Enhancing prediction of wildfire occurrence and behavior in Alaska using spatio-temporal clustering and ensemble machine learning
Mohamed Aymane Ahajjam, M. Allgaier, R. J. Chance, Emmanuel Chukwuemeka, Jaakko Putkonen, Timothy Pasch
Abstract
Wildfires are an integral part of Alaska’s ecological landscape, shaping its boreal forests and tundra. However, recent shifts in wildfire frequency, intensity, and seasonality pose unprecedented challenges for fire management in Alaska’s remote and ecologically vulnerable regions. This study addresses the challenge of wildfire occurrence and behavior prediction in Alaska by developing a comprehensive framework that leverages satellite-based data, geospatial features, advanced optimization, and machine learning (ML). First, NASA’s Fire Information for Resource Management System (FIRMS) dataset spanning +20 years is processed using a spatio-temporal clustering algorithm to create refined wildfire datasets. A sequential Genetic Algorithm (GA) is employed for cost-effective feature selection from 49 geospatial features, including remote sensing and reanalysis data. Histogram Gradient Boosting (HistGB) is then used for predictive modeling of wildfire occurrence, burnt area, and wildfire duration. This ensemble model’s performance is benchmarked across four prediction horizons (same-day, +7 days, +30 days, +90 days) and against various conventional ML and deep learning techniques. Results highlight key factors influencing wildfire dynamics in Alaska and demonstrate substantial improvements in prediction accuracy (e.g., an average improvement of 72.62 % in wildfire occurrence accuracy regardless of prediction horizon), offering valuable insights for risk assessment and resource allocation in wildfire management in Alaska. • Proposed a wildfires prediction framework with satellite data and machine learning. • Processed +20 years NASA FIRMS and 49 geospatial features. • Proposed HistGB model for wildfire occurrence, burnt area, and duration prediction. • Benchmarked model performance across 4 horizons using multiple prediction models. • Achieved a 72.62% improvement in wildfire occurrence prediction accuracy.