Automatic segmentation of skeleton drainage pattern using local basin graph based on graph neural network

The skeleton drainage pattern, with its elongated and complex structure, is crucial for hydrological analysis, land-use planning, and ecological conservation. Conventional methods struggle to model intricate river networks and represent localized features. This study proposes a GraphSAGE-based graph neural network built upon a local basin graph framework to automate skeleton drainage pattern segmentation. The method constructs a watershed graph structure by leveraging adjacency relationships among river catchments, thereby effectively capturing the spatial topology and hierarchical dependencies of the river network, which significantly enhances the model’s capacity to represent complex hydrological systems. To comprehensively describe drainage system attributes, it incorporates three types of features: basin connectivity, reach geometry, and elevation variability. The GraphSAGE model with a Mean aggregation strategy is constructed, significantly improving the accuracy of simulating local hydrological responses and global dynamic processes. Experiments achieve a validation accuracy of 98.18%, outperforming traditional machine learning models, existing GNNs, and dual-graph approaches in precision, recall, and F1-score, while showing robust generalization under incomplete features. These results highlight the model’s capability in accurately identifying and modeling complex drainage structures, offering a practical and generalizable framework for geospatial hydrological applications.