Assessing and Explaining Human–Nature Coordination at the County Scale in China: Combining an Interpretable Spatiotemporal Graph Neural Network and Difference‐In‐Differences

Coordinating high‐intensity human activities with fragile ecosystems remains a major challenge for global sustainability. Accurate assessment of human–nature coordination (HNC) is critical, yet county‐level variations, underlying drivers, and responses to ecological zoning have seldom been examined simultaneously within a unified framework. We construct an HNC index for Chinese county‐level units from 2001 to 2022 by integrating local and telecoupling coordination degree between ecosystem‐service supply and human activity. For 1817 counties with complete covariates, a multivariate time series graph neural network (MTGNN), integrated gradients (IG), and partial dependence plots (PDPs) characterize temporal and cross‐county predictive dependencies. A staggered difference‐in‐differences (DID) design then estimates the effect of National Key Ecological Function Zones (NKEFZ). National HNC declined slightly after 2016; 45.0% of counties improved and 54.7% deteriorated over the full study period. Precipitation, temperature, and slope contributed 34.7%, 12.3%, and 10.3% of model attribution, respectively, and precipitation was the dominant predictor in 66.1% of counties. The learned graph further identified nonlocal predictive communities. NKEFZ increased standardized HNC by 0.034. Formal channel tests support ecological restoration, reduced industrial activity pressure, and public‐service support, with the NDVI pathway accounting for approximately 30.9% of the total effect; the industrial‐structure pathway was not statistically supported. Policy effects differed across regions and limiting‐factor groups, while heterogeneity among learned‐graph communities remained relatively low. These findings connect county‐level HNC dynamics, predictive constraints, and ecological‐zoning effects, providing evidence for constraint‐sensitive policy design.