关键词: 植被动态, 轨迹, 华南喀斯特, XGBoost-SHAP, 可解释机器学习模型

Abstract: Terrestrial vegetation productivity is a key indicator of ecosystem function, encompassing but not limited to carbon storage, food supply, and soil preservation. In karst regions, vegetation productivity is constrained by the underlying rock matrix, leading to abrupt and nonlinear changes. Studying vegetation dynamics and their driving factors is crucial for formulating ecological restoration. In this study area, a multi-model trajectory diagnostic algorithm was employed to differentiate the diverse kNDVI change trajectories in the South China karst region from 2002 to 2022, including trend types (positive, negative, and no trend) and trajectory shapes (linear, curvilinear, and abrupt changes), and investigate how forest dynamics have influenced these trajectories. Additionally, we employed the XGBoost-SHAP model to investigate the nonlinear effects and underlying mechanisms of explanatory variables on abrupt changes in kNDVI. The results show that: (1) The study area shows eight distinct types of changes in vegetation productivity, with 74.12% of the area experiencing positive changes and 3.44% experiencing negative changes; (2) Abrupt changes in vegetation productivity are common in the region, particularly negative abrupt changes (59.5% of negative changes); (3) Forest restoration and protection promote positive linear changes in vegetation productivity, while forest disturbances encourage negative abrupt changes; (4) Climate, human activities, terrain, and soil factors jointly contribute to abrupt changes in vegetation productivity. The XGBoost-SHAP model results highlight the importance of threshold settings in identifying significant factors influencing vegetation changes. This study highlights the significance of diverse vegetation change trajectories and offers scientific support for ecological restoration.

Key words: vegetation dynamics, trajectories, South China Karst, XGBoost-SHAP, Interpretable machine learning