Hakim, W. L., Fadhillah, M. F., Park, S., & Lee, C. W. (2025). 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. International Journal of Applied Earth Observation and Geoinformation, 144, 104890. Click (a) Nationwide wildfire susceptibility map generated using SqueezeNet and a decade of FIRMS data. (b) 2025 wildfire burn area classification based on an SVM model integrating Sentinel-1 (ACD/CCD) and Sentinel-2 multi-sensor data. (c) ROC curve demonstrating the SqueezeNet model's reliable predictive performance (AUC 0.78) against actual 2025 wildfire occurrences. This study presents a dual-stage analytical framework to address the escalating wildfire risks in South Korea by integrating deep learning for susceptibility assessment and multi-sensor satellite classification for burn area delineation. In the first stage, a nationwide wildfire susceptibility model was constructed using a decade of NASA FIRMS hotspot data (2014–2024) alongside 12 conditioning factors. Among the tested deep learning architectures, SqueezeNet demonstrated the highest predictive performance, achieving an AUC of 0.83.In the second stage, active burn areas from the extreme March 2025 wildfires were mapped by fusing Sentinel-1 SAR parameters—specifically Amplitude Change Detection (ACD) and Coherence Change Detection (CCD)—with Sentinel-2 spectral indices. A Support Vector Machine (SVM) classifier yielded an overall accuracy of 97.5% and a Kappa coefficient of 0.95. Validating the susceptibility map against actual 2025 fire perimeters resulted in an AUC of 0.78, confirming the reliability of the integrated approach and providing a robust foundation for operational early warning systems. Continue reading...
