Spatiotemporal dynamics and drivers of vegetation net primary productivity in a karst region: Nonlinear responses and spatial-scale effects revealed by XGBoost–SHAP and MGWR

Vegetation net primary productivity (NPP) is a key indicator of vegetation productivity and carbon fixation. This study examined Wenshan and Honghe prefectures in Yunnan Province, a typical karst region of southwestern China. Theil–Sen trend estimation, the Mann–Kendall test, XGBoost–SHAP, and multiscale geographically weighted regression (MGWR) were integrated to characterize NPP dynamics during 2001–2020, identify nonlinear NPP responses to major explanatory variables, and assess spatial nonstationarity. Annual mean NPP fluctuated but showed an overall increasing tendency, with an estimated linear rate of 3.05 g C m −2 yr −2 ( p = 0.052). At the pixel scale, increasing trends covered 72.14% of the study area, including 34.82% with significant increases, whereas decreasing trends occurred across 26.42%. XGBoost–SHAP results showed that land use (LU) and the normalized difference built-up and soil index (NDBSI) had the greatest model-based relative importance, followed by elevation (DEM) and precipitation (PRE). The SHAP contributions of NDBSI and PRE changed direction at approximately −0.03 and 1208 mm, respectively. Both candidate response-transition positions were highly reproducible under spatial-block bootstrap resampling within the current model and sampling framework. MGWR revealed marked temporal variation and scale-dependent spatial nonstationarity in the associations between NPP and its explanatory variables. PRE–NPP associations differed among periods, whereas NDBSI was predominantly negatively associated with NPP and operated at a strongly localized spatial scale. These findings indicate that karst-region management should combine regional land-use planning and water-risk management with targeted restoration of locally degraded surface patches.