Machine learning-based burned area detection using Sentinel-2 imagery and spectral indices.

Wildfires are catastrophic events that cause significant damage worldwide. Accurate and rapid mapping of burned areas is crucial for post-fire management, planning, and monitoring vegetation recovery. This study compares the accuracy of burned area detection in the Farashband and Andika of Iran, using Machine Learning (ML) methods, Random Forest (RF), and Support Vector Machine (SVM), implemented on the Google Earth Engine platform under two scenarios. The first scenario uses original bands from post-fire Sentinel-2 imagery. The second scenario combines the original bands that had above-average importance in the first scenario with spectral indices (Burned Area Index for Sentinel-2 (BAIS2), Burned Area Detection Index (BADI), Mid-Infrared Burn Index (MIRBI), Normalized Difference Short-Wave Infrared Ratio (NDSWIR), and Normalized Burn Ratio (NBR)) to improve detection accuracy. Evaluation results showed that in Farashband, RF achieved an overall accuracy (OA) of 94.53% and a Kappa coefficient (KC) of 0.93 in the first scenario, which improved to 97.04% and 0.96 in the second. SVM improved from 96.43% and 0.96 to 97.96% and 0.99. In Andika, RF improved from 94.56% and 0.92 to 95.61% and 0.94, while SVM increased from 95.84% and 0.97 to 97.65% and 0.98. SWIR1-2 bands were most important for RF, while Blue and Red Edge-1 bands were most effective for SVM. NBR and NDSWIR were key for RF in Farashband, while MIRBI and NDSWIR were most effective in Andika. The proposed approach can accurately identify burned areas using Sentinel-2 imagery and ML techniques, aiding in the rehabilitation and recovery of affected regions.