(362s) Developing Risk Assessment Framework for Wildfire in the United States – a Deep Learning Approach

The summer of 2020 was a record season for wildfires in the western United States. Though large fires are not unheard of historically in this region, never has the destruction had such a high impact on communities. During this period in California, more structures were damaged or destroyed and more fatalities occurred than ever before due to wildfires. In addition, more of these large wildfires are found to be caused by humans. As the US population continues to expand, more people are moving into fire-prone areas which have been, historically, largely uninhabited. This, in combination with the seemingly increasing prevalence of wildfires in the American west, makes the study of such fires more important than ever. A model which can be used to predict fires’ maximum sizes, causes, and potential future occurrences would provide a valuable and potentially life-saving tool.

One such model created by Coffield et al. (2019) illustrates the potential for the use of machine learning for such predictions. The model made use of a decision tree method, accounting for both weather and vegetation cover, to predict the final size of a fire—either small, medium, or large—as early as the time of ignition for a given location in Alaska. It was trained specifically on Alaskan wildfires over a period of 17 years and has an accuracy of approximately 65% for “large” fires, defined by the researchers as greater than 19.8 square kilometers. The technique employed by Coffield et al. (2019) is an example of a shallow machine learning algorithm. For some applications, deep learning has shown more promise than traditional machine learning, particularly in cases where data sets are large and multivariate.

In this study, a highly efficient deep learning framework is constructed based on a spatial database of wildfire incidents that occurred in the United States from 1992 to 2018. It includes 2.17 million geo-referenced wildfire records, representing a total of 165 million acres burned during the 27-year period. The framework contains three parts: 1) predict the cause of a wildfire once there is a wildfire occurs by Deep Neural Networks (DNN). Compared with some wildly used machine learning algorithms (Naive Bayes, SVM, Random Forest, XGBoost), the DNN models achieved the highest accuracy of 67.1% for a 5-class classification. 2) predict the size of a wildfire once there is a wildfire occurs by DNN. By adjusting the undersampling ratio, the DNN models achieved accuracy of 68.4% for a 5-class classification. 3) forecast the likelihood of wildfire in the near future by Long Short Term Memory Networks (LSTM). Time series data was analyzed to perform the one-week forecast for different sized classes of wildfire in the United States. A specific model was constructed to perform the one-week forecast for California state. The coefficient of determination (R²) and root-mean-square error (RMSE) were calculated for statistical assessment and the developed LSTM model achieved satisfactory predictive capabilities. This developed deep learning framework can be used for emergency response planning and risk assessment.