基于神经网络的水系制图综合研究

Comprehensive Research on Water System Mapping Based on Neural Networks

摘要: 当前水系形态模式识别多依赖传统几何特征法，难以充分利用拓扑信息，样本有限或形态相似时泛化能力与精度受限，然而现有研究缺乏融合拓扑结构与多特征、借助数据增强提升分类稳定性的方法。本研究以OpenStreetMap与全国基础地理信息库河流数据为基础，构建含拓扑、几何及属性特征的图结构，结合图神经网络与数据增强技术识别水系形态，目标为构建标准化图结构数据、设计GCN分类模型、引入旋转与镜像增强策略。实验显示，增强后模型总体识别准确率达82.6%，较初始78.06%显著提升；辫状与网状水系易识别，平行状与放射状因样本少、形态相似难分类；各样本组准确率服从正态分布，验证方法稳定性。该研究为水系自动化识别提供新路径，可为流域规划等提供支撑。

Abstract: Current watercourse morphology pattern recognition heavily relies on traditional geometric feature methods, which struggle to fully leverage topological information. When samples are limited or morphologies are similar, their generalization capability and accuracy are constrained. However, existing research lacks methods that integrate topological structures with multiple features and utilize data augmentation to enhance classification stability. This study constructs a graph structure incorporating topological, geometric, and attribute features based on OpenStreetMap and national river data from the Basic Geographic Information Database. It combines graph neural networks with data augmentation techniques to identify watercourse morphology, aiming to establish standardized graph structure data, design a GCN classification model, and introduce rotation and mirroring augmentation strategies. Experiments demonstrate that the enhanced model achieves an overall recognition accuracy of 82.6%, a significant improvement over the initial 78.06%. Braided and networked water systems are easily identifiable, while parallel and radial systems are challenging to classify due to limited samples and morphological similarity. The accuracy rates across all sample groups follow a normal distribution, validating the method’s stability. This research provides a new pathway for automated water system recognition and can support applications such as watershed planning.