Ecological quality assessment and associated factors in the Yellow River basin using multi-vegetation indices RSEI and explainable machine learning

Ecological-quality assessment in the Yellow River Basin is complicated by pronounced spatial heterogeneity in vegetation cover and soil background, from sparsely vegetated arid surfaces to densely vegetated regions. Conventional RSEI relies mainly on NDVI, which is sensitive to soil-background effects at low vegetation cover and may saturate under dense vegetation. To test whether greenness-indicator choice alters ecological-quality assessment across these contrasting environmental conditions, we compared NDVI-, SAVI-, and kNDVI-based RSEI from 2000 to 2024 using correlation analysis, trajectory classification, and XGBoost–SHAP. All three models reproduced a consistent northwest-to-southeast ecological-quality gradient and an overall improving trend. Positive trajectories covered 37.66%–47.16% of the basin, versus 7.59%–13.63% for negative trajectories, with gradual changes dominating over abrupt shifts. SAVI-RSEI showed the lowest median temporal SD in sparse-vegetation and bare-soil areas (0.036, versus 0.043 for NDVI-RSEI and 0.049 for kNDVI-RSEI), whereas kNDVI-RSEI showed stronger spatial differentiation and the strongest agreement with interannual changes in the Ecological Environment Index (ρ = 0.548). Climatic predictors dominated 38.4%–40.6% of the basin, followed by soil predictors (31.6%–34.3%), while topographic and anthropogenic predictors occupied smaller proportions. Mean annual precipitation had the largest dominant spatial extent among individual predictors. SHAP analysis showed predominantly positive contributions from mean annual precipitation and soil organic carbon, but predominantly negative contributions from elevation and population density. No single greenness indicator was consistently superior across all evaluated conditions; performance varied with ecological background. Multi-greenness RSEI comparison therefore supports context-dependent indicator selection, uncertainty diagnosis, and spatially differentiated ecological monitoring in heterogeneous river basins.