High-Resolution Burned-Area Mapping and Vegetation Resilience in Heterogeneous Landscapes Using Sentinel-2 and Explainable Machine Learning

Accurate wildfire impact assessment and understanding post-disturbance recovery are essential for land management in fire-prone regions. This study develops a Sentinel-2–based burned-area extraction framework and integrates NDVI time-series analysis with explainable machine learning to quantify vegetation resilience across five fire-affected regions in China. The burned-area map achieves an overall accuracy of 99.8%, substantially outperforming MODIS products (77.9% and 92.7%) by better detecting fragmented patches in complex terrain. NDVI trajectories reveal three resilience pathways: compensatory recovery, stable recovery without compensation, and persistent degradation. Recovery times ranged from approximately 2 months to over 6 months, with some high-altitude areas showing no effective recovery. An XGBoost–SHAP model explains spatial recovery variability (R2 = 0.50–0.88) and identifies a consistent shift from early climate control to later topographic regulation. Landscape heterogeneity promotes recolonization only within intermediate thresholds, temperature exhibits optimal windows, and precipitation shows diminishing returns. Topography acts primarily by redistributing hydrothermal conditions rather than as an independent driver. The results demonstrate strong spatial variability in ecosystem stability and highlight nonlinear interactions among climate, terrain, and landscape structure as key determinants of resilience. The proposed framework improves burned-area monitoring and supports targeted ecological restoration and adaptive land-use planning in heterogeneous landscapes.