A comprehensive survey of the machine learning pipeline for wildfire risk prediction and assessment

Wildfires, intensified by climate change and human activity, present a growing global threat to ecosystems, economies, and public safety. This survey offers a comprehensive overview of machine learning approaches for wildfire risk prediction and assessment, encompassing the entire pipeline from data acquisition to model deployment. It highlights the integration of diverse data sources, including remote sensing, in-situ measurements, geospatial layers, and historical fire records and outlines pre-processing and feature engineering techniques to represent climatic, topographic, vegetation, anthropogenic, and temporal fire patterns. The paper categorizes a wide array of machine learning techniques applied in wildfire risk assessment, including traditional, deep learning, spatial, temporal, reinforcement learning, and hybrid approaches. It also examines post-processing strategies such as fire susceptibility mapping and uncertainty quantification. Key challenges, including data sparsity, interpretability, and integration of heterogeneous data, are discussed alongside prospects for ethical, adaptive, and real-time systems. By organizing the literature into a unified end-to-end pipeline, this work offers guidance for developing scalable, interpretable, and operational machine learning solutions for wildfire risk assessment.