由于全局线性模型无法捕捉空间非平稳性和复杂的交互效应，表征空间异质性干旱区生态质量的时空动态及其驱动机制仍具挑战性。本研究提出一种结合Light Gradient Boosting Machine与Shapley Additive Explanations（LGBM-SHAP）以及基于地理最优分区的异质性（GOZH）模型的集成建模方法，以探究2000年至2024年内蒙古遥感生态指数（RSEI）的动态变化。结果显示，空间极化现象加剧：主要位于西部干旱区的23.2%区域呈现持续退化轨迹，而东北部的56.1%区域则表现为持续改善。尽管全局LGBM-SHAP分析确定降水、数字高程模型（DEM）和放牧强度为主要驱动因子，但也揭示了辛普森悖论：由于资源追踪效应，放牧强度在全球尺度上与生态质量呈现误导性的正相关。GOZH模型通过划分12个数据驱动的生态独特分区解决了这一问题，表明在剔除水气候混杂效应后，放牧在高质量区域中作为严格的胁迫因子发挥作用。此外，研究确定了340毫米的临界降水阈值作为生态潜力的约束条件；发现人为胁迫因子（如裸地比例）为条件性驱动因子，仅在特定的水分受限区域内施加显著的退化压力（相对贡献率为11.0%）。这些发现表明，全局平均效应掩盖了关键的局部退化机制。所提出的方法为解耦自然与人为影响提供了稳健工具。

Characterizing the spatiotemporal dynamics and driving mechanisms of ecological quality in spatially heterogeneous drylands remains a challenge due to the inability of global linear models to capture spatial non-stationarity and complex interaction effects. This study proposes an integrated modeling approach combining the Light Gradient Boosting Machine with Shapley Additive Explanations (LGBM-SHAP) and the Geographically Optimal Zones-based Heterogeneity (GOZH) model to explore the dynamics of the Remote Sensing Ecological Index (RSEI) in Inner Mongolia from 2000 to 2024. Results reveal an intensifying spatial polarization: 23.2% of the region, primarily in the arid west, exhibits a Persistent Degradation trajectory, while 56.1% in the northeast exhibits Persistent Improvement. While the global LGBM-SHAP analysis identified Precipitation, DEM, and Grazing Intensity as dominant drivers, it also exposed a Simpson’s Paradox where grazing intensity showed a misleading positive correlation with ecological quality globally due to resource tracking effects. The GOZH model resolved this by delineating 12 data-driven ecologically distinct zones, demonstrating that grazing acts as a strict stressor in high-quality zones once hydro-climatic confounding effects are removed. Furthermore, the study identified a critical precipitation threshold of 340 mm as a constraint for ecological potential; anthropogenic stressors, such as the percentage of barren land, were found to be conditional drivers, exerting substantial degradation pressure (11.0% relative contribution) only within specific water-limited zones. These findings demonstrate that global averaging effects mask critical local degradation mechanisms. The proposed approach provides a robust tool for decoupling natural and anthropogenic impacts and offers a scientific basis for zoning-based ecosystem management that operates independently of administrative boundaries.