Predicting Daily Wildfire Emissions Using Machine Learning Techniques

By Laura Thapa, Department of Atmospheric and Oceanic Sciences, UCLA

Problem

Wildfire activity has increased dramatically over the past decade in the Western United States, with sixteen of California’s twenty most destructive wildfires occurring since 2016. Five of these fires occurred in 2020 alone, and current estimates put the 2020 burn area at greater than four million acres (Cal Fire). Wildfire smoke is known to negatively impact human health, climate and visibility (e.g., Wegesser et al 2009, Spracklen et al 2007). Accurate air quality forecasts can help mitigate some of these impacts yet forecasting episodes wildfire smoke remains difficult. One of the reasons for forecast failures is the way wildfire emissions are predicted.

Wildfire emissions are the gasses and particles released during combustion. The top five species are: carbon dioxide (CO2, 71.44%), water, (H2O, 20.97%), carbon monoxide (CO, 5.52%), particulate matter <2.5 microns in diameter (PM2.5, 0.47%), and nitric oxide (NO, 0.39%) (Liu et al, 2013). Smoke forecasting systems generally predict future fire emissions by persistence, copying emissions from the previous day into the next day. This is a good assumption if the fire behavior is not changing from day to day. However, wildfires are dynamic systems driven by meteorology, fuel availability, terrain, and human activity. Therefore, fires may experience both large and small growth days, which can cause the persistence assumption to fail. Figure 1 shows persistence failing during the 2018 Carr Fire.

In this article, we present two data-driven methods to predict daily wildfire emissions of organic carbon (OC) aerosol. We tackle the problem of forecasting future behavior of active fires in contrast to predicting fire ignition and suppression. Our domain of study is the Western US (latitude: 31–49 degrees and longitude: -120- -101 degrees) during the summer of 2019. Many fires that burned that year were observed in detail as part of the joint NASA/NOAA field campaign, Fire Influences on Regional to Global Environments and Air Quality (FIREX-AQ,https://csl.noaa.gov/projects/firex-aq/). We train the Random Forest (RF) and Support Vector Machine (SVM) algorithms on meteorology from the Modern Era Retrospective analysis for Research and Applications (MERRA) and emissions from the Quick Fire Emissions Dataset (QFED). We test these machine learning algorithms against persistence and find that the SVM is fast and accurate and that the RF reproduces spatial variability of the emissions.

Figure 1: Daily evolution of the Carr Fire burned area, a quantity which is correlated with emissions. Actual area is in blue and persistence prediction is in orange. Red boxes highlight periods when the persistence prediction was inaccurate (Cal Fire).

Data

Predicting wildfire emissions is a supervised regression problem. We define the target variable as the daily total of OC aerosol emitted (in kg/m²). There are many variables that one may choose as features to predict wildfire emissions, including fire and drought indices, land use and fuels data, and local meteorology (e.g., Liu et al 2014, Walker et al 2020). We restrict features to meteorological variables to cut down on the amount of data preprocessing and to efficiently train the two models.

We trained our models using MERRA, a dataset from NASA which combines observations and model outputs. This reanalysis contains fire emissions and surface meteorology hourly on a 0.5x0.625-degree grid. We pulled the Land Surface Forcing and Aerosol Diagnostics (extended) files for July-September of 2019. The Land Surface Forcing dataset contains information about surface meteorology, including wind speed, pressure, temperature, and specific humidity. The Aerosol Diagnostics (extended) dataset contains aerosol emissions from biomass burning sources derived from the QFED, and we pulled the OC variable from this dataset.

For each of the meteorological variables, we pulled out the daily average, minimum and maximum of the quantity. Each training data point contains these values for both the day before and the day of the model valid time. Additionally, the emissions from the day before the model valid time (i.e. the persistence prediction) were included as a feature in each training point. Each corresponding label was calculated as the daily total of emissions from the day the model is valid. Depending on the model being tested (SVM or RF), we chose to either scale or not scale the features (see Table 1). We treat each MERRA grid cell that contains emissions as a separate fire.

A histogram of emissions shows a high degree of right-skewness, so we log-transform the target variable (Figure 2). We removed days where persistence predicted no emissions, because we are concerned with predicting the evolution of active fires, rather than the ignition of new fires. The full dataset contained 30 features and 17,652 training points, 30% of which were held out as the test set. Before the training/testing split, we also set aside a single day (July 1, 2019) to compare the spatial distribution of emissions.

Figure 2. Illustration of the transformation applied to the target variable. The result is still skewed, but the values are within the same order of magnitude

Project Workflow

The workflow for this project proceeded as follows. Steps 2–6 were done independently for each algorithm tested, and fixed random states and seeds were used to ensure consistent data subsets for the model comparison. Please be aware that the grid search and model training for the RF can take 30–45 minutes. The RF and SVM models were chosen due to the size of the training dataset and because both methods are known to learn complex decision boundaries and regression lines. The models were trained and built using scikit-learn.

1. Load the data. Remove days where no emissions are predicted by persistence. Hold out one day (July 1st, 2019) for spatial analysis

2. Pre-process features and split into training and testing sets. RF: No feature scaling. SVM: Minmax normalization

3. Grid search to tune hyperparameters (done on 1/10th of the training set). RF: number of trees, tree depth, and number of features considered per split. SVM: regularization parameter C, RBF kernel parameter gamma, and regression cost function parameter epsilon

4. Chose the best grid search as the one which minimized correlation (R^2) between the predicted values and test set values.

5. Tune the final grid search parameters a to maximize R^2 (done on 1/10th of the data)

6. Fit the model to the full training set and predict using the test set

7. Predict July 1st and make a spatial plot using cartopy

Table 1 presents the final configuration for each model.

Table 1

Results

In analyzing model performance, we looked at the following three metrics:

· Correlation coefficient (R^2). This was the primary metric we used to judge model performance. Since our target variable was skewed and we were interested in predicting extreme as well as mean values, we engineered our models to maximize correlation between observed and modeled base-10 log of emissions.

· Mean Absolute Error (MAE). We used this metric as another way to assign a single value to model performance on the test set. This metric was preferred over root mean squared error (RMSE) because MAE is more robust than RMSE in the case of skewed data.

· Runtime. In addition to metrics which address model skill, we also looked at runtime of the model. A goal of this research is to use these methods operationally, so quick runtime is preferred.

Based on the above metrics applied to models trained on meteorology and persistence alone, we softly recommend the SVM over the RF for future study. The SVM MAE is lower than the RF or persistence MAE (SVM = 1.509, RF = 1.991, Persistence=1.697). Although the persistence R^2 remains the highest, the SVM R^2 is greater than the RF R^2 (SVM = 0.784, RF = 0.774, Persistence=0.837). Figure 3 shows scatter and contour plots of true emissions versus emissions predicted by each method. The region highlighted in the yellow circle on the persistence scatter plot represents days of extreme fire growth where the persistence assumption is known to underpredict emissions. The RF and SVM both generate improved predictions in this region, even if it is at the cost of overall R^2 or MAE. In addition, the SVM runtime is a fraction of the RF runtime. This is due to the fact that we use MAE as the cost function for the RF, and this setup is not efficiently implemented in the scikit-learn library.

Figure 3. Scatter plots and density plots of true emissions vs predicted emissions for persistence (a-b), SVR (c-d), and RF (e-f). Density plots illustrate where majority of scatter points lie.

Finally, we look at whether the models can represent the spatial variability in emissions. This is one place where the RF outperforms the SVM (Figure 4). The RF can replicate variation across clusters of emissions hotpots rather than predicting mean behavior everywhere as the SVM appears to do. These plots are for a single day, and more analysis is needed to see if this difference between RF and SVM is systematic.

Figure 4. Spatial plots of daily emissions for true values (a), SVM (b), and RF © for July 1, 2019. Color bar has units of kg/m²

Future Work

We plan to continue investigation into the best model with which to predict daily wildfire emissions. This will involve adding more features (such as fire weather and danger indices, fuels information, FIREX-AQ observations) and fire years to our training dataset. We expect that adding additional complexity will prevent our models from predicting mean behavior (as in Figure 4b). Since the SVM architecture is the fastest, we expect to continue using this model. We may also look into using a Neural Network to solve this problem or optimizing the RF to improve runtime. If this work is successful, we may test this data-driven emissions inventory in state of the art smoke forecasting systems to see if air quality forecasting systems improve.

References

Cal Fire, Welcome to CAL FIRE. (n.d.). Retrieved January 14, 2020, from https://www.fire.ca.gov/, (No Title). (Top 20 Most Destructive California Wildfires). Retrieved September 2020, from https://www.fire.ca.gov/media/5511/top20_destruction.pdf

Liu, Y., Goodrick, S., & Heilman, W. (2014). Wildland fire emissions, carbon, and climate: Wildfire-climate interactions. Forest Ecology and Management, 317, 80–96. https://doi.org/10.1016/j.foreco.2013.02.020

Spracklen, D. V., Logan, J. A., Mickley, L. J., Park, R. J., Yevich, R., Westerling, A. L., & Jaffe, D. A. (2007). Wildfires drive interannual variability of organic carbon aerosol in the western U.S. in summer. Geophysical Research Letters, 34(16). https://doi.org/10.1029/2007GL030037

Walker, X.J., Rogers, B.M., Veraverbeke, S. et al. Fuel availability not fire weather controls boreal wildfire severity and carbon emissions. Nat. Clim. Chang. 10, 1130–1136 (2020). https://doi.org/10.1038/s41558-020-00920-8

Wegesser, T. C., Pinkerton, K. E., & Last, J. A. (2009). California Wildfires of 2008: Coarse and Fine Particulate Matter Toxicity. Environmental Health Perspectives, 117(6), 893–897. https://doi.org/10.1289/ehp.0800166