Dual-stage wildfire risk analysis in South Korea: Susceptibility mapping from a decade of FIRMS data and 2025 burn area detection with multi-sensor classification

• Nationwide wildfire susceptibility maps developed by deep learning and FIRMS data. • SqueezeNet achieved AUC 0.83 (test) & 0.78 (2025 fire) in susceptibility prediction. • Burned areas mapped by combining Sentinel-1/2 data from the 2025 wildfire. • SVM classifier reached 97.5 % accuracy in burn area mapping in five provinces. • Dual-stage framework supports real-time risk assessment and post-fire mapping. Wildfire frequency and severity have escalated in South Korea, with the March 2025 event being the most destructive in its history. This study presents a dual-stage analytical framework that integrates deep learning to assess wildfire susceptibility and multi-sensor satellite classification to delineate burn areas. First, a nationwide wildfire susceptibility model was constructed using a decade of NASA FIRMS hotspot data (2014–2024) and 12 conditioning factors. Among the four tested deep learning models, SqueezeNet achieved the highest predictive performance, with an area under the curve (AUC) value of approximately 0.83 and minimal error metrics. Second, active burn areas from the 2025 wildfire were mapped by fusing Sentinel‑1 synthetic aperture radar (SAR), which includes amplitude and coherence change detection, and Sentinel‑2 spectral indices, enabling precise delineation of burn across five provinces. A support vector machine classifier yielded an overall accuracy of 97.5 % and a Kappa coefficient of 0.95. The susceptibility map, validated against the 2025 fire perimeters, achieved an AUC of 0.78, confirming the reliability of the proposed integrated framework. This approach provides a robust foundation for early warning systems and ecological risk assessments by combining multi-temporal fire patterns with validation against actual burn area.