Automatic segmentation of parallel drainage patterns supported by a graph convolution neural network. (January 2023)

• Record Type: Journal Article
• Authors: Yu, Huafei; Ai, Tinghua; Yang, Min; Huang, Lina; Gao, Aji
• Is Part Of: Expert systems with applications. Volume 211(2023)
• Publication Date: 2023-01
• Subjects: Parallel drainage pattern -- Automatic segmentation -- Graph theory -- Drainage features -- Graph sample and aggregate; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.118639
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: GraphSAGE was applied in the segmentation of parallel drainage pattern (SPDP). A method of building drainage dual graph (DDG) with hydrology knowledge is proposed. The GraphSAGE outperforms other machine learning methods and GCNNs in SPDP. The knowledge-based DDG reduces the quantity of DP samples for training. Abstract: Drainage pattern (DP) recognition is critical in hydrographic analysis, topography identification, and drainage characteristic detection. The traditional method is based on rule computation and self-similarity idea preliminarily performing the DP classification. However, DP segmentation is an uncertain spatial cognitive problem affected by enormous factors. To settle such a multi-conditions decision question, this study takes the segmentation of parallel drainage pattern (SPDP) as an example presenting a deep learning method, namely the graph convolution neural network (GCNN) based on Graph SAmple and aggreGatE (GraphSAGE). First, a directed graph and dual graph were used to construct a dual drainage graph recording spatial-cognition features of drainage. Second, nine drainage features were built to define the graph description from three perspectives: topological connectivity, meandering equilibrium, and directional unity. Finally, the GraphSAGE model was designed for SPDP and trained by typical samples to finish the segmentation works. The experiment examined the optimal feature combination and hyperparameter sensitivity, which can provide sufficient information for SPDP supported by GraphSAGE. Besides, our model outperformed other machine learning methods and GCNNs driven by a fixed quantity sampling mechanism and hydrological knowledge. This work provides a vital reference for hydrology research supported by combing hydrological knowledge with GCNNs.