Drainage Pattern Recognition Combined With Second‐Order Co‐Occurrence Matrices and Global Features in CNN Networks

Drainage pattern recognition is an important research problem in terrain knowledge mining, map generalization, and other fields. Accurate identification of river network patterns is helpful to better understand geographical phenomena and optimize the quality of cartography. In this study, a convolutional neural network (CNN) model based on drainage network co‐occurrence matrix and global features is developed to accurately identify drainage patterns. The characteristics of vector river network line elements are quantified efficiently. Co‐occurrence matrix is introduced into local pattern analysis to capture the spatial proximity between river segments and generate multiple co‐occurrence matrices representing the direction and attribute combination of river network. The co‐occurrence matrix is combined with the global properties of the drainage network as the input feature vector of the CNN model. Through the training of a large number of samples and optimization of the network structure, a CNN model specifically for drainage network pattern recognition is formed. To validate the effectiveness of the model, river data from the Boise area, the capital of Idaho, USA, were used for testing and compared with the tested results of the graph‐convolution recognition model. The experimental results show that the present model exhibits a high degree of accuracy and efficiency in river network pattern recognition and also confirms the value of the second‐order co‐occurrence matrix as an effective metric tool for unstructured spatial patterns.