Burned areas in the western US have increased ten-fold since 1980s, which are attributable to multiple factors, including increasing heat, changing precipitation patterns, and extended drought. To better understand how these factors contribute to large fire emissions (gridded monthly fire emissions >95th percentile of all the fire emissions in the western US; 0.009 Gg/month), we build a machine learning model to predict fire emissions (PM_2.5) over the western US at 0.25° resolution, interpreted using explainable artificial intelligence (XAI). From the predictor contributions derived from XAI, we conduct k-means clustering analysis to identify four clusters of predictor variables representing different drivers of large fire emissions. The four clusters feature the contributions of fuel load (Cluster 1) and different levels of dryness (Cluster 2–4), controlled by fuel moisture, drought condition, and fire-favorable large-scale meteorological patterns featuring high temperature, high pressure, and low relative humidity. In the past two decades, large fire emissions peak in summer. However, large fire emissions increased significantly in September and October in 2010–2020 relative to 2000–2009, extending the peak large fire emissions from summer to autumn. The larger enhancements of large fire emissions during autumn compared to summer are contributed by decreased fuel moisture, along with more frequent concurrent fire-favorable large-scale meteorological patterns and drought. These results highlight fuel drying as a common driver supported by multiple drivers, such as warmer temperature and more frequent synoptic patterns favorable for fires, in increasing the autumn risk of large fire emissions across the western US.

Key Points

Using explainable artificial intelligence and k-means clustering, we identify four clusters of different drivers of large fire emissions
The peak large fire emissions of the first three clusters extended from summer to autumn in 2010–2020 relative to 2000–2009
The increased fire emissions are caused by warmer temperature and concurrent drought and large-scale pattern of high pressure and low relative humidity

Plain Language Summary

Global warming has been raising temperature and drying out the western US. The increasingly warmer climate influences the seasonal water cycle over the western US and changes wildfire activity and its seasonality. Explainable artificial intelligence (XAI) is a set of useful tools for interpreting the predictions made by the machine learning (ML) models. Leveraging the power of XAI and a statistical clustering method, we built a ML model to predict the fire emissions over the western US and grouped the grids with large fire emissions by which predictors have larger contributions to the large fire emissions. We identified four groups of large fire emissions controlled by abundant fuel and extreme, moderate, and weak drying conditions, respectively. The drying conditions are contributed by multiple factors, including drought, local dryness, and fire-favorable large-scale meteorological patterns (high temperature, high pressure, and low relative humidity). Additionally, the large fire emission peak of the first three groups extends from summer to autumn. The increased fire emissions in autumn are caused by warmer temperature, decreased fuel moisture, along with concurrent fire-favorable large-scale meteorology and drought. These findings underscore the importance of drying in increasing the autumn risk of large fire emissions across the western US.

1 Introduction

Since the 1980s, burned area in the western US has increased by a factor of five, with the largest increases occurring in California, Washington, Oregon, and Idaho (MTBS, 2020). Fires are not only getting bigger but also occurring more frequently (Abatzoglou & Williams, 2016; Dennison et al., 2014). In 2020, a series of large fires burned across many states in the western US, including northern California, Oregon, and Washington. Extremely hot and dry conditions, along with strong and gusty winds, drove the development and spread of these fires, which burned more than 10 million acres and caused around $20 billion US dollars in damages and fire suppression costs (NIFC, 2020). In addition, the fires in California emitted 1,181 thousand short tons of fine particulates with a diameter of 2.5 μm or less (PM_2.5), which is also the largest PM_2.5 emissions from fires on record since 2000 (CARB, 2021). The smoke was transported to downwind regions, resulting in the worst air quality recorded in the regions since 2011 and increased respiratory mortality cases (Audrey, 2020; Liu et al., 2021).

Anthropogenic climate change has contributed to increased fire frequency and size in the western US (Dennison et al., 2014; Westerling et al., 2006; Williams et al., 2019; Zhuang et al., 2021). Under a warming climate, increases in temperature and vapor pressure deficit have escalated fuel aridity and flammability, facilitating fire activity (Abatzoglou & Williams, 2016; Holden et al., 2018). In addition, warming temperatures in winter and spring could lead to longer growing season, increasing vegetation growth, biomass, and fuel load, which would increase fire severity when fires occur (Halofsky et al., 2020). An increasing trend in drought duration and severity in the west coast could create compound warm and dry conditions favoring fire ignitions and spread, as found in the recent large wildfires (Andreadis & Lettenmaier, 2006; Ge et al., 2016). Winds also play an important role in controlling fires by aiding fires spread across the land quickly (Rothermel, 1972). Besides the meteorological condition at local or grid scale (hereafter “local meteorology”) and drought that directly influence fire behavior, large-scale meteorological patterns are also key controls of fires (Wang et al., 2021, 2022; Zhong et al., 2020). Large-scale meteorological patterns are weather patterns with spatial scales larger than mesoscale systems (with a spatial scale of 10–1,000 km and time scale of days to weeks) but smaller than the near-global scale (with a spatial scale ∼40,000 km at the equator and time scales of months and years) of some modes of climate variability, and they are often linked to extreme events (Grotjahn et al., 2016). For instance, prior studies have shown that large-scale meteorological patterns featuring low relative humidity, high temperature, persistent high pressure, and winds from inland provide a warmer and drier condition favorable for fires (Dong, Leung, Qian, et al., 2021; Dong, Leung, Song, et al., 2021; Wang et al., 2021; Zhong et al., 2020). An example of such weather phenomena are the Santa Ana winds, which occur during fall and are characterized by strong, dry downslope winds driven by large-scale meteorological patterns. These winds are notorious for exacerbating fire hazards in Southern California (Keeley & Syphard, 2019; Westerling et al., 2004). In addition to weather conditions, summer wildfires in the western US are also strongly associated with warming spring temperature and snow drought, which provide drier conditions conducive to summer fire development (Abolafia-Rosenzweig et al., 2022; Westerling et al., 2006). A later onset of rainy season in the observations (Luković et al., 2021) and a projected sharpening of the precipitation seasonal cycle (Dong, Leung, Lu, & Gao, 2019; Dong, Leung, Lu, & Song, 2019; Swain et al., 2018) may also contribute to lengthening of the dry season and increase fire risk (Swain, 2021). These studies show the individual or multiple factors that may associate with large fires or fire risk, where most highlight the importance of dry condition on fires, driven by various factors. Yet challenges remain in identifying the key factors and understanding their conjunctive contributions to changes in large fires in the western US. Hence this study addresses three key questions about fire emissions changes: (a) What are the factors control large fire emissions in the western US? (b) How have large fire emissions changed temporally and spatially in the recent decades? (c) What are the drivers of the large fire emissions changes?

Explainable artificial intelligence (XAI) is a class of novel methods allowing the users to understand and interpret the complex relationships between target and predictor variables revealed by machine learning (ML) models (Adadi & Berrada, 2018; Arrieta et al., 2020). XAI has been applied to understand the physical processes in the complex earth system (Abdollahi & Pradhan, 2021; Labe & Barnes, 2021; Toms et al., 2020). More specifically, XAI has been used to investigate wildfires in recent studies. For example, Kang et al. (2020) utilized an integrated ensemble model with SHapley Additive exPlanation (SHAP) to examine the predictor contributions to forest fire risk in South Korea. In their analysis, relative humidity contributes most to the forest fire risk. Besides relative humidity, other key contributing factors are related to accessibility, including road density, elevation, and population density, aligning with the fact that most of the fires in South Korea are human caused. Wang et al. (2021) used eXtreme Gradient Boosting (XGBoost) with SHAP to identify the key drivers of burned areas in different regions and the physical relationships between local meteorology and burned areas over the continental US. These two studies demonstrate the capability of XAI in understanding the drivers and their relationships with fires.

To understand the changes in large fire emissions and the driving factors, we leverage the power of the developed XAI model in Wang et al. (2022) and clustering analysis to cluster large fire emissions exceeding 95th percentile of all the fire emissions over the western US (with gridded monthly fire emissions >0.009 Gg/month) by their common key drivers. Our goal is to better understand the roles of different drivers, as delineated by the clusters of drivers in the observed changes of large fire emissions during 2000–2020. We first build an XGBoost model predicting fire emissions in the western US and apply k-means clustering to the predictor contributions from the XGBoost model using SHAP. We identify four clusters with different key drivers that jointly control large fire emissions in contrast to the small or zero fire emissions across the western US where fires frequently occur. To answer the second and third science questions of this study, we analyze the decadal changes and seasonal shift in mean large fire emissions for each cluster and identify the factors driving the changes of large fire emissions in the recent decades for each cluster. More details of the development and evaluation of the XGBoost model and descriptions of the SHAP and clustering methods are provided in Section 2. Results of cluster analysis and trends in large fire emissions are presented in Section 3, and Section 4 summarizes the results.

2 Data and Methods

2.1 PM_2.5 Fire Emission Data and Machine Learning Model

Wildfire activities depend on multiple factors, including local meteorological conditions, large-scale meteorological patterns, vegetation types and conditions, and human activity. Given numerous factors and the non-linear relationships between them and fires, ML is a useful tool to resolve the complex relationships between fires and the predictors. We utilize the ML model developed in Wang et al. (2022) that used local meteorology, large-scale meteorological patterns, land surface, and socioeconomic characteristics to predict monthly fire PM_2.5 emissions from Global Fire Emission Database version 4 with small fires (GFED4s) at 0.25° spatial resolution in Gg. Note that we include fire emissions from both large and small fires in GFED4s. We retrain the model over the western US instead of over the contiguous US. Data from 2000 to 2020 is used as determined by the GFED data availability.

The GFED fire PM_2.5 emissions are estimated by combining the burned area boosted by small fires (Randerson et al., 2012) and emission factors from Akagi et al. (2011) with a revised version of Carnegie-Ames-Stanford Approach (CASA) biogeochemical model that estimates fuel loads and combustion completeness for each monthly step (Giglio et al., 2013; van der Werf et al., 2017). The fire emissions after 2016 are estimated based on the burned area, derived from the relations between MODIS active fire detections and burned areas from GFED4s for 2013–2016 (van der Werf et al., 2017). Therefore, the fire emissions may not be consistent before and after 2016, which may contribute uncertainty in our analysis. It is also known that GFED has other uncertainties (Liu et al., 2021; Pan et al., 2020; Urbanski et al., 2018). For example, the emission factors used in GFED are based on midday sampling during peak fire emission rates (Akagi et al., 2011), which might lead to inaccurate fire emission estimations as the emission factors vary with the fire phases like smoldering or flaming (Prichard et al., 2020). While the GFED fire emission inventory is subject to uncertainties associated with fixed emission factors, any changes in the emission factors during different fire phases may have marginal effects on fire emissions at monthly timescales, as such changes typically occur at smaller timescales (hourly, daily, or sub-monthly scales). Thus, the aforementioned uncertainty does not affect our main conclusions. Additionally, the dynamic emission factors describing the temporal evolution of emissions often rely on field measurements, but the temporal variation in emission factors has not been explicitly well measured yet (Akagi et al., 2011; Saide et al., 2015; van der Werf et al., 2017). In this study, we focus on regions with high fire frequency to provide enough data samples for training of the ML model of fire emissions. Therefore, we only include grids with more than 8 months of non-zero fire emissions (in a total of 250 months). The grids hereafter refer to the grids of fire emissions and predictors at 0.25° spatial resolution. Although regions with longer fire return intervals are excluded in our analysis, the selected grids encompass 90% of the total fire emissions in the western US.

The framework of the ML model has been demonstrated in a prior study (Wang et al., 2022). The local meteorological predictors include monthly mean surface temperature, relative humidity at 2 m, precipitation, zonal (U) and meridional (V) components of wind at 10 m, 1,000-hr dead fuel moisture (FM1000), Energy Release Component (ERC), vapor pressure deficit (VPD), Standardized Precipitation Evapotranspiration Index (SPEI), and cloud-to-ground lightning flash density. ERC estimates the available energy released from forest fuels at the head of a fire's flaming front (Bradshaw et al., 1984). ERC is a composite of live and dead fuel moisture, considering the cumulative drying effect combining information from FM1000, relative humidity (RH), temperature, and precipitation duration, and it's a good reflection of dry conditions. The predictors of large-scale meteorological patterns are the standard deviation of the daily singular value decomposition (SVD), which are identified based on the day-to-day correlations between the regional mean fire PM_2.5 emissions and the five gridded daily meteorological variables. The first two SVD modes of three regions (northern California (NCA), southern Rocky Mountains (SRM), and southeastern US (SEUS)) are used to construct the predictors of large-scale meteorological patterns (Wang et al., 2021). SVD1_NCA and SVD1_RM represent low RH, high temperature, high pressure, and northeasterly winds over the NCA and SRM region. SVD2_NCA is featured by low RH, high temperature, and southwesterly winds while SVD2_RM is characterized by low RH and southerly winds. The two SVD_SElag2 represent a synoptic pattern (high RH and high pressure with southerly winds) of winter storm development. Larger SVD_SElag2 indicates larger variability in such patterns and thus more dry spell days favorable for spring fires in SEUS. We then calculate the monthly standard deviation of the daily SVD time series for the first two SVD modes, representing the day-to-day variations of synoptic fluctuations and atmospheric instability within a month. The detailed methods and discussions about the SVDs are provided in Wang et al. (2021). Note that the SVDs hereafter refer to the predictors representing the varability of the fire-favorable large-scale meteorological patterns. The land-surface predictors consist of monthly mean evapotranspiration (ET), surface soil moisture, land cover types, topography, leaf area index (LAI), vegetation fraction, and fuel load. The socioeconomic variables include the gross domestic product (GDP) and population density to represent human effects on fire emissions (e.g., ignitions and suppressions) (Bistinas et al., 2013; Parisien et al., 2016). All the predictors, including their sources and their original resolutions, are listed in Table S1 in Supporting Information S1.

In this study, we employ the eXtreme Gradient Boosting (XGBoost) model to predict fire emissions in the western US. The XGBoost model builds multiple decision trees that are added subsequently, and the new added tree learns the errors of the previous tree to reduce the loss and obtain the best prediction. The model performance is evaluated using 10-fold cross-validation. With this method, we randomly divide the data set (2000–2020 over western US) into ten equal-sized splits and train the model with nine splits to predict fire emissions for the remaining one split. The process is repeated 10 times for each split, and the predictions are evaluated by grids and regions using root mean square error (RMSE), coefficient of determination (R²), and the index of agreement (IoA). IoA is defined as:

$urn:x-wiley:23284277:media:eft21356:eft21356-math-0001$

where $urn:x-wiley:23284277:media:eft21356:eft21356-math-0002$ is the observations, $urn:x-wiley:23284277:media:eft21356:eft21356-math-0003$ is prediction, $urn:x-wiley:23284277:media:eft21356:eft21356-math-0004$ is the mean of the observations. The value of IoA ranges between 0 and 1, with values closer to 1 indicating a better fit. In addition to 10-fold cross validation, we also evaluate model's generalizability by training the model with data from 2000 to 2017 and 2019 to 2020 and testing it on the year 2018. We choose 2018 for testing because it has the second-highest fire emissions during 2000–2020 and is also within the range of training data.

2.2 Shapley Additive exPlanation (SHAP)

We apply SHAP to understand the important drivers of fire emissions identified by the ML model. SHAP is a novel model-agnostic approach to explain variable contribution based on game theory (Lundberg & Lee, 2017). The idea of SHAP is that each predictor variable is considered as a “player” in the game contributing to the “playout” (prediction). The SHAP variable importance measures the marginal contribution of each variable based on all possible combinations of the predictor variables. The marginal contribution of a predictor i is calculated by comparing the differences between the model fit $urn:x-wiley:23284277:media:eft21356:eft21356-math-0005$ including the predictor i and another model fit $urn:x-wiley:23284277:media:eft21356:eft21356-math-0006$ without predictor i, which is weighted by counting the number of permutations of the subset S.

SHAP is a consistent measurement of feature importance compared to other common feature importance methods (Lundberg et al., 2019). While most common feature importance methods only provide global importance that measures variable contributions based on all samples in the data set, SHAP has local feature importance that provides the contributions of predictors for each sample. Among all the grids in the western US, we select the grids with monthly fire emissions larger than 95th percentile of all the emissions over the western US (0.009 Gg/month), hereafter referred to as “large fire emission.” The SHAP local importance of the grids with large fire emission is used as the input data for the cluster analysis to identify different groups of controlling factors for large fire emissions.

2.3 Cluster Analysis

To understand the different groups of factors controlling large fire emissions in the western US, we apply the k-means clustering algorithm on the SHAP values of the predictor variables for the grids with large fire emission (>95th percentile; 0.009 Gg/month). We choose the SHAP values of the top 18 variables based on the absolute mean SHAP values, including temperature, ERC, VPD, FM1000, SPEI, u-wind, v-wind, lightning flash density, SVD1_NCA, SVD2_NCA, SVD1_RM, SVD2_RM, vegetation fraction, soil moisture, LAI, normalized fuel load, population, and GDP. K-means is one of the commonly used clustering algorithms that divide a multidimensional data set into k clusters, in which each data point belongs to a cluster with its cluster centroid (the center of a cluster; the cluster means) closest to the data point. K-means clustering determines the best centroids by a two-step process. The algorithm starts by randomly initializing the predefined number of centroids. First, it assigns data points to its nearest centroid based on their distance from the randomly selected centroid and computes the mean of all the points for each cluster. Second, it updates the cluster mean (centroid) until there is no change of data points assignment. In this study, we choose four clusters where the sum of squared distance falls suddenly (“elbow”), and the four clusters result in a reasonable agreement between complexity and representativeness. Based on the clustering results, Clusters 1–3 have higher mean fire emissions and more distinctive drivers, and they display larger decadal trends with an extension of peak fire emissions from summer to autumn than Cluster 4.

3 Results

3.1 Machine Learning Model Performance

The predictions produced by the ML model generally agree with GFED very well (Figure 1), where the R², IoA, and RMSE from 10-fold cross validation for monthly gridded prediction are 0.60, 0.86, and 0.17 Gg, respectively. Figures 1c and 1d show the maps of the average fire emissions over 2000–2020 for GFED and the prediction, respectively. The model correctly captures the hotspots of fires over California and Pacific Northwest, but generally overestimates the fire emissions. To test model generalization, we train the model using data from 2000 to 2017 and 2019 to 2020 and test on the year 2018, which has the second-highest fire emissions during 2000–2020. Compared to the 10-fold cross-validation results, the model performance for the year 2018 slightly degrades at the grid level, with R² of 0.33, IoA of 0.72, and RMSE of 0.33 Gg (Table S2 in Supporting Information S1). At the regional scale, the model successfully reproduces the temporal and spatial variability of fire emissions (Figure S1 in Supporting Information S1), demonstrating the model skills in predicting unseen data. Despite the large number of predictors, our model does not suffer from overfitting as demonstrated by the comparable model performance between training and testing process (Table S2 in Supporting Information S1). Note that compared to the ML model for predicting fire emissions in the CONUS (Wang et al., 2022), the ML model trained by fire emissions in western US for this study has lower RMSE, higher R², and higher IoA.

Details are in the caption following the image — **Figure 1**
Open in figure viewer PowerPoint

Model performance based on 10-fold cross validation. (a) scatter plot for the observation and prediction of the gridded monthly fire PM_2.5 emission (r = 0.78); (b) Time series of annual observed and predicted fire PM_2.5 emission (r = 0.98); Maps of (c) observed and (d) predicted fire PM_2.5 emission averaged over 2000–2020 (r = 0.97).

3.2 Factors Controlling Large Fire Emissions in the Western US

We identify four clusters of drivers contributing to significantly different large fire PM_2.5 emissions (>95th percentile). The mean fire PM_2.5 emissions decrease from Cluster 1 (2.48 Gg/mon) to Cluster 4 (0.07 Gg/mon) (Figure 2a; Table 1). The four clusters correspond to different combinations of leading predictors contributing to the large fire emissions in the western US (Figure 2b). The first cluster features the largest contribution of normalized fuel load, along with relatively large contributions of fuel moisture, drought as described by Standardized Precipitation-Evapotranspiration Index (SPEI), and large-scale meteorological patterns represented by Singular Value Decompositions (SVDs). Cluster 2 has the largest mean contributions of fuel moisture, SPEI, and SVDs among all clusters. Clusters 3 and 4 have smaller contributions of all variables, compared with the first two clusters. The scaled mean SHAP values are used to further reveal the differences between the clusters (Figure S2 in Supporting Information S1). Consistent with Figure 2b, Cluster 1 is dominated by the contributions of fuel load and Cluster 2 has the largest contributions of fuel moisture, followed by SPEI and SVDs. Similar to Cluster 2, large fire emissions in Cluster 3 are predominantly controlled by fuel moisture, SPEI, and SVDs, but with slightly larger contributions from variables of land surface states and characteristics (e.g., soil moisture, fuel load, and leaf area index (LAI)). Cluster 4 corresponds to significantly lower large fire emissions compared to the other clusters and features roughly equal contributions of different predictors with no prevailing drivers. Although winds are important factors in controlling fires, the monthly mean u- and v-wind do not emerge as dominant predictors in the cluster analysis (Figure 2). It may be explained by the fact that our modeling and cluster analysis are conducted at monthly scale, whereas the effect of winds on fires are more pronounced at the hourly or daily timescale. On the other hand, the SVD predictors, which capture the day-to-day variability of synoptic patterns within a month, have a great contribution to large fire emissions. Of these SVD predictors, SVD2_RM, characterized by westerly winds over Colorado (as depicted in Figure R1), is the SVD predictors associated with winds and also have larger contributions.

Table 1. The Mean Fire PM_2.5 Emissions, SPEI, Fuel Moisture, Normalized Fuel Load, and SVDs for the Four Clusters

	Mean fire PM_2.5 emissions (Gg/mon)	Mean SPEI values	Mean fuel moisture	Mean normalized fuel load	Mean SVDs
Cluster 1 (1%)	2.48	−0.78	8.94	1.21	4.27
Cluster 2 (4%)	1.81	−0.91	7.66	0.20	3.94
Cluster 3 (13%)	0.54	−0.70	8.79	0.18	4.03
Cluster 4 (82%)	0.07	−0.32	12.28	0.11	4.82

Note. The percentage distribution of grids with large fire emissions across the clusters is shown in the first column.

To examine the actual conditions of the predictors in each cluster, we compare the distributions of the predictor variables with larger contributions, including SPEI, fuel moisture, normalized fuel load, and the average of the SVD variables (Figure S3 in Supporting Information S1; Table 1). Smaller SPEI indicates more severe drought conditions, and smaller SVDs represent less variability in (i.e., more stagnant) fire-favorable large-scale circulation patterns. The patterns are featured by low RH and high temperature, providing warmer and drier conditions that are favorable for fires. Cluster 1 has the largest fuel load but the third smallest SPEI, fuel moisture, SVDs, and normalized fuel load, compared to the other three clusters. Cluster 1 corresponds to forest area in California and Oregon where fuel load is heavy and abundant (Figure 2c). Cluster 2 represents the driest conditions, with the smallest average values of SPEI, fuel moisture, and SVDs over Sierra Nevada and the southwestern US where drought occur more frequently in summer and fall (Chen et al., 2019) (Figure 2d). Cluster 3 has similar drought and dryness conditions as Cluster 1 but much smaller fuel loads, corresponding to smaller amount of fire emissions across broader regions of western US (Figure 2e). Lastly, Cluster 4 has the least severe drought and dryness condition and the smallest amount of fuel load. Generally, larger predictor contributions (i.e., SHAP values) correspond to conditions favorable for fires (i.e., drier and more fuel load).

To summarize, Cluster 1 represents large fire emissions mainly attributed to fuel load, while Cluster 2 features large fire emissions driven by extreme aridity, more stagnant large-scale meteorology, and drought conditions. Cluster 3 shows large fire emissions contributed by moderate dryness, fuel load, and land surface characteristics, and Cluster 4 includes relatively less large fire emissions, with more frequent occurrence caused by less intense aridity, drought, and large-scale meteorological patterns conducive to fires.

3.3 Decadal Changes and Seasonal Shift in Mean Fire Emissions

To determine the changes in large fire emissions driven by the common key drivers within each cluster from 2000 to 2020, we calculate the annual mean large fire emissions for each cluster by aggregating the emissions within each cluster. Figure 3 shows the time series of annual mean large fire PM_2.5 emissions for the four clusters. The annual mean large fire emission in each year is calculated by averaging the monthly fire emissions (in Gg) over all the fire grids in each cluster during that year. All four clusters show increasing trends in annual mean fire PM_2.5 emissions over the western US, which are statistically significant (p < 0.05) except for Cluster 4. By removing 2018 to 2020 with very large fire emissions, the increasing trends still exist between 2000 and 2017 but they are not statistically significant (Figure S4 in Supporting Information S1). Although the trends are sensitive to whether the high emissions in recent years are included, the observed trends using slightly different start and end years (e.g., 2001–2018 and 2002–2019) remain significant for Clusters 2 and 3. The results show some robustness of the decadal trends despite the relatively short time period and large interannual variability. The increasing trends indicate there has been an increasing trend in the averaged fire emissions controlled by different sets of drivers since 2000.

The differences in the mean fire PM_2.5 emissions between 2000-2009 and 2010–2020 are displayed for each month and cluster in Figure 4. Large fire emission data for 2000–2009 (12,162 grids) and 2010–2020 (13,926 grids) provide substantial and comparable sample sizes for investigating decadal changes. For Cluster 1, peak emissions in 2000–2009 occurred in June and August (Figure 4a). Compared to 2000–2009, fire emissions increased the most in September during 2010–2020 and mainly occurred over Oregon and parts of northern California (Figure S5a in Supporting Information S1). Similar to Cluster 1, Cluster 2 shows peak emissions in August in 2000–2009, with significantly large enhancements in September and October in the recent decade (Figure 4b). The largest increases in September mainly occurred in mixed conifer over northern California and in prairie areas over northeastern Colorado (Figure S5b in Supporting Information S1). Cluster 3 has peak emissions in July, August, and September in 2000–2009, and fire emissions increased significantly at 0.05 level in fall (September and October) in the recent decade (Figure 4c) over the deciduous broadleaf forest in northern California, Pacific Northwest, and Idaho (Figures S5c and S5e in Supporting Information S1). Although the fire emissions of Cluster 4 also peaked in the summer of 2000–2009, the increase in fire emissions was mainly confined within summer with a delayed peak rather than an obvious extension in the fire season.

Overall, large fire emissions significantly increased in September and October for Clusters 1, 2, and 3, with enhancements found over different regions in the western US. The annual mean fire emissions in autumn months (September, October, November) also show significant increasing trends of fire emissions for the four clusters (Figure S6 in Supporting Information S1). Hereafter, we focus on Clusters 1–3 that have higher mean fire emissions due to their distinguishing features compared to Cluster 4: (a) the mean fire emissions of Clusters 1–3 are at least five times larger than that of Cluster 4 (Table 1); (b) Clusters 1–3 have more distinctive drivers compared to Cluster 4 with equal contributions across the multiple drivers; and (c) Clusters 1–3 display larger decadal trends, featuring an extension of peak fire emissions from summer to autumn, which is not found in Cluster 4.

3.4 Factors Contributing to Increased Large Fire Emissions in Autumn

The decadal trends (Figure 3) and seasonality changes (Figure 4) of large fire emissions between 2000-2009 and 2010–2020 suggest changes in the contributions of the key drivers within each cluster (Figure 2b), leading to enhanced large fire emissions in autumn over the recent decades. To reveal how the variable contributions change in autumn between the two time periods, we compare the mean SHAP values of the key predictor variables between 2010-2020 and 2000–2009. Note that the SHAP values represent the variable contributions to the prediction (i.e., fire emissions). For each variable, we convert the difference to percentage by dividing the difference by the mean SHAP value over 2000–2009 (Figure 5). Multiple factors contribute to the changes in large fire emissions between the two time periods. The enhanced fire emissions in September for Clusters 1–3 are attributed to larger contributions of large-scale meteorological patterns (SVD predictors), which increase by 113%–1700% relative to their 2000–2009 mean SHAP values (Figures 5a and Table 2). Besides SVD predictors, for Cluster 1, the contributions of temperature, drought, and u-wind increase by 61%–612%, supporting the increased fire emissions in September during 2010–2020 (Figure 5a and Table 2). Along with SVD, the contribution of fuel moisture is significant to the increased fire emissions in autumn for Cluster 2 and 3, with a relatively large contribution (Figure 2). Moreover, the contribution of fuel moisture increases by 20% and 33% for Clusters 2 and 3, respectively, during the period of 2010–2020. The contribution of temperature also increases slightly for the two clusters by around 10%–22%. To determine whether the large percentage changes of key drivers are driven by large changes in the SHAP values or small SHAP values in the mean, Figure 5b shows the changes in SHAP values for the drivers. The changes in contributions (SHAP values) are generally consistent with the percentage changes, showing the larger contributions from large-scale meteorological patterns. Overall, the reduced contributions from drivers such as normalized fuel load for Cluster 1 and SPEI for Cluster 2 are overwhelmed by the increased contributions from multiple drivers that conspire to increase large fire emissions in the more recent decade.

Table 2. Average Number of Grids With Large Fire Emissions and Mean Large Fire Emissions for 2000–2009 (Left) and 2010–2020 (Right), and the Changes in Predictor Contributions for Clusters in September and October Between the Two Periods

	Average number of grids per year with large fire emissions		Mean large fire emissions (Gg/mon)		SVD1_NCA (%)	SVD2_NCA (%)	SVD1_RM (%)	SVD2_RM (%)	Fuel moisture (%)	Temp (%)	SPEI (%)	U wind (%)
Cluster1_Sep	1	7.2	0.26	6.29	13.94	73.55	18.04	52.60	−0.42	29.94	20.67	25.78
Cluster2_Sep	0.3	17.1	0.67	5.48	5.48	65.59	26.36	56.99	26.18	4.95	−43.65	−13.05
Cluster3_Sep	16.3	52	0.45	1.06	1.06	1.96	9.72	13.19	12.75	3.75	15.91	−3.90
Cluster2_Oct	1.4	1.3	0.49	3.76	3.76	−84.69	−24.97	−23.02	103.95	153.99	13.02	531.64
Cluster3_Oct	7.6	9.7	0.16	0.98	0.98	−67.33	259.39	76.94	27.34	53.64	−31.35	306.25

In October, significant enhancements of fire emissions are observed only in Clusters 2 and 3 (Figure 4). These enhancements are mainly attributed to the u-wind, whose contribution to emissions has increased by a factor of 6 and 4 for Clusters 2 and 3, respectively, in the 2010–2020 period compared to 2000–2009 (Figures 5b and Table 2). In addition to u-wind, the contributions of normalized fuel load, temperature, and fuel moisture increase in the recent decade for Clusters 2 and 3, which also play a part in the enhanced fire emissions in October of 2010–2020 (Figures 5c and 5d and Table 2). Within the scope of the clusters, the contributions of the predominant variables related to dryness level (SVD and fuel moisture) increases in autumn over 2010–2020, along with more common large fire events (i.e., more grids with large fire emissions) and larger fire emissions for different set of drivers (Table 2).

For the variables identified by the SHAP analysis to play important roles in the changes in large fire emissions, we examine their trends to understand how those variables contribute to the emission seasonality changes. As shown in Figure 6, significantly larger negative trends of fuel moisture over western US are found in autumn (−0.085/yr and −0.114/yr for September and October, respectively) than summer (−0.066/yr and −0.049/yr for July and August, respectively), consistent with the larger warming trends in autumn. One the other hand, no significant trends are found for precipitation. Although the trends of SVDs and SPEI are not significant for September, there are more concurrent drought and fire-favorable large-scale meteorological patterns over western US in 2010–2020, compared to 2000–2009 (Figure S7a in Supporting Information S1). Similar to SVDs and SPEI, there is no significant trend of u-wind over western US for October, but more grids experienced large u-wind (|u-wind| > 3 m/s) in October (Figure S7b in Supporting Information S1), which may explain its increased contribution to large fire emissions in 2010–2020.

In summary, utilizing SHAP values, we find that the enhanced large fire emissions in September for Clusters 1–3 are contributed by increased frequency of concurrent fire-favorable large-scale meteorological patterns represented by SVDs and drought in the recent decade. For Cluster 2–3 in October, the increased large fire emissions are contributed by u-wind, corresponding to more regions experiencing large u-wind, which is not the case for summer, where SVD1_NCA has the largest increased contribution but the contributions of wind-related predictors (u-wind, v-wind, and SVD2_RM) generally decreased in 2010–2020, compared to 2000–2009 (Figure S8 in Supporting Information S1). In addition to u-wind, the large fire emissions in October are also contributed by increased temperature and decreased fuel moisture.

4 Discussion and Conclusions

This study addresses three science questions related to factors controlling large fire emissions in the western US, the spatiotemporal changes of large fire emissions, and the drivers of the observed changes. Although previous studies have investigated the changes in fire weather or meteorological conditions in autumn, connections between the meteorological changes and emissions changes could not be made without analysis of the corresponding changes in fire activities (Goss et al., 2020; Swain, 2021). Using an ML model of fire emissions and applying k-means clustering to the SHAP values of predictors, we identify and analyze clusters of common key drivers contributing to the large fire emissions in the western US. The approach of ML and SHAP directly considers the contributions of predictors and their interactions to the large fire emissions, given that the SHAP values represent the predictor contributions of the corresponding fire emissions. Cluster analysis identifies four clusters of large fire emissions driven by different sets of predictors: the first cluster features the largest contribution of fuel load; the second cluster is characterized by aridity (represented by fuel moisture, ERC, and VPD), followed by SVDs and drought; the leading factors for the third cluster are similar to the second one, but with smaller magnitudes of contributions and relatively more contributions from land surface states and characteristics; the fourth cluster has the smallest contributions of all the variables featured by equal contributions from the key variables. The SHAP values of aridity, SVDs, and drought are relatively more dominant compared to other predictor variables for the first three clusters, indicating that dryness is a critical factor for large fire emissions in the western US (Barbero et al., 2014; Holden et al., 2018; Wang et al., 2021).

Large fire emission significantly increased in autumn (September, October, November; SON) for the first three clusters during the last two decades, consistent with previous findings of increased fire activity across the western US in recent decades (Abatzoglou & Williams, 2016; Westerling et al., 2006). Additionally, September and October show significant enhancement in fire emissions in 2010–2020 compared to 2000–2009 for the three clusters. In September, the increased fire emissions for Cluster 1 occurred mostly over Oregon. For Clusters 2 and 3, the enhancements in September occurred over northern California. The enhancements for Clusters 1–3 in September are mainly contributed by the SVD predictors (SVD2_NCA and SVD2_RM) (Figure 5), which strengthen dryness level during the dry seasons. The SVD2_NCA and SVD2_RM represent variations of the large-scale meteorological patterns characterized by low relative humidity (RH) over northern California and southern Rocky Mountains, respectively (Wang et al., 2021). The smaller SVD values corresponding to more persistent synoptic patterns favorable for fires. While the SVD values over 2010–2020 did not decrease compared to 2000–2009, the low SVD values (i.e., fire-favorable large-scale meteorology) concurrent with drought happened more frequently in September in 2010–2020 than in 2000–2009 (Figure S7a in Supporting Information S1). In October, increased fire emissions over northern California and Colorado in Clusters 2 and 3 are caused by increased contribution of u-wind. Although the trend of the mean u-wind is not significant, there is a statistically significant increasing trend of the number of grids with large u-wind (|u-wind| > 3 m/s) for October (Figure S7b in Supporting Information S1). In summary, besides the contributions from SVD for September and u-wind for October, decreasing fuel moisture and increasing temperature are the common variables leading to increased fire emissions in September and October in 2010–2020 for Clusters 2 and 3, as shown in Figure 6. Such coincidence indicates the combined effects of a warming autumn and decreasing fuel moisture resulting in increased fire activities in autumn and extension of peak fire emissions (Williams et al., 2019). Hence, in the backdrop of increasing dryness and more concurrent extreme conditions (e.g., drought with fire-favorable large-scale meteorological patterns), fuel moisture is a common and key driver of the extension of peak fire emissions comparing 2000–2009 and 2010–2020, where larger decreasing trends of fuel moisture in autumn than summer (∼30%–130%; Figure 6a).

With a primary focus on large fire emissions, which have significant impact on human health and climate, fuel load is highlighted as a key driver (e.g., Cluster 1) because fire emissions are relatively more dependent than burned area on fuel load and land cover types (Battye & Battye, 2002). Similar analysis but focusing on burned area may reveal different relative importance of the key drivers, with implications for understanding the changes in burned area. Human factors are important for wildfires, but only simple predictors, such as population density, GDP, and land cover types, are used in the ML model. Not fully representing the human impacts on fires may lead to an underestimation of the role of humans in seasonality changes of peak fire emissions (Balch et al., 2017; Parisien et al., 2016). Considering the influence of meteorology and fuel, the enhanced fire emissions in autumn in the first three clusters, mostly over northern California, are primarily driven by drier conditions mainly represented by lower fuel moisture in autumn in the recent decades (Goss et al., 2020).

In sum, using fire emission data from satellite observations and an integrated machine learning approach, this study provides observational evidence and analysis of drivers for an extension of the peak season with large fire emissions in the western US. As climate model projections suggest a sharpened seasonal cycle of California precipitation, with drier spring and autumn under future warming, fire risk in autumn may further increase in the future (Dong, Leung, Lu, & Gao, 2019; Dong, Leung, Lu, & Song, 2019; Swain et al., 2018). Although autumn precipitation shows no trends during the shorter study period of 2000–2020, fuel moisture is found to decrease significantly, along with increases in temperature (Figure 6). While the observed changes in fires and their predictors in the last two decades are consistent with model projections of a warming climate, internal variability likely also plays an important role given its dominant influence on the observed precipitation changes in California in the past decades (Dong, Leung, Song, et al., 2021; Zhuang et al., 2021). Whether driven by external forcing and/or internal variability, increasing risk of large fire emissions in fall extending from the historical summer peak season poses challenges in fire management and planning.

With fires producing large emissions not only in summer but even more in autumn, exposure of populations, especially in the wildfire-urban-interface to fire smoke, will increase adverse health effects (Burke et al., 2021; Wilmot et al., 2021). To prevent more damages or health concerns due to the increasing autumn fires and emissions, useful tools such as prescribed fires that reduce fuels and improve ecosystem health can be a solution to compensate for the increased fire risks (Arkle & Pilliod, 2010; Schultz & Moseley, 2019). Future works will focus on a better understanding of interactions among humans, meteorology, hydroclimate, and vegetation under the changing climate and increasing autumn fire risk.

Acknowledgments

This research is supported by the U.S. Department of Energy Office of Science Biological and Environmental Research as part of the Regional and Global Model Analysis and Multisectoral Dynamics program areas. PNNL is operated for the Department of Energy by Battelle Memorial Institute under contract DE-AC05-76RL01830.

Open Research

Data Availability Statement

The ML prediction and predictor data set used in this study are publicly accessible online at https://zenodo.org/record/5780388#.YbjOpX3ML7E.

Supporting Information

References

Abatzoglou, J. T., & Williams, A. P. (2016). Impact of anthropogenic climate change on wildfire across Western US forests. Proceedings of the National Academy of Sciences, 113(42), 11770–11775. https://doi.org/10.1073/pnas.1607171113
10.1073/pnas.1607171113
CASADSPubMedWeb of Science®Google Scholar
Abdollahi, A., & Pradhan, B. (2021). Urban vegetation mapping from aerial imagery using explainable AI (XAI). Sensors, 21(14), 4738. https://doi.org/10.3390/s21144738
10.3390/s21144738
ADSWeb of Science®Google Scholar
Abolafia-Rosenzweig, R., He, C., & Chen, F. (2022). Winter and spring climate explains a large portion of interannual variability and trend in Western U.S. summer fire burned area. Environmental Research Letters, 17(5), 054030. https://doi.org/10.1088/1748-9326/ac6886
10.1088/1748-9326/ac6886
Web of Science®Google Scholar
Adadi, A., & Berrada, M. (2018). Peeking inside the black-box: A survey on explainable artificial intelligence (XAI). IEEE Access, 6, 52138–52160. https://doi.org/10.1109/ACCESS.2018.2870052
10.1109/ACCESS.2018.2870052
Web of Science®Google Scholar
Akagi, S. K., Yokelson, R. J., Wiedinmyer, C., Alvarado, M. J., Reid, J. S., Karl, T., et al. (2011). Emission factors for open and domestic biomass burning for use in atmospheric models. Atmospheric Chemistry and Physics, 11(9), 4039–4072. https://doi.org/10.5194/acp-11-4039-2011
10.5194/acp-11-4039-2011
CASADSWeb of Science®Google Scholar
Andreadis, K. M., & Lettenmaier, D. P. (2006). Trends in 20th century drought over the continental United States. Geophysical Research Letters, 33(10). https://doi.org/10.1029/2006GL025711
10.1029/2006GL025711
PubMedWeb of Science®Google Scholar
Arkle, R. S., & Pilliod, D. S. (2010). Prescribed fires as ecological surrogates for wildfires: A stream and riparian perspective. Forest Ecology and Management, 259(5), 893–903. https://doi.org/10.1016/j.foreco.2009.11.029
10.1016/j.foreco.2009.11.029
Web of Science®Google Scholar
Arrieta, A. B., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado, A., et al. (2020). Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Information Fusion, 58, 82–115. https://doi.org/10.1016/j.inffus.2019.12.012
10.1016/j.inffus.2019.12.012
Web of Science®Google Scholar
Audrey, C. (2020). 1 in 7 Americans have experienced dangerous air quality due to wildfires this year. Retrieved from https://www.npr.org/2020/09/23/915723316/1-in-7-americans-have-experienced-dangerous-air-quality-due-to-wildfires-this-ye
Google Scholar
Balch, J. K., Bradley, B. A., Abatzoglou, J. T., Nagy, R. C., Fusco, E. J., & Mahood, A. L. (2017). Human-started wildfires expand the fire niche across the United States. Proceedings of the National Academy of Sciences, 114(11), 2946–2951. https://doi.org/10.1073/pnas.1617394114
10.1073/pnas.1617394114
CASADSPubMedWeb of Science®Google Scholar
Barbero, R., Abatzoglou, J. T., Steel, E. A., & K Larkin, N. (2014). Modeling very large-fire occurrences over the continental United States from weather and climate forcing. Environmental Research Letters, 9(12), 124009. https://doi.org/10.1088/1748-9326/9/12/124009
10.1088/1748-9326/9/12/124009
ADSWeb of Science®Google Scholar
Battye, W., & Battye, R. (2002). Development of emissions inventory methods for wildland fire (p. 91). U.S. Environmental Protection Agency. Retrieved from https://www.epa.gov/sites/default/files/2020-10/documents/development_of_emissions_inventory_methods_for_wildland_fire.pdf
Google Scholar
Bistinas, I., Oom, D., Sá, A. C. L., Harrison, S. P., Prentice, I. C., & Pereira, J. M. C. (2013). Relationships between human population density and burned area at continental and global scales. PLoS One, 8(12), e81188. https://doi.org/10.1371/journal.pone.0081188
10.1371/journal.pone.0081188
CASADSPubMedWeb of Science®Google Scholar
Bradshaw, L. S., Deeming, J. E., Burgan, R. E., & Cohen, J. D. (1984). The 1978 National fire-danger rating system: Technical documentation. General technical report INT-169 (Vol. 44, p. 169). U.S. Department of Agriculture, Forest Service, Intermountain Forest and Range Experiment Station. https://doi.org/10.2737/INT-GTR-169
Google Scholar
Burke, M., Driscoll, A., Heft-Neal, S., Xue, J., Burney, J., & Wara, M. (2021). The changing risk and burden of wildfire in the United States. Proceedings of the National Academy of Sciences of the USA, 118(2). https://doi.org/10.1073/pnas.2011048118
10.1073/pnas.2011048118
Web of Science®Google Scholar
CARB. (2021). Wildfire emission estimates for 2020. California air Resources board. Retrieved from https://ww2.arb.ca.gov/sites/default/files/2021-07/Wildfire%20Emission%20Estimates%20for%202020%20_Final.pdf
Google Scholar
Chen, L. G., Gottschalck, J., Hartman, A., Miskus, D., Tinker, R., & Artusa, A. (2019). Flash drought characteristics based on U.S. Drought monitor. Atmosphere, 10(9), 498. https://doi.org/10.3390/atmos10090498
10.3390/atmos10090498
ADSWeb of Science®Google Scholar
Dennison, P. E., Brewer, S. C., Arnold, J. D., & Moritz, M. A. (2014). Large wildfire trends in the Western United States, 1984–2011. Geophysical Research Letters, 41(8), 2928–2933. https://doi.org/10.1002/2014GL059576
10.1002/2014GL059576
ADSWeb of Science®Google Scholar
Dong, L., Leung, L. R., Lu, J., & Gao, Y. (2019). Contributions of extreme and non-extreme precipitation to California precipitation seasonality changes under warming. Geophysical Research Letters, 46(22), 13470–13478. https://doi.org/10.1029/2019GL084225
10.1029/2019GL084225
ADSWeb of Science®Google Scholar
Dong, L., Leung, L. R., Lu, J., & Song, F. (2019). Mechanisms for an amplified precipitation seasonal cycle in the U.S. West coast under global warming. Journal of Climate, 32(15), 4681–4698. https://doi.org/10.1175/JCLI-D-19-0093.1
10.1175/JCLI-D-19-0093.1
ADSWeb of Science®Google Scholar
Dong, L., Leung, L. R., Qian, Y., Zou, Y., Song, F., & Chen, X. (2021). Meteorological environments associated with California wildfires and their potential roles in wildfire changes during 1984–2017. Journal of Geophysical Research: Atmospheres, 126(5), e2020JD033180. https://doi.org/10.1029/2020JD033180
10.1029/2020JD033180
ADSWeb of Science®Google Scholar
Dong, L., Leung, L. R., Song, F., & Lu, J. (2021). Uncertainty in El Niño-like warming and California precipitation changes linked by the IPO. Nature Communications, 12(1), 6484. https://doi.org/10.1038/s41467-021-26797-5
10.1038/s41467-021-26797-5
CASADSPubMedWeb of Science®Google Scholar
Ge, Y., Apurv, T., & Cai, X. (2016). Spatial and temporal patterns of drought in the Continental U.S. during the past century. Geophysical Research Letters, 43(12), 6294–6303. https://doi.org/10.1002/2016GL069660
10.1002/2016GL069660
ADSWeb of Science®Google Scholar
Giglio, L., Randerson, J. T., & van der Werf, G. R. (2013). Analysis of daily, monthly, and annual burned area using the fourth-generation global fire emissions database (GFED4). Journal of Geophysical Research: Biogeosciences, 118(1), 317–328. https://doi.org/10.1002/jgrg.20042
10.1002/jgrg.20042
ADSWeb of Science®Google Scholar
Goss, M., Swain, D. L., Abatzoglou, J. T., Sarhadi, A., Kolden, C. A., Williams, A. P., & Diffenbaugh, N. S. (2020). Climate change is increasing the likelihood of extreme autumn wildfire conditions across California. Environmental Research Letters, 15(9), 094016. https://doi.org/10.1088/1748-9326/ab83a7
10.1088/1748-9326/ab83a7
Web of Science®Google Scholar
Grotjahn, R., Black, R., Leung, R., Wehner, M. F., Barlow, M., Bosilovich, M., et al. (2016). North American extreme temperature events and related large scale meteorological patterns: A review of statistical methods, dynamics, modeling, and trends. Climate Dynamics, 46(3), 1151–1184. https://doi.org/10.1007/s00382-015-2638-6
10.1007/s00382-015-2638-6
ADSWeb of Science®Google Scholar
Halofsky, J. E., Peterson, D. L., & Harvey, B. J. (2020). Changing wildfire, changing forests: The effects of climate change on fire regimes and vegetation in the Pacific northwest, USA. Fire Ecology, 16(1), 4. https://doi.org/10.1186/s42408-019-0062-8
10.1186/s42408-019-0062-8
Web of Science®Google Scholar
Holden, Z. A., Swanson, A., Luce, C. H., Jolly, W. M., Maneta, M., Oyler, J. W., et al. (2018). Decreasing fire season precipitation increased recent Western US forest wildfire activity. Proceedings of the National Academy of Sciences of the USA, 115(36), E8349–E8357. https://doi.org/10.1073/pnas.1802316115
10.1073/pnas.1802316115
CASPubMedWeb of Science®Google Scholar
Kang, Y., Jang, E., Im, J., Kwon, C., & Kim, S. (2020). Developing a new hourly forest fire risk index based on catboost in South Korea. Applied Sciences, 10(22), 8213. https://doi.org/10.3390/app10228213
10.3390/app10228213
CASGoogle Scholar
Keeley, J. E., & Syphard, A. D. (2019). Twenty-first century California, USA, wildfires: Fuel-dominated vs. wind-dominated fires. Fire Ecology, 15(1), 24. https://doi.org/10.1186/s42408-019-0041-0
10.1186/s42408-019-0041-0
Web of Science®Google Scholar
Labe, Z. M., & Barnes, E. A. (2021). Detecting climate signals using explainable AI with single-forcing large ensembles. Journal of Advances in Modeling Earth Systems, 13(6), e2021MS002464. https://doi.org/10.1029/2021MS002464
10.1029/2021MS002464
ADSWeb of Science®Google Scholar
Liu, Y., Austin, E., Xiang, J., Gould, T., Larson, T., & Seto, E. (2021). Health impact assessment of the 2020 Washington state wildfire smoke episode: Excess health burden attributable to increased PM_2.5 exposures and potential exposure reductions. GeoHealth, 5(5), e2020GH000359. https://doi.org/10.1029/2020GH000359
10.1029/2020GH000359
PubMedWeb of Science®Google Scholar
Luković, J., Chiang, J. C. H., Blagojević, D., & Sekulić, A. (2021). A later onset of the rainy season in California. Geophysical Research Letters, 48(4), e2020GL090350. https://doi.org/10.1029/2020GL090350
10.1029/2020GL090350
ADSWeb of Science®Google Scholar
Lundberg, S., & Lee, S. (2017). A unified approach to interpreting model predictions. Presented at the Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
Google Scholar
Lundberg, S. M., Erion, G. G., & Lee, S.-I. (2019). Consistent individualized feature attribution for tree ensembles. ArXiv:1802.03888 [Cs, Stat]. Retrieved from http://arxiv.org/abs/1802.03888
Google Scholar
MTBS. (2020). MTBS (Monitoring trends in burn severity). Retrieved from http://mtbs.gov/
Google Scholar
NIFC. (2020). 2020 National large incident year-to-date report. National Interagency Fire Center. Retrieved from https://web.archive.org/web/20201229021815/https://gacc.nifc.gov/sacc/predictive/intelligence/NationalLargeIncidentYTDReport.pdf
Google Scholar
Pan, X., Ichoku, C., Chin, M., Bian, H., Darmenov, A., Colarco, P., et al. (2020). Six global biomass burning emission datasets: Intercomparison and application in one global aerosol model. Atmospheric Chemistry and Physics, 20(2), 969–994. https://doi.org/10.5194/acp-20-969-2020
10.5194/acp-20-969-2020
CASADSWeb of Science®Google Scholar
Parisien, M.-A., Miller, C., Parks, S. A., De Lancey, E. R., Robinne, F.-N., & Flannigan, M. D. (2016). The spatially varying influence of humans on fire probability in North America. Environmental Research Letters, 11(7), 075005. https://doi.org/10.1088/1748-9326/11/7/075005
10.1088/1748-9326/11/7/075005
Web of Science®Google Scholar
Prichard, S., O’Neill, S., Eagle, P., Andreu, A., Drye, B., Dubowey, J., et al. (2020). Wildland fire emission factors in North America: Synthesis of existing data, measurement needs and management applications. International Journal of Wildland Fire, 29(2), 132–147. https://doi.org/10.1071/WF19066
10.1071/WF19066
Web of Science®Google Scholar
Randerson, J. T., Chen, Y., van der WerfWerf, G. R., Rogers, B. M., & Morton, D. C. (2012). Global burned area and biomass burning emissions from small fires. Journal of Geophysical Research, 117(G4). https://doi.org/10.1029/2012JG002128
10.1029/2012JG002128
Web of Science®Google Scholar
Rothermel, R. C. (1972). A mathematical model for predicting fire spread in wildland fuels. Research paper INT-115 (Vol. 40, p. 115). U.S. Department of Agriculture, Intermountain Forest and Range Experiment Station. Retrieved from https://www.fs.usda.gov/research/treesearch/32533
Google Scholar
Saide, P. E., Peterson, D. A., da Silva, A., Anderson, B., Ziemba, L. D., Diskin, G., et al. (2015). Revealing important nocturnal and day-to-day variations in fire smoke emissions through a multiplatform inversion. Geophysical Research Letters, 42(9), 3609–3618. https://doi.org/10.1002/2015GL063737
10.1002/2015GL063737
ADSWeb of Science®Google Scholar
Schultz, C. A., & Moseley, C. (2019). Collaborations and capacities to transform fire management. Science, 366(6461), 38–40. https://doi.org/10.1126/science.aay3727
10.1126/science.aay3727
CASADSPubMedWeb of Science®Google Scholar
Swain, D. L. (2021). A shorter, sharper rainy season amplifies California wildfire risk. Geophysical Research Letters, 48(5), e2021GL092843. https://doi.org/10.1029/2021GL092843
10.1029/2021GL092843
ADSWeb of Science®Google Scholar
Swain, D. L., Langenbrunner, B., Neelin, J. D., & Hall, A. (2018). Increasing precipitation volatility in twenty-first-century California. Nature Climate Change, 8(5), 427–433. https://doi.org/10.1038/s41558-018-0140-y
10.1038/s41558-018-0140-y
ADSWeb of Science®Google Scholar
Toms, B. A., Barnes, E. A., & Ebert-Uphoff, I. (2020). Physically interpretable neural networks for the geosciences: Applications to Earth system variability. Journal of Advances in Modeling Earth Systems, 12(9), e2019MS002002. https://doi.org/10.1029/2019MS002002
10.1029/2019MS002002
ADSWeb of Science®Google Scholar
Urbanski, S. P., Reeves, M. C., Corley, R. E., Silverstein, R. P., & Hao, W. M. (2018). Contiguous United States wildland fire emission estimates during 2003–2015. Earth System Science Data, 10(4), 2241–2274. https://doi.org/10.5194/essd-10-2241-2018
10.5194/essd-10-2241-2018
ADSWeb of Science®Google Scholar
van der Werf, G. R., Randerson, J. T., Giglio, L., van Leeuwen, T. T., Chen, Y., Rogers, B. M., et al. (2017). Global fire emissions estimates during 1997–2016. Earth System Science Data, 9(2), 697–720. https://doi.org/10.5194/essd-9-697-2017
10.5194/essd-9-697-2017
ADSWeb of Science®Google Scholar
Wang, S. S.-C., Qian, Y., Leung, L. R., & Zhang, Y. (2021). Identifying key drivers of wildfires in the contiguous US using machine learning and game theory interpretation. Earth's Future, 9(6), e2020EF001910. https://doi.org/10.1029/2020EF001910
10.1029/2020EF001910
ADSPubMedWeb of Science®Google Scholar
Wang, S. S.-C., Qian, Y., Leung, L. R., & Zhang, Y. (2022). Interpreting machine learning prediction of fire emissions and comparison with FireMIP process-based models. Atmospheric Chemistry and Physics, 22(5), 3445–3468. https://doi.org/10.5194/acp-22-3445-2022
10.5194/acp-22-3445-2022
CASADSWeb of Science®Google Scholar
Westerling, A. L., Cayan, D. R., Brown, T. J., Hall, B. L., & Riddle, L. G. (2004). Climate, Santa Ana Winds and autumn wildfires in southern California. Eos, Transactions American Geophysical Union, 85(31), 289–296. https://doi.org/10.1029/2004EO310001
10.1029/2004EO310001
ADSGoogle Scholar
Westerling, A. L., Hidalgo, H. G., Cayan, D. R., & Swetnam, T. W. (2006). Warming and earlier spring increase Western U.S. Forest wildfire activity. Science, 313(5789), 940–943. https://doi.org/10.1126/science.1128834
10.1126/science.1128834
CASADSPubMedWeb of Science®Google Scholar
Williams, A. P., Abatzoglou, J. T., Gershunov, A., Guzman-Morales, J., Bishop, D. A., Balch, J. K., & Lettenmaier, D. P. (2019). Observed impacts of anthropogenic climate change on wildfire in California. Earth's Future, 7(8), 892–910. https://doi.org/10.1029/2019EF001210
10.1029/2019EF001210
ADSWeb of Science®Google Scholar
Wilmot, T. Y., Hallar, A. G., Lin, J. C., & Mallia, D. V. (2021). Expanding number of Western US urban centers face declining summertime air quality due to enhanced wildland fire activity. Environmental Research Letters, 16(5), 054036. https://doi.org/10.1088/1748-9326/abf966
10.1088/1748-9326/abf966
Web of Science®Google Scholar
Zhong, S., Yu, L., Heilman, W. E., Bian, X., & Fromm, H. (2020). Synoptic weather patterns for large wildfires in the northwestern United States—A climatological analysis using three classification methods. Theoretical and Applied Climatology, 76(3–4), 1057–1073. https://doi.org/10.1007/s00704-020-03235-y
10.1007/s00704-020-03235-y
ADSGoogle Scholar
Zhuang, Y., Fu, R., Santer, B. D., Dickinson, R. E., & Hall, A. (2021). Quantifying contributions of natural variability and anthropogenic forcings on increased fire weather risk over the Western United States. Proceedings of the National Academy of Sciences of the USA, 118(45). https://doi.org/10.1073/pnas.2111875118
10.1073/pnas.2111875118
Web of Science®Google Scholar

Volume11, Issue7

July 2023

e2022EF003086

Extension of Large Fire Emissions From Summer to Autumn and Its Drivers in the Western US

Abstract

Key Points

Plain Language Summary

1 Introduction

2 Data and Methods