Supply–Demand Differentiation and Ecological Functional Zoning of Ecosystem Services in the Pearl River Basin Based on Explainable Ensemble Machine Learning

Regional-scale ecosystem service surpluses may mask expanding local deficits in rapidly urbanizing areas, weakening targeted ecological management in the basin. Taking the Pearl River Basin as a case study, we integrated the InVEST model, the ecosystem service supply–demand ratio (ESDR), an MLP-XGBoost weighted ensemble optimized with the black-winged kite algorithm (BKA), SHAP analysis, and Gaussian mixture model (GMM) clustering to characterize changes in the supply and demand of water yield, carbon storage, soil conservation, and habitat quality, their contrasting drivers, and ecological functional zoning across the basin from 2000 to 2020. Water-yield and soil-conservation supply increased by 3.75% and 3.03%, respectively, with the northeastern and eastern parts of the basin as the common growth areas. Carbon storage and habitat quality supply decreased by 1.28% and 4.30%, respectively, with marked declines in the mountainous northwest and urban southeast. Carbon-storage demand increased by 245.45%, the largest demand-side increase among the four services, and southeastern cities were the principal areas of growth. Although the basin-wide mean ESDRs of all four services remained positive, those for carbon storage, habitat quality, and water yield declined overall or during specific periods. The water-yield ESDR decreased from 0.162 to 0.078, the carbon-storage ESDR from 0.303 to 0.090, and the habitat-quality ESDR from 0.555 to 0.503, whereas the soil-conservation ESDR increased from 0.015 to 0.02. Local deficits were concentrated in the Foshan–Guangzhou metropolitan area and around the urban nodes of Nanning, Liuzhou, and Guilin. The MLP-XGBoost weighted ensemble outperformed the individual models, with greater predictive accuracy and generalizability. Its test-set R2 values ranged from 0.211 to 0.737, with the highest predictive performance for water yield. SHAP analysis showed that precipitation was the dominant predictor of the water-yield ESDR, slope and precipitation jointly influenced the soil conservation ESDR, for carbon storage, the dominant controls shifted from vegetation alone to a combination of vegetation, population, and economic activity, and habitat quality was jointly influenced by population, vegetation, elevation, temperature, and road disturbance. Both the GMM and the Bayesian Gaussian mixture model (BGMM) were used to identify service types. Based on clustering performance and stability, we adopted the GMM to divide the basin into five functional zones, namely the Integrated Conservation Zone, the Key Soil and Water Conservation Zone, the Carbon–Water Regulation and Habitat Restoration Zone, the Integrated Enhancement Zone, and the Priority Restoration Zone. This study demonstrates that integrating spatiotemporal dynamics of ecosystem service supply and demand with explainable machine learning and stable clustering enables the identification of localized risks masked by regional average surpluses. These findings reveal heterogeneous governance needs, providing a scientific basis for ecological functional zoning and territorial spatial planning in the Pearl River Basin.