A Combined Statistical and Machine Learning Approach for Spatial Prediction of Extreme US Wildfire Frequencies and Sizes

• Activity Number: 335 - Recent Advances in Spatial and Spatio-Temporal Modeling
• Type: Contributed
• Date/Time: Tuesday, August 9, 2022 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Daniela Cisneros* and Raphael Huser and Yan Gong and Rishikesh Yadav and Arnab Hazra
• Companies: KAUST and King Abdullah University of Science and Technology (KAUST) and KAUST and King Abdullah University of Science and Technology and Indian Institute of Technology Kanpur
• Keywords: Approximate Bayesian inference; Extreme wildfire frequencies and size; GMRF; Random Forests; SPDE; Machine learning

Abstract

Wildfires in the United States (US) have led to considerable economic losses and social stresses. Moreover, there is concern that climate change may increase the intensity, duration, and frequency of wildfires. Wildfire prediction is an important component of wildfire management because it impacts resource distribution, mitigation of adverse effects, and recovery efforts. Therefore it is of crucial importance to develop resilient statistical methods that can reliably predict extreme wildfire events over space and time. Our approach relies on a four-stage high-dimensional bivariate sparse spatial model for zero-inflated data, which is developed using stochastic partial differential equations. In Stage 1, the observations are categorized in zero/nonzero categories. In Stage 2, smoothed parameter surfaces are obtained from empirical estimates using fixed rank kriging. In Stage 3, the standardized log-transformed positive observations are modeled using a spatial Gaussian process. Finally, in Stage 4, the predicted values are rectified using Random Forests. Our final model was shown to effectively predict low to high quantiles of US wildfire frequencies and sizes at unobserved sites.