs.fed.us/pnw/rma/fia-topics/state-stats/
Griffith, J. D.; Kochenderfer, M. J.; Moss, R. J.; Mišić, V. V.; California/index.php.
Gupta, V.; and Bertsimas, D. 2017. Automated dynamic re-
source allocation for wildfire suppression. Lincoln Labora-
tory Journal 22(2).
Joseph, M. B.; Rossi, M. W.; Mietkiewicz, N. P.; Mahood,
A. L.; Cattau, M. E.; St. Denis, L. A.; Nagy, R. C.; Iglesias,
V.; Abatzoglou, J. T.; and Balch, J. K. 2019. Spatiotempo-
ral prediction of wildfire size extremes with Bayesian finite
sample maxima. Ecological Applications 29(6).
Julian, K. D.; and Kochenderfer, M. J. 2019. Distributed
wildfire surveillance with autonomous aircraft using deep
reinforcement learning. Journal of Guidance, Control, and
Dynamics 42(8): 1768–1778.
Kochenderfer, M. J. 2015. Decision Making Under Uncer-
tainty: Theory and Application. MIT Press.
Liaw, A.; Wiener, M.; et al. 2002. Classification and regres-
sion by randomForest. R news 2(3): 18–22.
Martin-fernÁndez, S.; Martı́nez-Falero, E.; and Pérez-
González, J. M. 2002. Optimization of the resources man-
agement in fighting wildfires. Environmental Management
30(3): 352–364.