Leveraging explainable artificial intelligence and satellite time-series for unravelling nonlinear environmental drivers of post-wildfire vegetation recovery

Understanding post-wildfire vegetation recovery is critical for ecological restoration under climate change. Using the severe March 2024 Yajiang forest fire in China's alpine ecosystem as a case study, we integrated time-series forecasting with interpretable machine learning, augmenting multi-source predictors with a Landsat-derived differenced normalized burn ratio (dNBR). Benchmarking eleven models under an identical chronological split, random forest most accurately forecast vegetation-index residuals (test-set R² = 0.842, RMSE = 0.0230). SHAP analysis yielded three principal insights. First, recovery was governed by short-term accumulated rather than instantaneous meteorological conditions: the three-period rolling mean of atmospheric pressure was the strongest predictor (41.6% of total SHAP importance), while dNBR ranked first among standalone factors and second overall as a fire-legacy control. Second, recovery followed strongly nonlinear dynamics centred on the first post-fire year, as burn severity's suppressive effect peaked in early-to-middle recovery and decayed by the second growing season, after which lagged hydrothermal drivers rebounded. Third, synergistic enhancement was pronounced: soil–moisture interactions with evapotranspiration and recovery duration reached 4.5- and 3.9-fold the standalone moisture effect, confirming water regulates resilience through cumulative coupling. This integrated, mechanism-driven approach informs post-disaster ecosystem management in fire-prone alpine regions.