Machine Learning Reveals Source–Sink Regulation by Green Infrastructure Morphology Across Scales and Rainfall Scenarios
摘要
Green infrastructure (GI) enhances infiltration and interception, regulating rainfall–runoff processes. This regulation is often framed as a "source–sink" balance, but most methods label each landscape unit as a fixed source or sink. Such a binary view cannot resolve how GI partitions rainfall into runoff and retention. We therefore built a framework that derives source efficiency (SRC) and sink efficiency (SNK) from SWAT + water-balance fluxes. Taking the Dongting Lake Basin as a case, we combined MSPA, Random Forest, and SHAP to reveal how the effect of GI morphology on source–sink efficiency varies with scale and rainfall type. Results showed that: (1) Core dominated source–sink regulation and was strongest for SNK; larger, connected Core areas promoted infiltration and storage, lowering SRC and raising SNK. (2) Beyond Core, fragmented classes were strongly scale and rainfall dependent: at small scales, Islet and Perforation were the main non-Core drivers (importance values of 0.129 and 0.139), while Edge gained importance at the largest scale (0.105). (3) Several classes reversed their effect beyond critical proportions, including a U-shaped response for Perforation and repeated sign reversals for Islet, because scattered patches lack the continuous surface needed to store water and instead shift rainfall toward runoff. These findings show that rainfall type and spatial scale together govern how GI regulates runoff, and suggest that protecting connected Core areas and limiting fragmentation is essential for watershed planning, climate adaptation, and flood risk mitigation.