A Skeleton-Line-Based Graph Convolutional Neural Network for Areal Settlements’ Shape Classification

Abstract

Among the geographic elements, shape recognition and classification is one of the im portant elements of map cartographic generalization, and the shape classification of an areal settlement is an important part of geospatial vector data. However, there is currently no relatively simple and efficient way to achieve areal settlement classification. Therefore, we combined the skeleton line vector data of an areal settlement and the graph convolutional neural network to propose an areal settlement shape classification method that (1) extracts the skeleton line of the areal settlement to form a dual graph with nodes as edges, (2) extracts multiple features to obtain a graph representation of the shape, (3) extracts and aggregates the shape information represented by the areal settlement skeleton line using the graph convolutional neural network for multiple rounds to extract high-dimensional shape information, and (4) completes the shape classification of the high-dimensional shape information. The experiment used 240 samples, and the classification accuracy was 93.3%, with areal settlement shapes of E-, F-, and H-type achieving F-measures of 96.5%, 92.3%, and 100%, respectively. The result shows that the classification method of the areal settlement shape has high accuracy.