Leveraging Open-Source Data and Tools to Predict Wildfire Risk Using Machine Learning

With recent severe fire seasons in the United States, making informed and accurate predictions of wildfires has become increasingly important to risk management and preventing future destruction and human fatalities. Utilizing machine learning techniques with geospatial data allows the creation of high-quality models that can lead to real-time solutions. Because predictive models can be expensive and time-consuming to produce and run, the Wildfire Hazard Model was built to create a static risk score map using only open-source data and tools. The model examines how machine learning can effectively utilize data from past fires to predict areas at risk for future fires. As inputs, the model uses 50 years of historical wildfire perimeters from the National Interagency Fire Center (NIFC), along with other risk indicators such as precipitation averages and temperature maxima (NASA Earthdata), and population density (U.S. Census), aggregated to a 10-km resolution grid across the contiguous Unites States. A comparative analysis between different types of machine learning models found that an Extreme Gradient Boosting model led to the highest accuracy scores and fine-tuning of individual model constraints, and input data honed in on the model’s most important features. The model makes predictions for fire frequency; classifying grid cells as fire or no-fire; and severity, which categorizes based on burn acreage. For the three most recent years of fires, the model attains a frequency accuracy of 94.5% and a severity accuracy of 93.7%. This model exemplifies how machine learning can be utilized freely and successfully to aid in the identification of high-risk areas, ultimately allowing for more effective deployment of fire mitigation strategies. In this presentation, we will demonstrate how the model was built, tested, and how the results can be visualized.