GCCM-Based Causal Screening and Interpretable Machine Learning for Abrupt Vegetation Changes in Complex Mountainous Regions

Terrestrial vegetation productivity is vital for carbon sequestration and food provision. In Yunnan Province, complex terrain and diverse climates cause abrupt and nonlinear vegetation changes that linear trend analyses fail to capture, while existing attribution methods rarely distinguish causal drivers from spurious correlations or quantify thresholds with uncertainty. Using MODIS NDVI data (2002–2022), we developed an integrated framework combining multi-model trajectory diagnosis, land-use spatial analysis, GCCM causal screening, XGBoost-SHAP attribution, and GAM threshold refinement. Positive changes dominated 76.50% of the area; positive abrupt changes peaked in 2011–2014, negative in 2008–2009. GCCM excluded NL_change, retaining 12 factors. The positive model achieved F1 = 0.521 and AUC = 0.689, with climate factors accounting for 62.8% of mean absolute SHAP importance. The negative model achieved AUC = 0.858 (random CV) and 0.821 (spatial block CV), with human activity accounting for 17.9%—nearly double its positive share. Carbon_change ranked 10th for positive but 3rd for negative changes, consistent with an asymmetric association pattern. GAM-smoothed SHAP curves identified model-response thresholds with 95% confidence intervals: PRE_trend showed a stable zero-crossing near −8 mm/yr and a peak at 0–5 mm/yr (R2 = 0.72); DEM showed a monotonic negative relationship with no stable zero-crossing (R2 = 0.83); TEM_trend exhibited a stable zero-crossing near 0.02 °C/yr (R2 = 0.57). Spatial block cross-validation confirmed acceptable ranking ability and stable SHAP rankings. This framework links change identification, model-based attribution, and model-response threshold identification, offering a transferable paradigm for complex terrain regions.