As an effective tool for handling nonlinear and complex relationships, machine learning has drawn growing attention in urban ecology in recent years. Taking this study as an example of applying machine learning to the research on the nonlinear relationships of urban green-blue landscapes, it explored the impact of the blue-green space patterns of five high-density cities (Nanjing, Hefei, Wuhan, Suzhou, Shanghai) along the Yangtze River on carbon sequestration benefits. By using the CASA model to estimate the net primary productivity (NPP) and quantify the characteristics of the blue-green space patterns of these five cities, the XGBoost-SHAP model was employed to identify the key influencing factors of the patch layer and the type layer on carbon sequestration benefits. The research findings indicate that the characteristics of the blue-green space patterns in different cities have a significant yet differentiated impact on carbon sequestration benefits. Specifically, connectivity, shape complexity, area size and aggregation are the main factors influencing carbon sequestration benefits. In particular, indices such as the proximity index (CONTIG), the fractal dimension (FRAC) and the landscape shape index (LSI) have a significant impact on carbon sequestration. Considering the characteristics of the blue-green spaces in different cities, this study proposed planning suggestions for optimizing the layout of blue-green spaces. This study not only demonstrates the application potential of the XGBoost-SHAP model in factor analysis within the field of urban ecology but also provides theoretical support for urban ecological planning.