arXiv:2312.04291v1 [physics.ao-ph] 7 Dec 2023

Simulating the Air Quality Impact of Prescribed Fires
Using a Graph Neural Network-Based PM2.5
Emissions Forecasting System

Kyleen Liao
Saratoga High School
CA, USA
kyleenliao@gmail.com

Jatan Buch
Columbia University
NY, USA
jb4625@columbia.edu

Kara Lamb
Columbia University
NY, USA
kl3231@columbia.edu

Pierre Gentine
Columbia University
NY, USA
pg2328@columbia.edu

Abstract
The increasing size and severity of wildfires across western North America have
generated dangerous levels of PM2.5 pollution in recent years. In a warming climate,
expanding the use of prescribed fires is widely considered to be the most robust fire
mitigation strategy. However, reliably forecasting the potential air quality impact
from these prescribed fires, a critical ingredient in determining the fires’ location
and time, at hourly to daily time scales remains a challenging problem. This paper
proposes a novel integration of prescribed fire simulation with a spatio-temporal
graph neural network-based PM2.5 forecasting model. The experiments in this
work focus on determining the optimal time for implementing prescribed fires in
California as well as quantifying the potential air quality trade-offs involved in
conducting more prescribed fires outside the fire season.

1

Introduction

Across many parts of the western United States (WUS), wildfire size, severity, and fire season length
have increased due to climate change. Wildfires across the WUS have led to the largest daily mean
PM2.5 (particulate matter < 2.5 microns) concentrations observed by ground-based sensors in recent
years [1], and exposure to PM2.5 is responsible for 4.2 million premature deaths worldwide per year
[2]. Within California, additional PM2.5 emissions from extreme wildfires over the past 8 years have
reversed nearly two decades of decline in ambient PM2.5 concentrations [3].
Prescribed fires, or controlled burns, have been widely accepted as an effective land management tool
and could have the potential to reduce the resulting smoke from future wildfires [4]. Since air quality
is a major public concern surrounding prescribed fires [5], land managers conducting these burns
require access to robust, near real-time predictions of downwind air pollution in order to determine
suitable locations and burn windows.
This paper introduces a graph neural network (GNN) model that incorporates satellite observations
of fire behavior in order to forecast PM2.5 emissions from ambient sources, observed fires, and
simulated controlled burns. Our GNN-based forecasting system can aid land managers in minimizing
the PM2.5 exposure of vulnerable populations during controlled burns and inform the public when
proposing prescribed fires.
37th Conference on Neural Information Processing Systems (NeurIPS 2023).

2

Related Work

Previous works studying the effect of prescribed fires used chemical transport models (CTMs) like
the Community Multiscale Air Quality (CMAQ) and Goddard Earth Observing System Atmospheric
Chemistry (GEOS-Chem) models to calculate the PM2.5 impact of prescribed fires at different
locations [4]. While CTMs can model the chemical processes in PM2.5 transport, generating accurate
predictions requires a large volume of information because of the complex chemical interactions.
Furthermore, the extensive calculations in CTMs make it challenging to explore a large range of
parameters for simulating prescribed burns [6, 7, 8]. In air quality predictions, machine learning
models have been shown to outperform CTMs in terms of accuracy and computational burden [9].
While several past works use machine learning to forecast air quality [10, 11], ours is the first research
paper to our knowledge that utilizes machine learning to predict the PM2.5 concentrations from
simulated prescribed fires.
Our research builds upon the GNN machine learning model from Wang et al. (2020) [11], which was
used to forecast non-wildfire-influenced PM2.5 pollution in China. In contrast, this work focuses on
simulating the effect of prescribed fires and predicting fire-influenced PM2.5 in California.

3

Methods

3.1

Dataset

Our dataset consists of PM2.5 , meteorological, and fire data at an hourly resolution over 5 years
(2017-2021). The PM2.5 concentration data, at a total of 112 air quality sensor locations in California,
is collected from both the California Air Resources Board as well as the Environmental Protection
Agency [12, 13]. The data for the 7 meteorological variables, which include u and v horizontal
components of wind, total precipitation, and air temperature, are retrieved from the ERA5 Reanalysis
database [14]. The full list of predictors is in Table 1. Though the meteorological variables may
capture the diurnal PM2.5 cycles and seasonal patterns, the Julian date and hour of the day are also
included as predictors in order to provide the model with additional context.
The fire radiative power (FRP) provides information about the fire intensity. The FRP at each fire
location is taken from the Visible Infrared Imaging Radiometer Suite (VIIRS) [15] instrument on
board the Suomi satellites. In order to assess the impact of nearby fires at the location of a PM2.5
monitor, we aggregate the FRP values of all active fires within radii of 25km, 50km, 100km, and
500km. To emphasize the fires that would likely have a more substantial downwind effect on PM2.5
concentrations, we use inverse distance weighting (IDW) and wind-based weighting in the FRP
aggregation. This process is described in more detail in Appendix A. The number of fires within
500km of a PM2.5 site is also included in the dataset. After compilation, the dataset’s missing values
were imputed using the MissForrest algorithm [16].
The prescribed fire data, retrieved from Cal Fire [17], is not represented as a variable in the training
dataset, but instead used in Experiments 1 and 2 when simulating prescribed fires.

Table 1: GNN Predictors
Predictor Name

Unit

Source

Planetary Boundary Layer Height (PBLH)
u-component of wind
v-component of wind
2m Temperature
Dewpoint temperature
Surface pressure
Total precipitation
WIDW FRP within 25km, 50km, 100km, 500km
Number of fires within 500km
Julian date
Time of day

m
m/s
m/s
K
K
Pa
m
MW
1
1
1

ERA5 Reanalysis
ERA5 Reanalysis
ERA5 Reanalysis
ERA5 Reanalysis
ERA5 Reanalysis
ERA5 Reanalysis
ERA5 Reanalysis
VIIRS
VIIRS
N/A
N/A

2

Table 2: Training/Validation/Testing Split

3.2

Training

Validation

Testing

1/1/2017 - 12/31/2018

1/1/2020 - 12/31/2020

1/1/2021 - 12/31/2021

Graph Neural Network (GNN)

We trained a spatio-temporal graph neural network (GNN) model based on Wang et al. (2020) [11]
to predict PM2.5 concentrations at an hourly temporal resolution. The GNN model is integrated
with a recurrent neural network (RNN) component, such that the model is able to capture both the
spatial and temporal propagation of PM2.5 . The model’s node features include the meteorological
and fire-related variables, while the edge attributes include the wind direction and speed at the source
node and the direction and distance between any two locations.
As shown in Table 2, for the GNN model, two years are used for training and one year each for
validation and testing. The year 2019 is excluded during training, validation, and testing because the
2019 fire season was an outlier and was less damaging than the other years. Validating and testing
the model on the years 2020 and 2021 respectively would help us gain a better understanding of the
model’s performance during intense fires. Our model produces forecasts for a prediction window of
48 hours into the future based on a historical window of 240 hours.
3.3

Prescribed Fire Simulations

The main contribution of this work is the novel simulation of the effect of prescribed fires in
conjunction with the GNN-based prediction of the resulting PM2.5 concentrations. This pipeline is
illustrated in Figure 1. The prescribed fires are simulated by transposing historical controlled burns to
target times, which are selected by matching the Cal Fire prescribed fire data with the VIIRS FRP
data. The transposed prescribed fire FRP information is combined with the observed meteorological
data at the target times and inputted into the GNN model, which produces the PM2.5 predictions.
Using this pipeline, we perform two model experiments. Experiment 1 demonstrates how the GNN
forecasting system can determine the optimal time to implement prescribed fires and focuses on
the short-term pollution effect of prescribed fires. Experiment 2, on the other hand, focuses on
quantifying the pollution impact of prescribed fires across months. In the rest of the section, we
discuss each experiment in more detail:

Figure 1: Prescribed Fire Simulation Pipeline

3.3.1

Experiment 1: Minimizing Prescribed Fire PM2.5 Impact

To determine the optimal time to implement prescribed fires, we consider the immediate effect of
prescribed fires. That is, we transpose the FRP values from actual prescribed fire events to target time
points and add them to the observed FRP values at those points. As these FRP values are combined,
they are aggregated using inverse distance and wind-based weighting, as outlined in Section 3.1 and
Appendix A.
In this experiment, we transpose a window of time containing prescribed fires (1/3/21 - 1/15/21) to
target times throughout the year 2021 at 24-hour time steps to simulate the air quality impacts of
controlled burns. This window contains ten prescribed fires with burned areas above 100 acres.
3

Table 3: Results of the GNN, LSTM, and MLP Models
MAE
RMSE

GNN

LSTM

MLP

5.23
6.72

5.73
7.32

6.24
7.83

Table 4: Results of PM2.5 Predictions Based On Simulated Prescribed Fires
Mean (µg/m3 )
Max (µg/m3 )

3.3.2

Jan

Feb

Mar

Apr

May

Jun

Jul

Aug

Sep

Oct

Nov

Dec

15.62
36.49

15.64
39.06

14.69
39.10

15.18
38.76

14.76
40.85

15.92
42.04

18.44
47.25

21.73
60.13

18.55
43.44

16.95
45.61

19.94
40.54

18.73
45.32

Experiment 2: Quantifying Prescribed Fire PM2.5 Trade-Off

This experiment aims to quantify the pollution trade-off of implementing prescribed fires by simulating
the effect of controlled burns in 2021 at the location of the Caldor fire. We employ two simulation
techniques: one corresponding to the immediate effect of prescribed fires, and another related to
simulating the longer-term effect of prescribed burning. For the former, we transpose historical
prescribed fires near the location of the Caldor fire to the spring of 2021. In the latter case, we
simulate the effect of controlled burns later in the year by excluding all FRP values within 25km of
the Caldor fire between 8/14/21 and 10/21/21, implicitly assuming that a prescribed fire implemented
earlier in the year (or even during the previous 1 to 2 fire seasons) could effectively mitigate a large
fire in the same location a few months later. The PM2.5 predictions from this counterfactual scenario
are compared to baseline predictions derived using observed meteorological and fire inputs from 2021
without any prescribed fires around the Caldor fire locations. For more details, see Appendix B.

4

Results

The GNN model resulted in the lowest mean absolute error (MAE) and root mean squared error
(RMSE) when compared with the long-short term memory (LSTM) and multilayer perception (MLP)
models, as shown in Table 3. As a reference for the error, very unhealthy and hazardous PM2.5 levels
are ≥ 150.5 µg/m3 . Experiments 1 and 2 use the GNN and are the main focus of the results section.
4.1

Experiment 1: Minimizing Prescribed Fire PM2.5 Impact

As shown in Table 4, the fall season, especially the month of August, was the least optimal time to
implement prescribed fires since it resulted in the most significant PM2.5 concentrations. As August
is during the peak wildfire season, implementing prescribed fires would only exacerbate the already
hazardous air quality. August’s mean was 29.61% greater than the average mean of other months,
and August’s maximum was 44.27% greater than the average maximum of other months. The mean
and maximum calculations are described in Appendix C. On the other hand, March, which had the
lowest mean value, seemed to be an optimal month to implement prescribed fires.
4.2

Experiment 2: Quantifying Prescribed Fire PM2.5 Trade-Off

The results support that, though prescribed fires may increase PM2.5 in the short term, the prescribed
fires reduce future PM2.5 resulting from wildfires. As shown in Table 5, the simulated prescribed
burns led the mean of the PM2.5 predictions to be increased by an average of 0.31 µg/m3 and the
maximum PM2.5 prediction to be increased by 3.07%. Details on the mean and maximum calculations
are included in Appendix C. Table 5 also quantifies that the maximum of the predictions with the
Caldor fire’s influence removed was 52.85% lower than the maximum of the baseline predictions.
Thus, the magnitude of the immediate PM2.5 increase from the prescribed fire was significantly
lower than the magnitude of the PM2.5 decrease experienced during the fire season. Furthermore,
excluding the influence of the Caldor Fire reduced the number of days with an unhealthy daily average
PM2.5 concentration from a mean of 3.54 days to 0.70 days. The reduction in PM2.5 pollution after
excluding the Caldor fire influence is illustrated in Figure 2, where the PM2.5 monitoring sites are
color-coded depending on the PM2.5 pollution’s US AQI level.
4

Table 5: Comparing the Predicted PM2.5 Effect of Simulated Prescribed Burns With Baseline PM2.5
Predictions
3/21/21 - 5/31/21
Simulated
Prescribed Burn
Mean (µg/m3 )
Max (µg/m3 )

6.83
55.78

8/14/21 - 10/21/21

Without Prescribed
Burn (Baseline)

Removed Caldor
Fire

With Caldor Fire
(Baseline)

6.52
54.12

10.49
83.61

16.24
177.32

(a) Removed Caldor Fire

(b) With Caldor Fire (Baseline)

Figure 2: Maximum PM2.5 predictions per site from 8/14/21 - 12/31/21 under condition (a) with
prescribed burn at the Caldor Fire location during the spring and without Caldor Fire during the
wildfire season and (b) without prescribed burn at the Caldor Fire location and with the Caldor Fire
during the wildfire season

5

Conclusion and Future Work

To our knowledge, this is the first research paper to apply machine learning for simulating the
PM2.5 impact of prescribed fires, which is significant as machine learning is less computationally
expensive than CTMs and requires lower expert curation of input variables. The primary contribution
of this work is the prescribed fire simulation pipeline, which integrates prescribed fire simulations
with GNN-based PM2.5 predictions. Future work will focus on improving the fire simulation by
incorporating physics-based modeling in the GNN framework. Our pipeline provides land managers
and the fire service with a useful tool to minimize the PM2.5 exposure of vulnerable populations,
while also informing local communities of potential air quality impacts as well as beneficial trade-offs
when implementing controlled burns.

6

Acknowledgements

We acknowledge funding from NSF through the Learning the Earth with Artificial Intelligence and
Physics (LEAP) Science and Technology Center (STC) (Award #2019625). Jatan Buch, Kara Lamb,
and Pierre Gentine were also supported in part by the Zegar Family Foundation.

5

References
[1] M. Burke, A. Driscoll, S. Heft-Neal, J. Xue, J. Burney, and M. Wara, “The changing risk and
burden of wildfire in the United States,” Proceedings of the National Academy of Sciences,
vol. 118, no. 2, 2021.
[2] World Health Organization, “Ambient (outdoor) air pollution,” 2022.
[3] M. Burke, M. Childs, B. de la Cuesta, et al., “The contribution of wildfire to PM2.5 trends in
the USA,” Nature, 2023.
[4] M. M. Kelp, M. C. Carroll, T. Liu, R. M. Yantosca, H. E. Hockenberry, and L. J. Mickley,
“Prescribed burns as a tool to mitigate future wildfire smoke exposure: Lessons for states and
rural environmental justice communities,” Earth’s Future, vol. 11, no. 6, 2023.
[5] S. M. McCaffrey, “Prescribed fire: What influences public approval,” in Fire in eastern oak
forests: delivering science to land managers, proceedings of a conference (M. B. Dickinson,
ed.), (Columbus, OH), pp. 192–198, U.S. Department of Agriculture, Forest Service, Northern
Research Station, 2006. Gen. Tech. Rep. NRS-P-1. Newtown Square, PA.
[6] M. H. Askariyeh, H. Khreis, and S. Vallamsundar, Air pollution monitoring and modeling,
p. 111–135. Elsevier, 2020.
[7] D. Byun and K. L. Schere, “Review of the governing equations, computational algorithms
and other components of the Models-3 Community Multiscale Air Quality (CMAQ) modeling
system,” Applied Mechanics Reviews, vol. 59, p. 51–76, 2006.
[8] N. Zaini, L. W. Ean, A. N. Ahmed, M. Abdul Malek, and M. F. Chow, “PM2.5 forecasting for
an urban area based on deep learning and decomposition method,” Scientific Reports, vol. 12,
no. 1, 2022.
[9] Y. Rybarczyk and R. Zalakeviciute, “Machine learning approaches for outdoor air quality
modelling: A systematic review,” Applied Sciences, 2018.
[10] L. Li, J. Wang, M. Franklin, et al., “Improving air quality assessment using physics-inspired
deep graph learning,” npj Climate and Atmospheric Science, vol. 6, p. 152, 2023.
[11] S. Wang, Y. Li, J. Zhang, Q. Meng, L. Meng, and F. Gao, “PM2.5 -GNN: A domain knowledge
enhanced graph neural network for PM2.5 forecasting,” in Proceedings of the 28th International
Conference on Advances in Geographic Information Systems, 2020.
[12] California Air Resources Board, “Air quality and meteorological information system.” https:
//www.arb.ca.gov/aqmis2/aqmis2.php, 2023. Accessed June 7, 2023.
[13] US Environmental Protection Agency, “Air quality system data mart.” https://www.epa.
gov/outdoor-air-quality-data, 2023. Accessed June 7, 2023.
[14] H. Hersbach, B. Bell, P. Berrisford, et al., “The ERA5 global reanalysis,” Quarterly Journal of
the Royal Meteorological Society, vol. 146, pp. 1999–2049, 2020.
[15] W. Schroeder, P. Oliva, L. Giglio, and I. A. Csiszar, “The new VIIRS 375 m active fire detection
data product: Algorithm description and initial assessment,” Remote Sensing of Environment,
vol. 143, pp. 85–96, 2014.
[16] D. J. Stekhoven and P. Bühlmann, “Missforest—non-parametric missing value imputation for
mixed-type data,” Bioinformatics, vol. 28, no. 1, pp. 112–118, 2011.
[17] CalFire, “Prescribed burns.” https://data.ca.gov/dataset/prescribed-burns, 2023.
Accessed June 20, 2023.

6

Figure 3: FRP Aggregation

A

Fire Radiative Power (FRP) Aggregation

When creating the dataset, for each PM2.5 monitor location, aggregations are performed on radii of
25km, 50km, 100km, and 500km to derive the wind and inverse-distance weighted (WIDW) FRP
using the process described in Figure 3 and Equation 1,
FWIDW =

n
X
Fi |Vi | cos (|αi |)

4πRi 2

i=1

(1)

where n is the number of fire locations within a certain radius of the PM2.5 monitor site, F is the
FRP value at the fire location, |V | is the magnitude of the wind speed at the fire location, α is the
relative angle between the wind direction and the direction from the fire to the PM2.5 monitor, and R
is the distance between the fire site and PM2.5 monitor.

B

Additional Details on Experiment 2 Methodology

This experiment simulates the effect of three prescribed fires, which were all within 20km of the 2021
Caldor fire, were active from 3/21 - 5/31 in 2018, 2019, and 2020 respectively, and burned around
6,300 acres each. Since the Caldor fire burned around 221,835 acres, we assume that preventing a
fire of that scale would require a larger controlled burn. Thus, the FRP values from the prescribed
fires are both artificially increased by a factor of 100 and transposed together to 2021, emulating a
large prescribed fire from 3/21/21 to 5/31/21. As described in Section 3.3.1, the prescribed fires are
transposed by combining the FRP values of the prescribed fires with the observed FRP values at the
target time point, and then by aggregating those values using inverse distance weighting and the wind
information at the target time points.
As mentioned in Section 3.3.2, for all time points after 5/31/21, the effect of prescribed burns later
in the year are simulated by excluding all FRP values within 25km of the Caldor fire location. To
further remove the Caldor fire influence, PM2.5 values from 2018 are used as inputs instead of the
Caldor-influenced, observed PM2.5 values from 2021. The 2018 PM2.5 data is chosen because, in
comparison to the other years, the 2018 fire season most closely resembles the 2021 fire activity
without fires at the Caldor fire location.

C

Mean and Maximum Calculation Details

For Experiment 1, the mean and maximum values are calculated by averaging the mean and maximum
PM2.5 predictions of the locations whose PM2.5 observations were ≥ 50 µg/m3 during the window
1/3/21 - 1/15/21. As the PM2.5 observations at those locations are elevated during 1/3/21 - 1/15/21,
they are likely influenced by the fire events that were transposed across the year 2021.
For Experiment 2, the mean and maximum values are calculated by averaging the mean and maximum
PM2.5 predictions of the 13 PM2.5 monitor locations within 100km of the Caldor fire.
7