Objective The intensification of climate change has led to a significant escalation in flood risk within shallow mountainous areas, posing a severe threat to human life, health, and ecological security. These transitional areas, often situated at the interface between mountainous terrain and urbanized plains, are uniquely vulnerable to the hydrological impacts of extreme precipitation. Existing research has established that green infrastructure (GI), through its influence on fundamental hydrological processes such as the rainfall – runoff and runoff – sediment relationships, can play a pivotal role in stormwater management. However, the current body of literature predominantly focuses on two main scales: the effectiveness of individual GI elements at the localized plot level and the impact of the broader green space matrix at the large basin scale. Consequently, a critical knowledge gap persists concerning the influence of the spatial configuration of GI patches — such as their shape, size, and degree of fragmentation — on hydrological responses at the finer, sub-basin scale, which is the most relevant scale for understanding flood generation. Clarifying the mechanisms through which GI spatial patterns affect mountainous stormwater runoff and subsequently optimizing these patterns are crucial steps toward enhancing the flood prevention and control capabilities of shallow mountainous areas. This research aims to bridge the knowledge gap by elucidating these mechanisms and developing an optimization framework to mitigate the adverse effects of extreme rainfall in the sensitive shallow mountainous areas.

Methods This research adopts a two-stage research framework, comprising the two stages of mechanism exploration and pattern optimization. In the stage of exploration of hydrological mechanisms, two sample basins are selected within the shallow mountainous area of Beijing and, based on historical meteorological data and land cover data, the SWAT (soil and water assessment tool) model is used to simulate runoff generation in mountainous sub-basins with high spatiotemporal resolution. Meanwhile, machine learning methods, specifically an XGBoost-based model, are applied to the sample data to construct a high-accuracy predictive model for stormwater runoff generation, with a focus on GI spatial pattern characteristics as predictor variables. To interpret the machine learning results, the SHAP (SHapley Additive exPlanations) framework is employed to quantitatively elucidate the impact mechanisms of various GI spatial pattern metrics on mountainous stormwater runoff. In the pattern optimization stage, key GI spatial metrics are identified as optimization variables based on their hydrological influence. Under a dual-objective framework emphasizing both cost-effectiveness and flood mitigation efficacy, the NSGA-Ⅱ (nondominated sorting genetic algorithm Ⅱ) is used to optimize GI configuration for a representative shallow mountainous area. The effectiveness of these optimizations in reducing flood risks is validated through extreme historical rainfall scenarios.

Results The resulting predictive model for mountainous runoff generation demonstrates excellent simulation and forecasting capabilities, especially in modeling the influence of GI spatial pattern changes on runoff processes in complex mountainous terrains. The interpretive analysis using SHAP on the trained model provides crucial insights into the underlying mechanisms. Among the numerous GI landscape metrics evaluated, two features emerge as the most critical drivers positively correlated with increased mountainous stormwater runoff: the patch density (PD) of closed-canopy deciduous broad-leaved forests and the percent of landscape (PLAND) occupied by grasslands. The analysis reveals that an increase in either of the aforesaid two metrics consistently contributes to higher predicted runoff volumes. In contrast, the spatial pattern characteristics of other vegetation types, such as closed-canopy evergreen coniferous forests and closed-canopy deciduous coniferous forests, are found with a comparatively weak and less significant influence on the hydrological response. During the multi-objective pattern optimization process, using the two most influential metrics (PD and PLAND) as adjustable variables for a typical area, the optimized spatial pattern is able to reduce flood risk by 13.5% under the scenario of once-in-a-century extreme rainfall.

Conclusion The XGBoost machine learning model displays outstanding applicability for flood risk assessment and hydrological scenario simulation in shallow mountainous areas. An in-depth analysis of the GI spatial metrics identified by SHAP interpretation suggests that the fragmentation resulting from increased PD of closed-canopy deciduous broad-leaved forests, together with the impact of grassland PLAND on the runoff coefficient, are the core driving factors of stormwater runoff generation in these mountainous contexts. Additionally, the shape and configuration of grassland patches may further promote stormwater runoff. Accordingly, in the process of optimizing GI spatial arrangements in shallow mountainous areas, enhancing the connectivity of closed-canopy deciduous broad-leaved forest while reducing the size of large grassland patches is found conducive to forming optimal GI layouts that reduce flood risk under extreme precipitation. Through the application of interpretable machine learning techniques, this research reveals the underlying mechanisms by which different GI spatial pattern metrics influence mountain runoff generation and, based on these findings, effectively reduces regional flood risk during extreme rainfall events. The methodological approach and practical guidance provided by this research offer robust technical support for flood-mitigating green space planning in similar shallow mountain terrains and contribute valuable experience for regional adaptation to intensified climate-driven stormwater challenges.