Using Machine Learning and Aggregated Remote Sensing Data for Wildfire Occurrence Prediction and Feature Selection: A Case Study in California

Using Machine Learning and Aggregated Remote Sensing Data for Wildfire Occurrence Prediction and Feature Selection: A Case Study in California

Timothy Gao (Amador Valley High School)

Advisors: Lufang Wang (Florida International University), Xiang Gao (Massachusetts Institute of Technology)

Due to global warming, wildfires are becoming increasingly frequent and destructive, threatening environmental and human well-being on a global scale. Wildfire prediction has long posed a scientific challenge due to the complex interactions between its multitude of causes. Recently, the increasing availability of remote sensing data and machine learning (ML) advancements have provided an unprecedented opportunity to tackle this challenge. In the emerging field of ML-driven wildfire prediction, a wide range of features have been employed in prior studies, yet both a comprehensive feature importance ranking and feature subset selection for wildfire occurrence classification models remain elusive. Using California as a case study, I harvested and compiled over 100 relevant features from heterogeneous databases, including remote sensing data from multiple NASA satellites, the US Census Bureau, and OpenStreetMap. These features span from meteorological and climate conditions to vegetation and human factors. Seven classification models and four feature selection methods including additive explainers and model-specific coefficient importance algorithms were selected and tested to compute a comprehensive feature importance ranking. A simulated-annealing-based feature selection algorithm was implemented to identify a 20-feature subset that optimizes latent feature interactions and would significantly reduce data collection efforts. Combining this feature subset with the XG-Boost classifier yielded the best prediction performance with a 0.9598 F-score. These results shed valuable insight on the underlying complexities behind important wildfire causes, provide essential decision support for improved wildfire preparedness and response for governments, and can be applied in future ML wildfire prediction research and modeling.