Terrestrial water storage (TWSA) depletion poses a critical challenge to the sustainability of global drylands. The Loess Plateau presents a unique paradox as it hosts the world's largest ecological restoration program while simultaneously serving as a major coal energy base. Yet current attribution studies often overlook the spatial heterogeneity of industrial stressors, leading to uncertainties regarding the true drivers of local water deficits. Here, we integrated GRACE satellite data (2003-2022) with machine learning to disentangle the impact intensity versus spatial dominance of competing stressors. Results reveal that the most severe depletion is concentrated in the High-intensity Mining Area (HMA) with a decline rate of -16.38 mm yr(-1), significantly outpacing the Key Vegetation Restoration Area (KAR) rate of -11.34 mm yr(-1) (p < 0.05). Attribution analysis reveals distinct driver regimes: in KAR, while vegetation greening (NDVI) dominates the spatial extent (51.4% coverage), mining activity exerts the highest mean impact intensity (25.19 mm), surpassing both NDVI (18.10 mm) and Vapor Pressure Deficit (13.33 mm). In contrast, mining establishes absolute dominance in HMA with an intensity of 68.62 mm, serving as the primary driver of rapid depletion. Furthermore, we identified a compounding mechanism where Vapor Pressure Deficit amplifies the hydrological impact of mining through non-linear interactions. These findings expand the prevailing vegetation-centric view of regional water loss and highlight a spatial duality defined by widespread ecological costs versus acute industrial collapse, necessitating differentiated management strategies that strictly regulate mining intensity alongside vegetation sustainability.