Gated Sparse Graph Transformer integrating topographic and lithological features for drainage pattern recognition

Drainage pattern recognition is valuable for watershed hydrological simulation and geological structure exploration. Most existing methods merely focus on topological structures and geometric forms of drainage, and fail to balance global and local features. Accordingly, we formulate drainage pattern recognition as a graph-classification task by representing vector river networks as directed graphs and integrating hierarchical, geometric, basin, topographic, and lithological features into a Gated Sparse Graph Transformer (GSGT) that jointly captures local topological patterns and long-range structural dependencies. The GSGT combines GatedGCN for local feature extraction and sparse multi-head attention to model long-range dependencies. Experimental results demonstrate that lithology has a prominent correlation with the development of drainage patterns and serves as a critical feature for improving model performance. Compared with baseline algorithms, the GSGT reaches an accuracy of 95.72% on the US dataset and 92.32% in generalization tests, offering a technical solution for intelligent automatic recognition of drainage patterns.