A graph Transformer method for river network pattern recognition incorporating three-dimensional terrain features

doi:10.11947/j.AGCS.2026.20260139

Abstract

Abstract:

This study proposes a graph Transformer method for river network pattern recognition incorporating three-dimensional terrain features. First, a dual graph is constructed from vector drainage network data, in which river segments are represented as graph nodes and the topological connections between river segments are represented as graph edges. Second, in addition to two-dimensional topological and geometric morphological features, three-dimensional terrain features, including elevation difference, slope, and channel gradient, are introduced to enhance the representation of the terrain context associated with river-segment nodes. Third, GraphSAGE is employed to sample and aggregate neighborhood information, thereby learning local confluence relationships as well as morphological and terrain differences between adjacent river segments. A Transformer-based self-attention mechanism is then introduced to model global structural dependencies within the drainage network graph, enabling the collaborative representation of local neighborhood relationships and overall spatial structural patterns. Finally, graph classification is performed to identify five drainage patterns: dendritic, fan-shaped, parallel, skeletal, and grid-like patterns. Experimental results show that the proposed method achieves an accuracy of 95.36% in recognizing the five drainage patterns, outperforming GCN, GraphSAGE, and the Transformer methods. The proposed method provides an approach to drainage network spatial structure representation, watershed classification, and hydrological process simulation.

Key words: Transformer, GraphSAGE, river network pattern recognition, terrain features

CLC Number:

P208

Runze Chai, Jingzhong Li, Haowen Yan, Xiaolong Wang, Zhenyue Liu, Ling Wang. A graph Transformer method for river network pattern recognition incorporating three-dimensional terrain features[J]. Acta Geodaetica et Cartographica Sinica, 2026, 55(9): 1695-1705.

Add to citation manager EndNote|Reference Manager|ProCite|BibTeX|RefWorks

URL: http://xb.chinasmp.com/EN/10.11947/j.AGCS.2026.20260139

http://xb.chinasmp.com/EN/Y2026/V55/I9/1695

Figures/Tables 13

Fig. 1

Fig. 2

Fig. 3

Fig. 4

Fig. 5

Fig. 6

Fig. 7

Fig. 8

Tab. 1

Tab. 2

Tab. 3

Tab. 4

Tab. 5

References 24

[1]	艾廷华, 刘耀林, 黄亚锋. 河网汇水区域的层次化剖分与地图综合[J]. 测绘学报, 2007, 36(2): 231-236, 243.
	Ai Tinghua, Liu Yaolin, Huang Yafeng. The hierarchical watershed partitioning and generalization of river network[J]. Acta Geodaetica et Cartographica Sinica, 2007, 36(2): 231-236, 243.
[2]	刘呈熠, 武芳, 巩现勇, 等. 规则池塘群的复杂直线模式识别[J]. 测绘学报, 2020, 49(2): 256-266. DOI: . doi: 10.11947/j.AGCS.2020.20190030
	Liu Chengyi, Wu Fang, Gong Xianyong, et al. Complex linear pattern recognition for regular pond groups[J]. Acta Geodaetica et Cartographica Sinica, 2020, 49(2): 256-266. DOI: . doi: 10.11947/j.AGCS.2020.20190030
[3]	王文宁. 多形态河系的流域基本单元特征分析、识别及相似度计算[D]. 兰州: 兰州交通大学, 2023.
	Wang Wenning. Feature analysis, recognition and similarity calculation of watershed basic units in multi-morphology river systems[D]. Lanzhou: Lanzhou Jiaotong University, 2023.
[4]	Yu Huafei, Ai Tinghua, Yang Min, et al. A recognition method for drainage patterns using a graph convolutional network[J]. International Journal of Applied Earth Observation and Geoinformation, 2022, 107: 102696.
[5]	Strahler A N. Quantitative analysis of watershed geomorphology[J]. Eos, Transactions American Geophysical Union, 1957, 38(6): 913-920.
[6]	张园玉, 李霖, 金玉平, 等. 基于图论的树状河系结构化绘制模型研究[J]. 武汉大学学报(信息科学版), 2004, 29(6): 537-539, 543.
	Zhang Yuanyu, Li Lin, Jin Yuping, et al. Structured design of dendritic river networks based on graph[J]. Geomatics and Information Science of Wuhan University, 2004, 29(6): 537-539, 543.
[7]	杜清运, 杨品福, 谭仁春. 基于空间统计特征的河网结构分类[J]. 武汉大学学报(信息科学版), 2006, 31(5): 419-422.
	Du Qingyun, Yang Pinfu, Tan Renchun. Classification of river networks structure based on spatial statistical character[J]. Geomatics and Information Science of Wuhan University, 2006, 31(5): 419-422.
[8]	Liu Chengyi, Zhai Renjian, Qian Haizhong, et al. Identification of drainage patterns using a graph convolutional neural network[J]. Transactions in GIS, 2023, 27(3): 752-776.
[9]	Wang Wenning, Yan Haowen, Lu Xiaomin, et al. Drainage pattern recognition method considering local basin shape based on graph neural network[J]. International Journal of Digital Earth, 2023, 16(1): 593-619.
[10]	Yu Huafei, Ai Tinghua, Yang Min, et al. A graph autoencoder network to measure the geometric similarity of drainage networks in scaling transformation[J]. International Journal of Digital Earth, 2023, 16(1): 1828-1852.
[11]	翟仁健, 武芳, 邓红艳, 等. 基于遗传多目标优化的人工河网自动选取模型[J]. 测绘学报, 2008, 37(1): 108-113.
	Zhai Renjian, Wu Fang, Deng Hongyan, et al. An automated selection model of ditch based on multi-objective optimization by genetic algorithm[J]. Acta Geodaetica et Cartographica Sinica, 2008, 37(1): 108-113.
[12]	邵黎霞, 何宗宜, 艾自兴, 等. 基于BP神经网络的河系自动综合研究[J]. 武汉大学学报(信息科学版), 2004, 29(6): 555-557.
	Shao Lixia, He Zongyi, Ai Zixing, et al. Automatic generalization of river network based on BP neural network techniques[J]. Geomatics and Information Science of Wuhan University, 2004, 29(6): 555-557.
[13]	黄哲琨, 钱海忠, 蔡中祥, 等. 基于图神经网络的多尺度网状河系分类匹配方法[J]. 测绘学报, 2025, 54(2): 371-384. DOI: . doi: 10.11947/j.AGCS.2025.20240145
	Huang Zhekun, Qian Haizhong, Cai Zhongxiang, et al. A multi-scale mesh river system classification matching method based on graph neural network[J]. Acta Geodaetica et Cartographica Sinica, 2025, 54(2): 371-384. DOI: . doi: 10.11947/j.AGCS.2025.20240145
[14]	Zhou Jie, Cui Ganqu, Hu Shengding, et al. Graph neural networks: a review of methods and applications[J]. AI Open, 2020, 1: 57-81.
[15]	王米琪. 基于图卷积神经网络的道路网模式识别[D]. 武汉: 武汉大学, 2021.
	Wang Miqi. Road network pattern recognition based on graph convolutional neural network[D]. Wuhan: Wuhan University, 2021.
[16]	刘慧敏, 胡文柯, 唐建波, 等. 顾及功能语义特征的建筑物空间分布模式识别方法[J]. 测绘学报, 2020, 49(5): 622-631. DOI: . doi: 10.11947/j.AGCS.2020.20190222
	Liu Huimin, Hu Wenke, Tang Jianbo, et al. A method for recognizing building clusters by considering functional features of buildings[J]. Acta Geodaetica et Cartographica Sinica, 2020, 49(5): 622-631. DOI: . doi: 10.11947/j.AGCS.2020.20190222
[17]	Yan Xiongfeng, Ai Tinghua, Yang Min, et al. A graph convolutional neural network for classification of building patterns using spatial vector data[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2019, 150: 259-273.
[18]	Xu Xiaofeng, Liu Pengcheng, Guo Mingwu. Drainage pattern recognition of river network based on graph convolutional neural network[J]. ISPRS International Journal of Geo-Information, 2023, 12(7): 253.
[19]	Yu Huafei, Ai Tinghua, Yang Min, et al. Integrating domain knowledge and graph convolutional neural networks to support river network selection[J]. Transactions in GIS, 2023, 27(7): 1898-1927.
[20]	Yu Huafei, Ai Tinghua, Yang Min, et al. Automatic segmentation of parallel drainage patterns supported by a graph convolution neural network[J]. Expert Systems with Applications, 2023, 211: 118639.
[21]	Qiang Bo, Liu Tao, Du Ping, et al. Drainage pattern recognition method using graph convolutional networks combined with three-dimensional elevation features[J]. Transactions in GIS, 2025, 29: e70002.
[22]	王亚青, 王中辉. 异构图卷积网络支持下的河系自动选取方法[J]. 测绘学报, 2025, 54(7): 1332-1345. DOI: . doi: 10.11947/j.AGCS.2025.20240337
	Wang Yaqing, Wang Zhonghui. River network automated selection method based on heterogeneous graph convolutional networks[J]. Acta Geodaetica et Cartographica Sinica, 2025, 54(7): 1332-1345. DOI: . doi: 10.11947/j.AGCS.2025.20240337
[23]	余华飞, 邱天奇, 周哲, 等. 图Transformer支持下的河网模式识别[J]. 测绘学报, 2024, 53(11): 2075-2085. DOI: . doi: 10.11947/j.AGCS.2024.20230582
	Yu Huafei, Qiu Tianqi, Zhou Zhe, et al. Drainage pattern recognition supported by graph Transformer[J]. Acta Geodaetica et Cartographica Sinica, 2024, 53(11): 2075-2085. DOI: . doi: 10.11947/j.AGCS.2024.20230582
[24]	Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[C]//Proceedings of 2017 Advances in Neural Information Processing Systems. Long Beach: Curran Associates, 2017: 5998-6008.

方法	准确率/（%）	召回率/（%）	F₁值/（%）	训练时间/min
GCN	85.52	84.24	84.61	4.32
GraphSAGE	89.57	88.21	89.36	5.18
Transformer	92.25	91.82	92.03	7.64
本文方法	95.36	95.08	95.24	8.52

编号	类别	方法	模式概率					预测类别
编号	类别	方法	树状	扇状	平行状	骨架状	格网状	预测类别
（a）	扇状	GCN	0.142	0.386	0.421	0.039	0.012	平行状
		GraphSAGE	0.118	0.514	0.319	0.036	0.013	扇状
		Transformer	0.071	0.761	0.126	0.030	0.012	扇状
		本文方法	0.048	0.806	0.116	0.019	0.011	扇状
（b）	骨架状	GCN	0.248	0.054	0.038	0.577	0.083	骨架状
		GraphSAGE	0.621	0.046	0.034	0.209	0.090	树状
		Transformer	0.184	0.037	0.026	0.701	0.052	骨架状
		本文方法	0.096	0.026	0.018	0.833	0.027	骨架状
（c）	树状	GCN	0.316	0.109	0.026	0.538	0.011	骨架状
		GraphSAGE	0.672	0.092	0.023	0.201	0.012	树状
		Transformer	0.748	0.061	0.017	0.164	0.010	树状
		本文方法	0.879	0.038	0.011	0.062	0.010	树状
（d）	平行状	GCN	0.071	0.392	0.384	0.083	0.070	扇状
		GraphSAGE	0.053	0.315	0.507	0.060	0.065	平行状
		Transformer	0.031	0.216	0.719	0.018	0.016	平行状
		本文方法	0.018	0.126	0.835	0.010	0.011	平行状
（e）	格网状	GCN	0.028	0.034	0.122	0.052	0.764	格网状
		GraphSAGE	0.020	0.025	0.091	0.040	0.824	格网状
		Transformer	0.012	0.014	0.053	0.025	0.896	格网状
		本文方法	0.008	0.010	0.031	0.014	0.937	格网状

试验设置	GraphSAGE局部学习模块	Transformer全局学习模块	三维地形特征	准确率/（%）	F₁值/（%）
去除Transformer全局学习模块	√	×	√	89.57	89.36
去除GraphSAGE局部学习模块	×	√	√	92.25	92.03
去除三维地形特征	√	√	×	93.13	92.86
本文方法	√	√	√	95.36	95.24

特征组合	准确率	召回率	F₁值
无地形特征	93.13	92.75	92.86
高程差	93.78	93.52	93.71
坡度	93.85	93.65	93.73
比降	94.08	93.95	93.84
高程差-坡度	94.47	94.28	94.03
高程差-比降	94.16	93.98	93.72
坡度-比降	94.92	94.76	94.48
高程差-坡度-比降	95.36	95.08	95.24

区域	准确率	召回率	F₁值
原始数据集	95.36	95.08	95.24
长江流域	95.19	94.75	94.88
澜沧江流域	95.27	94.84	95.20