Disentangling the contributions of ecological restoration and climate variability to regional air quality improvement remains a critical challenge for nature-based solution assessments. Here, we integrated satellite NDVI, ground-based air quality monitoring (PM10, SO2, NO2, CO, O3), and meteorological data to construct a decadal (2015-2024) site-year panel across China's Loess Plateau. Using random forest meteorological adjustment, two-way fixed-effects panel models, structural equation modeling, and threshold regression, we provide four novel contributions. First, we isolate the independent net effect of vegetation restoration: after removing meteorological fluctuations and spatial-temporal fixed effects, each 0.1-unit NDVI increase reduces PM10 by 9.06 mu g & sdot;m- 3 (95% CI: -11.42 to -6.70) in semi-arid zones and 6.88 mu g & sdot;m- 3 (95% CI: -9.23 to -4.53) in semi-humid zones, with no significant effect in arid zones. Second, we identify a critical NDVI threshold (approximate to 0.45): below this level, PM10 reduction is negligible; above it, the purification effect accelerates sharply. Third, we quantify mediation pathways: vegetation reduces PM10 indirectly by increasing humidity and reducing wind speed and radiation (mediation proportion: 34-47%). Fourth, we diagnose the O3 trade-off: greening is associated with significant O3 increases only in southeastern, high-NOx zones where high-BVOC (biogenic volatile organic compounds) species (Robinia pseudoacacia, Populus spp.) dominate; this effect is statistically mediated by increased temperature and radiation. Our findings provide the first regional-scale causal evidence that ecological restoration independently improves air quality under adequate moisture conditions, establish a quantitative coverage target (NDVI >= 0.45) for maximizing co-benefits, and demonstrate that O3 outcomes are speciesdependent-requiring BVOC-sensitive afforestation in polluted areas. The analytical framework is transferable to other water-limited restoration regions globally.