Diagnosing ecosystem service mismatches through service flows and explainable AI: evidence from the Weihe River Basin, China

Ecosystem service flows (ESFs) connect ecological supply with human demand, yet the links among service transmission, supply-demand imbalance, and nonlinear response thresholds remain difficult to diagnose in rapidly urbanizing regions. To provide a transferable diagnostic approach, we developed an explainable AI-assisted supply-demand-flow framework. We tested it in the Weihe River Basin, a representative ecological-economic transition zone of Northwest China, from 2000 to 2020. We quantified the supply, demand, and ecosystem service supply-demand ratios (ESDRs) for water yield, soil conservation, carbon sequestration, and food provision; simulated service-specific flows; and used XGBoost-SHAP-PDP to identify dominant drivers and threshold intervals. Ecological sources in the Qinling, Longshan and loess hilly regions remained separated from demand centers on the Guanzhong urban plain. Water yield supply increased from 9.41 to 11.99 billion m 3 , and soil conservation supply from 4.36 to 5.93 billion t, while their respective demands declined by 21.5% and 24.1%. Water and soil conservation flows were concentrated along terrain-controlled river corridors, carbon sequestration showed strong outward spillover and weak internal inflow, and food flows shifted toward a multi-center network linking eastern Gansu and the Guanzhong Plain. SHAP identified population density as the leading driver of water and food ESDRs and forestland proportion as the dominant driver of soil and carbon ESDRs. By coupling service-flow mechanisms with interpretable machine learning, the framework offers a reusable basis for environmental impact assessment, ecological compensation, corridor protection, demand-side regulation, and threshold-based territorial planning in comparable urbanizing regions.