关键词: 机器学习, 水供给服务, XGBoost-SHAP模型, 驱动机制

Abstract: Clarifying the spatiotemporal variations and driving mechanisms of water supply services holds significant guiding implications for the sustainable management of water resources and the sustainable development of ecosystems in Sichuan Province. We used the InVEST model’s water yield module to map spatiotemporal distributions of water yield in Sichuan for the years 2000, 2005, 2010, 2015, and 2020. We selected nine driving factors, including annual precipitation, annual potential evapotranspiration, annual average temperature, soil type, land use type, normalized difference vegetation index, digital elevation model, slope, and gross domestic product. Combined with global spatial autocorrelation and the XGBoost-SHAP model, we identified the spatiotemporal variations and driving factors of water yield in Sichuan Province. The results showed that the annual average water yield in Sichuan Province exhibited an N-shaped dynamic trend of increase-decrease-increase from 2000 to 2020, reaching its peak in 2020 (5.75×10⁵ m³). Water yield showed a clear pattern of being higher in the east and lower in the west. The high-value areas primarily concentrated in the southwestern part of Chengdu, the eastern part of Ya’an, and the regions of Meishan and Leshan. Over time, high-value areas gradually expanded to cover the southeastern part of Sichuan Province. Annual precipitation had the greatest contribution to the change of water yield in Sichuan region. The interaction between annual precipitation and annual potential evapotranspiration directly determined the generation and variation of water yield, representing the most dominant interactive factor. This research confirmed the effectiveness of the XGBoost-SHAP model in revealing the nonlinear impacts of driving factors on water yield, and could more directly and clearly identify the key driving factors.

Key words: machine learning, water supply service, XGBoost-SHAP model, driving mechanism