Machine learning study demonstrates high-accuracy segmentation of complex river networks using local basin graphs, highlighting improved hydrological mapping and land-use analysis.
Key Points
To automate the segmentation of complex skeleton drainage patterns by developing a graph neural network framework built upon local basin graphs.
Constructed a watershed graph structure using catchment adjacency relationships to encode spatial topology and hierarchical river dependencies.
Integrated basin connectivity, reach geometry, and elevation variability into a GraphSAGE model utilizing a Mean aggregation strategy.
Achieved a validation accuracy of 98.18% for skeleton drainage pattern segmentation.
Surpassed traditional machine learning models, baseline graph neural networks, and dual-graph frameworks across precision, recall, and F1-score, while retaining robustness under incomplete features.