To elucidate drought-driving mechanisms under different vegetation change states and to advance the understanding of drought processes in typical arid regions, this study takes Ningxia, China, as the study area and conducts a quantitative analysis by integrating spatial statistical methods with machine learning techniques based on multi-source data from 2001 to 2021. Vegetation restoration and degradation areas are identified through cluster analysis based on differences in fractional vegetation cover, thereby clarifying their spatial patterns. An XGBoost regression model coupled with SHAP interpretation is then constructed to systematically investigate the driving effects of multiple environmental factors on drought dynamics and to compare differences in the effects of driving factors across temporal scales. The results indicate pronounced differences between vegetation change states and drought spatial distribution patterns across periods, with the XGBoost models exhibiting strong fitting performance (best test-set R 2 = 0.879 for restoration areas and R 2 = 0.886 for degradation areas), demonstrating robust regression capability. Air temperature and elevation are identified as the most stable dominant driving factors in both types of areas. In certain periods, factors such as distance to transportation also exhibit strong localized explanatory power, reflecting their important role in shaping spatial variations in drought. During 2017-2021, air temperature exhibits the highest mean SHAP value in restoration areas (SHAP mn = 0.157), whereas elevation contributes most prominently in degradation areas over the same period (SHAP mn = 0.137). Nighttime light intensity, slope, and aspect generally display weaker explanatory power, though they exhibit stage-specific importance in particular year intervals. From a spatial heterogeneity perspective, this study reveals both commonalities and differences in drought drivers between vegetation restoration and degradation areas, providing methodological insights and theoretical support for risk regulation strategies in drought-prone regions.
Song et al. (Wed,) studied this question.