Predicting Wildfire Burn Severity from Pre-Fire SAR Signatures: A Deep Learning Approach

Climate-driven increases in wildfire frequency and severity necessitate improved prediction capabilities for proactive risk management. Existing wildfire models rely heavily on realtime atmospheric data and often overfit to specific fire conditions, while pre-fire landscape assessment remains underdeveloped. This study addresses a fundamental question: can static landscape features captured before ignition predict spatial patterns of burn severity? We propose a deep learning framework using L-band polarimetric Synthetic Aperture Radar (SAR) data from NASA's Uninhabited Aerial Vehicle SAR (UAVSAR) to predict four-class burn severity from pre-fire conditions alone. Using the 2020 Bobcat Fire (115,997 acres, Southern California) as a case study, we integrate six features: HH, HV, and VV polarization backscatter, HH/HV ratio, Alpha 1 Angle from Cloude-Pottier decomposition, and terrain slope. A U-Net convolutional neural network with patch-based semantic segmentation achieved 51% accuracy (F1-macro: 0.48, ROC-AUC: 0.74-0.82), representing 2.04× improvement over random baseline (25%) and substantially outperforming pixel-wise baselines (Random Forest: 0.36, XGBoost: 0.37). The model prioritizes detection sensitivity over classification precision, achieving 64% recall for high-severity areas despite 26% precision, making it suitable for risk screening rather than definitive prediction. This framework establishes that pre-fire SAR features encode meaningful information about potential burn severity, providing a foundation for integration with real-time fire weather models. The recent launch of NASAISRO SAR (NISAR) in July 2025, offering systematic 12-day global L-band coverage, makes this approach particularly timely for operational wildfire management.