Ecosystem services that enhance synergy through coordinated spatial zoning and management: A network analysis and machine learning approach

Elucidating the complex interactions within the water–food–ecology nexus is crucial for ensuring the sustainable management of fragile ecosystems. However, traditional pairwise analyses often fail to capture system network dynamics and nonlinear driver responses. The Loess Plateau, where the largest ecological restoration project in the world has been implemented, is facing increasing conflicts over ecosystem services (ESs). In this study, the spatiotemporal dynamics of four key ESs (water yield, soil conservation, carbon fixation, and food provision) on the Loess Plateau from 2000 to 2024 were assessed at the county level. Signed partial correlation networks were used to quantify the trade-off degree (TD) and synergy degree (SD). Extreme gradient boosting coupled with out-of-fold SHapley additive exPlanations (XGBoost–SHAP) was used to identify predictive drivers and model response variations. A stability-screened self-organizing map delineated ES bundles and classification-sensitive transition zones. Our results showed that carbon fixation, soil conservation, and food provision increased; however, the network remained dominated by trade-offs. Carbon fixation–soil conservation had the most extensive synergy, whereas food provision–soil conservation had the most extensive trade-off degree. Moreover, 37.31% of counties experienced increasing TDs and declining SDs. TD was most strongly associated with the impervious surface proportion (IMP; 24.0%), grassland cover (13.8%), and population density (12.2%), whereas SD was driven primarily by precipitation (26.0%), slope (10.5%), and fractional vegetation cover (9.8%). The IMP response changed its sign near 0.033, while the FVC transition varied from approximately 0.47 to 0.51 across predictor sets and hydrothermal strata, indicating a conditional assessment range rather than a universal vegetation cover cap. By integrating bundle function, classification stability, and model response signals, we provide an adaptive basis for spatial management. This study provides a reference framework for alleviating socio-ecological conflicts in water-scarce regions around the world.