Seismically triggered landslides disrupt montane ecosystems, where vegetation recovery is critical to ecological resilience and mitigating secondary hazards. However, the theoretical potential for vegetation regrowth remains poorly quantified. To quantitatively investigate this recovery process, this study proposes a survival-analysis-based inference framework to characterize and evaluate a newly introduced metric: the Potential Vegetation Greenness Recovery Timeline (Potential-VGRT). Unlike traditional regression or other machine learning-based estimation approaches, this study employs the Random Survival Forests (RSF) model to integrate both recovered events and right-censored observations (unrecovered areas), thereby capturing the observed spectrum of recovery dynamics within the observed environmental conditions. Focusing on Maoxian County—a region heavily impacted by the 2008 Wenchuan earthquake, the RSF-based estimation model demonstrated robust spatial discrimination (C-index = 0.82; Mean Time-dependent AUC = 0.89), effectively characterizing spatial gradients between rapid regeneration and arrested succession. To investigate the drivers of Potential-VGRT, we applied Shapley Additive Explanations (SHAP) to quantify both global nonlinear responses and spatially explicit local contributions. SHAP revealed strong nonlinear threshold responses and interactions among precipitation, temperature, solar radiation, and potential evapotranspiration, while spatial mapping of SHAP values further demonstrated pronounced spatial variability in their localized effects. By jointly quantifying recovery probability and duration within a unified survival framework, this approach establishes a spatially explicit statistical baseline for vegetation recovery potential under observed environmental conditions. This baseline provides a process-informed reference for interpreting recovery capacity and supports site-specific ecological assessment within the observed spatiotemporal domain in earthquake-prone mountainous regions. • Potential-VGRT is introduced to quantify vegetation recovery after landslides. • A survival-analysis-based machine learning framework was proposed in this study. • The framework adopts the RSF model to cover quick recovery and arrested succession. • RSF model shows robust spatial discrimination in the observed domain (C-index = 0.82). • SHAP reveals nonlinear thresholds and spatial heterogeneity in climatic controls.
Wan et al. (Mon,) studied this question.