Seeing the Spark Before the Flame

Wildfire Risk Detection via U-Nets

Abstract

Wildfires increasingly threaten ecosystems, infrastructure, and human life, with climate change intensifying their frequency and severity. This project presents a multimodal two-head U-Net framework for pre-ignition wildfire risk prediction using weather, NDVI, terrain, and historical wildfire records. By treating wildfire risk mapping as a pixel-level segmentation problem, the model produces spatially explicit daily heatmaps that identify areas with elevated ignition susceptibility. Evaluation on temporally held-out wildfire seasons (2024–2025) shows strong early-warning performance, with top configurations capturing roughly 52-percent of near-term fire pixels while inspecting only 5-percent of the spatial area. Results indicate that combining dynamic environmental signals with static susceptibility information yields meaningful improvements over single-modality baselines and supports more proactive wildfire preparedness and decision-making. These results highlight the model’s potential for early-warning systems, enabling proactive resource allocation and improved wildfire preparedness at scale.