关键词: 地图综合, 河系选取, 深度学习, 河系异构图, RGCN

Abstract:

River network selection is a map generalization process in which important rivers are selected and other rivers are discarded, due to space limitations when scaling down from large-scale to small-scale maps. Traditional deep learning methods typically focus on homogeneous graphs with a single type of relationship between river segments, which limits their ability to fully utilize the complex connectivity information between segments. This often results in lower selection accuracy and compromised topological connectivity. To address these issues, this paper introduces an automated river network selection method based on heterogeneous graph convolutional networks. In this method, river segments are represented as nodes, and their connections as edges. These edges are categorized into three types based on different relationship characteristics, creating a heterogeneous graph of the river network. The river network data and corresponding selection labels are input into the relational graph convolutional networks (RGCN) model, which aggregates information and classifies the nodes. River segments are selected based on the classification results, achieving automation in the selection process. Experiments using river network data at scales of 1∶24 000 and 1∶250 000 show that the proposed method significantly improves selection accuracy. Key performance metrics, including precision, recall, F₁ score and AUC, all exceed 92%. Additionally, the method reduces river network discontinuities and better preserves the topological connectivity and shape similarity of the river network.

Key words: map generalization, river network selection, deep learning, heterogeneous graph of the river network, RGCN