The south central United States region is predicted to have a high risk of wildfires in the latter half of the 21st century. Understanding fire regimes, i.e. the size and occurrences of fire throughout a fire season, is key to predicting and planning for future fires. Researchers from the University of Houston, funded in part by CPO’s Atmospheric Chemistry Carbon Cycle and Climate (AC4) program, have developed a fire prediction model which incorporates multiple machine learning algorithms to better predict areas burned by wildfire at a smaller spatial scale. Their work, published in the journal Atmospheric Chemistry and Physics, finds that for both the summer and winter-spring fire seasons in the US south central region non-average relative humidity and drought conditions in the preceding 3-5 months of a fire season are the top two key variables in predicting burned area. This information offers valuable direction for future fire management and prediction efforts.Further, their model demonstrates the impact that climate variability can have on the magnitude of wildfires and provides more tools to model developers for predicting wildfire in a changing climate.
While current modeling techniques can adequately predict fire area as a whole, models tend to simulate areas of land measured in hundreds of kilometers. This can result in the underprediction of burned land for specific sub-areas within the larger total wildfire area predicted. Researchers Sally Wang and Yuxuan Wang developed a model to predict burn area at a much finer scale (i.e. 50 km vs. 700 km) over a fire-prone region encompassing Texas, Oklahoma, Louisiana, and Arkansas. They then analyzed the results of their model, which predicts the area burned by wildfire each month from 2002-2015, in order to identify the relative importance various environmental factors play in determining the magnitude of burned area.
Wang and Wang grouped environmental variables into four categories: weather (e.g. precipitation), climate (e.g. indicators of when precipitation deviated from long-term average), fuel (e.g., soil moisture), and fixed geospatial variables (e.g. land cover type). They investigated each variable’s individual contribution to wildfire prediction in their model, but also the relative importance of the overall categories as a whole. Climate had the largest absolute contribution to the burned area for both fire seasons. For both fire seasons, relative humidity outside the average expected range was the top predictor variable, followed by drought conditions present in the preceding months. Interestingly, fuel considerations only appeared as important factors for the winter-spring fire season. Wang and Wang speculate that the general abundance of vegetation during the summer fire season helps explain why temperature variability and pre-fire season drought outweigh fuel in relative importance.
Previous research has shown that regional climate variability is a key driver for wildfires and Wang and Wang show that the south central US region is no exception. Their prediction model improves the ability to predict the where and when of wildfire burns at a finer scale. However, Wang and Wang caution that there are limitations to their model due to model structure and the machine learning algorithms used. Their work also does not incorporate human actions that affect wildfires, a critical factor which they encourage future work on the subject to include.