A recognition method for drainage patterns using a graph convolutional network

Drainage pattern recognition (DPR) is a classic and challenging problem in hydrographic system analysis, topographical knowledge mining, and map generalization. An outstanding issue for traditional DPR methods is that the rules used to extract patterns based on certain geometric measures are limited, not accessing the effects of manual recognition. In this study, a graph convolutional network (GCN) was introduced for DPR. First, a dual graph of drainage was built based on the channel connection and hierarchical structure after constructing typical sample data. Second, its features were extracted as inputs of the GCN from three scales, namely, global unity at a macroscale, hierarchical connectivity at a mesoscale, and local equilibrium at a microscale. Finally, the model architecture based on the GCN was designed for DPR. Typical pattern samples (i.e. dendritic, distributary, parallel, skeleton, and rectangular drainage) from OpenStreetMap and USGS were used to implement the training and testing of the model, respectively. The results show that our approach outperforms other machine learning methods, including convolutional neural network, with an accuracy of 85.0%. In summary, the GCN has considerable potential for DPR and a wide scope for further improvement.

Automated summary

Drainage pattern recognition (DPR) is a challenging problem in various fields. In this study, a graph convolutional network (GCN) was introduced to improve traditional DPR methods by considering manual recognition effects. The proposed approach achieved higher accuracy compared to other machine learning methods, demonstrating the potential and room for improvement of GCN in DPR.