Interferometric coherence outperforms backscatter and texture in explainable SAR-Based burned area classification: a case study of the 2022 Oak fire, California

Synthetic Aperture Radar (SAR) is widely used for burned area mapping because it operates independently of cloud cover and daylight, but most classification schemes rely on backscatter amplitude and texture, with interferometric coherence rarely evaluated as a standalone predictor. This study assesses the contribution of Sentinel-1 backscatter ratios, Grey-Level Co-occurrence Matrix (GLCM) texture, and interferometric coherence to burned area classification for the 2022 Oak Fire (Mariposa County, California), using exclusively free Sentinel-1 and Sentinel-2 data and validating the optical reference against the independent Monitoring Trends in Burn Severity (MTBS) product. An Otsu threshold computed directly from the study area’s dNBR distribution (0.31) showed substantial agreement with the MTBS reference perimeter (IoU = 0.785), with disagreement concentrated in low severity burn classes. An initial model incorporating optical spectral indices produced inflated accuracy (F1 = 0.970) traced to shared spectral-band dependency with the reference label; excluding these indices and pruning 226 highly correlated SAR feature pairs yielded a final 43-feature, SAR-only predictor set with no meaningful loss in performance (Random Forest F1 = 0.852, XGBoost F1 = 0.864, ROC-AUC = 0.946). SHAP-based interpretation showed that interferometric coherence, particularly the shorter-baseline pair spanning the fire’s ignition and initial spread, ranked above backscatter ratios and GLCM texture in both classifiers, despite representing a small share of the final feature set. In contrast, misclassified pixels showed disproportionate reliance on amplitude-based rather than coherence features. These results indicate that C-band coherence from freely available Sentinel-1 data can provide burned-area discrimination not captured by backscatter and texture alone, particularly when the interferometric pair spans the fire event itself.