Since 1984, satellites have observed a growing trend in summer wildfire activity in the western United States, with the total burned area increasing by 104,000 acres (42,100 hectares) per year on average [Abolafia-Rosenzweig et al., 2022]. From 1984 to 2000, wildfires across an area including all or parts of 11 states burned about 27.4 million acres in total, whereas from 2001 to 2018, this figure grew to about 55.9 million acres. In 2020 alone, the burned area jumped to roughly 8.7 million acres—equivalent to 32% of the cumulative area burned from 1984 to 2000—and the 2020 and 2021 fire seasons combined burned almost 15 million acres of the western United States, an area nearly as large as West Virginia.
This trend is largely attributable to longer and drier fire seasons caused by human-induced global warming [Abatzoglou and Williams, 2016; Zhuang et al., 2021]—and it is likely to accelerate. Projections out to 2050 suggest that the climate of the U.S. West will be twice as conducive to forest fires compared with that of the 30-year period from 1991 to 2020 [Abatzoglou et al., 2021].
In spring 2020, Jimy Dudhia, a scientist at the National Center for Atmospheric Research, asked us whether established relationships between climate and fire can be used to forecast fire activity accurately. This question ignited our curiosity, fueling research to find out whether weather in the winter and spring can reliably predict the severity of the fire season the following summer.
The Climate-Fire Connection
The western United States is in the midst of an unprecedented period of widespread megadrought and fire activity that exceeds the severity of any other period observed in available millennia-long paleorecords [Williams et al., 2022; Higuera et al., 2021]. We have seen and felt the impacts of wildfires firsthand from our drought-stricken hometown of Boulder, Colo., over the past 2 years. We have evacuated our homes to escape several wildfires, and we witnessed neighborhoods burn to the ground amid the devastating Marshall Fire this past winter.
Fires across the western United States cause thousands of smoke-related deaths and destroy thousands of homes each year, they have increased COVID-19 mortality, and they have led to persistent changes in ecosystems and water supplies.
On a national level, these fires cause thousands of smoke-related deaths and destroy thousands of homes each year, they have increased COVID-19 mortality, and they have led to persistent changes in ecosystems and water supplies. Suppressing these fires requires government expenditures that frequently exceed $1 billion annually. Thus, accurate forecasts of fire activity across a broad scale are becoming increasingly important for efficiently allocating resources required for wildland firefighting. Close relationships between climate and fire in the West, which are leveraged by our fire forecasting systems, may also motivate policy intended to reduce greenhouse gas emissions, which are heavily contributing to heating, drying, and increasingly severe wildfire seasons in the region [Abatzoglou and Williams, 2016; Zhuang et al., 2021].
The tight coupling between climate and fire in the western United States has been enhanced by the legacy of fire suppression and a lack of prescribed burning since the time of Euro-American colonization around 1800, leading to historically dense forests. Fire prediction models based on climatic conditions can help validate whether land management strategies like tree thinning and prescribed burning can counter the effects of the heating and drying and weaken the climate-fire coupling. For instance, if climate-based predictions of fire activity become less skillful following large-scale forest management strategies, then the strategies used are likely useful for mitigation.
Previous research established that most of the year-to-year variability and the overall trend in fire season severity over the past 4 decades in the western United States can be explained by climate fluctuations [Riley et al., 2013; Abatzoglou and Kolden, 2013; Williams et al., 2019; Abolafia-Rosenzweig et al., 2022; Westerling et al., 2006]. This correlation exists because the flammability of trees, grasses, and brush (i.e., the fuel for fires) and the rate of fire spread fundamentally depend on how dry the fuels and surrounding environment are. Fire spread involves a series of ignitions caused as heat from a fire raises neighboring fuels to their ignition temperature (Figure 1). Once enough heat has been transferred, the fuels combust. When fuels are moist, extra energy (latent heat) is required to evaporate the water and dry them before their temperature can reach the ignition point. This basic concept of thermodynamics has played a key role in regulating the year-to-year variability of broad-scale fire activity in the western United States, and it is expected to continue to do so as long as there is an abundance of fuel [Abatzoglou et al., 2021].
Fig. 1. Wildfires can spread after an initial ignition source (e.g., a campfire or lightning) sets fuel on fire. (a) In the case illustrated here, heat from the burning tree dries neighboring trees via evaporation, consuming a large amount of energy (latent heat of vaporization, 2,260 joules per gram). (b) After the neighboring trees are dry, additional heat coming from the ignited tree raises the neighboring trees’ temperatures (via sensible heat flux) until they reach the ignition temperature. The specific heat of timber (2 joules per gram per °C) determines the energy required to produce this temperature increase. (c) At the ignition temperature, these trees begin to combust, and the series of heat transfers and ignitions continues.
Machine Learning Quantifies Fire Burning
We posited that much of the variability and trend in summer burned area across the West is explained by prefire climate conditions alone.
We reframed Dudhia’s question as our central hypothesis, positing that much of the variability and trend in summer burned area across the West is explained by prefire climate conditions alone. To test this hypothesis, we developed and evaluated statistical models that “learn” historical relationships between presummer climate and summer fire activity in the western United States [Abolafia-Rosenzweig et al., 2022].
The machine learning methodology we used involved inputting combinations of presummer (i.e., winter and spring) climate conditions, or predictors—each averaged over varying presummer periods—into generalized additive models (GAMs). GAMs are linear models adapted to learn features in and model nonlinear data sets. The predictors considered included precipitation, temperature, evapotranspiration (the movement of water from the soil and plants into the air), potential evapotranspiration (the atmospheric demand for water), vapor pressure deficit (the dryness of air), and drought severity and area.
From more than 100,000 potential models, each based on a unique combination of predictors, we selected the 100 best for use in an ensemble prediction system on the basis of how well each could fit satellite observations of burned area (as determined by minimizing a metric called the Akaike information criterion). We then evaluated the predictive ability of our multimodel ensemble by comparing the burned areas it predicted for past years held out of the model training data with observed burned areas during those years. Accuracy on held-out data—referred to as cross validation—is considered the gold standard for machine learning evaluation. Namely, we performed both leave-one-year-out and retroactive forecasting cross validations.
In leave-one-year-out cross validations, models are trained with data from all years except the target predicted year, which is left out. Because this target year is not used to train the model, the prediction made for this year is called an out-of-bag prediction. This procedure is performed for every year in the study period to assess a full record of out-of-bag predictions. Retroactive forecasting mimics an operational forecasting system in which models are trained with records only from years prior to a target year to predict the target-year fire activity. In our case, we produced retroactive forecasts for the years 2002–2020, with the 2002 retroactive forecast, for example, using models trained on data from 1984 to 2001.
Both cross-validation methods depicted robust relationships and predictability between presummer climate and summer burned area across the western United States in our model ensemble (r ≥ 0.73, where 1.0 represents perfect correlation), supporting our central hypothesis [Abolafia-Rosenzweig et al., 2022].
The 2022 Summer Forecast
We recently applied the methodology described above to a new model ensemble tasked with forecasting the total western U.S. burned area for summer 2022 (June–September). This experimental forecasting effort examined the portion of the contiguous United States west of 104°W and contained within the four western regions defined by the National Integrated Drought Information System (NIDIS) drought early warning systems (DEWS)—Pacific Northwest, California-Nevada, Missouri River Basin, and Intermountain West. Predictions were made with a 1-month lead time, considering prefire climate predictors from November 2021 through April 2022.
The model ensemble mean predicts nearly half of the interannual variability of summer burned area from 1984 to 2021 on the basis of pre–fire season climate (r = 0.7; Figure 2). And the predicted trend in burned area over this time suggests an increase of 62,000 acres per year, explaining 60% of the observed trend of 104,000 acres per year. Furthermore, the model ensemble predicts whether the burned area in a given year was above or below average with 82% accuracy. These models thus explain much of the year-to-year variability and trend in fire activity over the past 4 decades. To our knowledge, no other macroscale forecasting system for burned area has outperformed our new ensemble.
Fig. 2. (a) A time series of observed annual summer fire season burned areas shows the increasing trend since 1984, which has been punctuated by extreme seasons, such as in 2020. The experimentally predicted fire season burned area for 2022 is 3.8 million acres (black star), which is the eighth largest in the model’s 39-year record (and ninth largest in the satellite record). All results shown here are from a leave-one-year-out cross validation with 1-month lead time predictions. Gray shading shows the 95% ensemble range. Black and red dashed lines show the linearized trends of modeled and observed burned area from 1984 to 2021. (b) Spring vapor pressure deficit (VPD) normalized anomalies (σ) relative to the 1984–2022 record are shown across the experimental domain considered in this work, which is bounded by the dashed line. Blue and red shades represent lower and higher VPDs, respectively, relative to the long-term average. (c) Dots (colored by year) represent model-predicted versus observed summer burned areas. The overall prediction accuracy is reported as correlation (r), mean bias, and mean absolute bias.
The portion of variability in observed burned area that is not explained by our approach is likely explained by a combination of factors that were not used to train our model. Such factors include fuel availability, local wind patterns, ignition sources, and the rapid onset of midsummer or late-summer drought. Part of the explanation for the model ensemble’s underestimation of the trend in increasing burn area is that our presummer climate models do not directly account for the relatively substantial trend in summer warming—summer temperatures across the West have warmed by 0.033°C per year on average from 1984 to 2021—which is not represented by winter and spring temperature trends (<0.014°C per year).
Large underestimates of the total areas burned during the 2020 and 2021 fire seasons also partially contributed to the underestimated trend by our model ensemble. Our statistical models inadequately accounted for late-summer drought intensification across the western United States in 2020 that enabled unprecedented late-summer fire activity. This activity included the devastating August Complex Fire in California, which consumed more than 1 million acres. The failure to capture the extreme extent of burned area in 2021 is likely because of the rare severity of the drought that year, which is not reflected in the model training data.
In general, statistical models that forecast fire activity on the basis of historical relationships may exhibit diminishing accuracy in a rapidly changing climate system because future conditions could frequently fall outside the range of historic variability. Indeed, four out of five of our ensemble’s least accurate predictions were for anomalously active fire seasons in just the past 10 years (2012, 2017, 2020, and 2021). Yet such models, which can continually be improved with additional training, are still highly valuable and useful, especially in the absence of other reliable means to forecast the intensity of coming fire seasons.
For the 2022 fire season, our machine learning approach experimentally forecasts a burned area of 3.8 million acres, an area roughly the size of Connecticut.
For the 2022 fire season, our machine learning approach experimentally forecasts a burned area of 3.8 million acres—an area roughly the size of Connecticut—although this figure could range from 1.9 million to 5.3 million acres considering the uncertainty quantified from the range of predictions from the full model ensemble. This forecast corresponds to the eighth-largest total burned area over the western United States in the model’s 1984–2022 record (Figure 2a), and it is 38% larger than the average summer (June–September) burned area from the simulated 1984–2021 record.
The severity of this summer’s predicted burned area is attributable to below-average winter–spring precipitation and above-average winter–spring temperatures resulting in widespread drought conditions (Figure 2b). As of May 2022, the United States Drought Monitor (USDM) reported that drought was affecting 68%–100% of the western DEWS regions.
Using and Improving Fire Forecasting Systems
Forecasts of cumulative burn area and fire activity can inform resource allocation decisions in advance and on a national scale. Specifically, for example, forecasts of the extent of summer fire activity can inform the amount of money allocated to suppress fires during the peak fire season. These forecasts are also important in helping convey that broad-scale fire activity in the western United States has been and is expected to continue to be driven primarily by climate. Amid discussions of greenhouse gas emissions targets, climate, and corresponding legislation, policymakers can apply this understanding to craft sensible legislation aimed at reversing trends of worsening wildfires.
The abilities of statistical models to predict western U.S. wildfires accurately are challenged by rapidly changing climate, fire regimes, human behavior, and corresponding vegetation changes. To help mitigate model inaccuracies, scientists can account for fuel availability using satellite-observed vegetation indices and should consider using physically based weather models to further train fire prediction models with information about potential summer hydrometeorology (e.g., precipitation), local wind patterns, and natural ignition sources (i.e., lightning). Researchers should continue to investigate how climate-fire relationships are strengthened or weakened by human-induced ecosystem changes and by new climate and fire regimes in the Anthropocene [Littell, 2018]. With such ongoing research and model development efforts, we hope the region will be better equipped to foresee and respond to future wildfires and to reduce their devastating impacts on people and nature.
The work described here was supported by NOAA MAPP (Modeling, Analysis, Predictions, and Projections) grants NA18OAR4310134 and NA20OAR4310421 and the NCAR (National Center for Atmospheric Research) Water System Program. The authors thank Karen Slater for help in improving the writing and Betty Abolafia-Rosenzweig for illustrating Figure 1. NCAR is sponsored by the National Science Foundation. Burned-area observations are obtained from NASA’s Landsat-based Monitoring Trends in Burn Severity (MTBS) data set for 1984–2020 and MODIS Terra and Aqua MCD64A1 burned-area maps for 2021. Pre–fire season predictors used for the 2022 experimental forecast are obtained from NASA’s operational North American Land Data Assimilation System (NLDAS-2). Vapor pressure deficit shown in Figure 2b is from gridMET.
Abatzoglou, J. T., and C. A. Kolden (2013), Relationships between climate and macroscale area burned in the western United States, Int. J. Wildland Fire, 22(7), 1,003–1,020, https://doi.org/10.1071/WF13019.
Abatzoglou, J. T., and A. P. Williams (2016), Impact of anthropogenic climate change on wildfire across western US forests, Proc. Natl. Acad. Sci. U. S. A., 113(42), 11,770–11,775, https://doi.org/10.1073/pnas.1607171113.
Abatzoglou, J. T., et al. (2021), Projected increases in western US forest fire despite growing fuel constraints, Commun. Earth Environ., 2, 227, https://doi.org/10.1038/s43247-021-00299-0.
Abolafia-Rosenzweig, R., C. He, and F. Chen (2022), Winter and spring climate explains a large portion of interannual variability and trend in western U.S. summer fire burned area, Environ. Res. Lett., 17(5), 054030, https://doi.org/10.1088/1748-9326/ac6886.
Higuera, P. E., B. N. Shuman, and K. D. Wolf (2021), Rocky Mountain subalpine forests now burning more than any time in recent millennia, Proc. Natl. Acad. Sci. U. S. A., 118(25), e2103135118, https://doi.org/10.1073/pnas.2103135118.
Littell, J. S. (2018), Drought and fire in the western USA: Is climate attribution enough?, Curr. Clim. Change Rep., 4, 396–406, https://doi.org/10.1007/s40641-018-0109-y.
Riley, K. L., et al. (2013), The relationship of large fire occurrence with drought and fire danger indices in the western USA, 1984–2008: The role of temporal scale, Int. J. Wildland Fire, 22(7), 894–909, https://doi.org/10.1071/WF12149.
Westerling, A. L., et al. (2006), Warming and earlier spring increase western U.S. forest wildfire activity, Science, 313(5789), 940–943, https://doi.org/10.1126/science.1128834.
Williams, A. P., et al. (2019), Observed impacts of anthropogenic climate change on wildfire in California, Earth’s Future, 7, 892–910, https://doi.org/10.1029/2019EF001210.
Williams, A. P., B. I. Cook, and J. E. Smerdon (2022), Rapid intensification of the emerging southwestern North American megadrought in 2020–2021, Nat. Clim. Change, 12, 232–234, https://doi.org/10.1038/s41558-022-01290-z.
Zhuang, Y., et al. (2021), Quantifying contributions of natural variability and anthropogenic forcings on increased fire weather risk over the western United States, Proc. Natl. Acad. Sci. U. S. A., 118(45), e2111875118, https://doi.org/10.1073/pnas.2111875118.
Ronnie Abolafia-Rosenzweig (email@example.com), Cenlin He, and Fei Chen, National Center for Atmospheric Research, Boulder, Colo.
Citation: Abolafia-Rosenzweig, R., C. He, and F. Chen (2022), For western wildfires, the immediate past is prologue, Eos, 103, https://doi.org/10.1029/2022EO220319. Published on 13 July 2022.
Text © 2022. The authors. CC BY-NC-ND 3.0
Except where otherwise noted, images are subject to copyright. Any reuse without express permission from the copyright owner is prohibited.