A self-supervised matching method for polygonal geographic entity based on a three-branch attention network

doi:10.11947/j.AGCS.2026.20250296

Abstract

Abstract:

Geographic entity are important data outputs of 3D realistic geospatial scenes. Due to scale differences in the spatial expression of polygonal entities with the same name across datasets from different scenarios, geographic entity matching technology is required to support data fusion and updating. Aiming at the current situation that existing polygonal entity matching methods still have room for optimization in breaking away from manual dependence and realizing refined and differentiated modeling of multiple matching relationships (1∶1, 1∶M, M∶N), this paper proposes a self-supervised matching method for polygonal geographic entity based on a three-branch attention network. First, it calculates the similarity of four types of features: size, distance, shape, and direction. For each type of feature, the standard deviation is computed based on the number of matched entity pairs under different thresholds. A loss function is constructed using this standard deviation to train the model and obtain decision thresholds, and entity pairs whose similarity meets the decision thresholds are converted into pseudo-labels. Second, a matching network with a three-branch structure is built to handle 1∶1, 1∶M, and M∶N matching relationships respectively. This network integrates the attention mechanism and gradient-weighted class activation mapping (Grad-CAM) to adaptively assign weights to each feature. Finally, two types of data—construction land and engineering project land in Huangshan city, Anhui province—are selected as experimental data to conduct verification on the pseudo-labels, feature weight assignment, and the three-branch network framework respectively. The experimental results show that compared with existing methods, the proposed method in this paper does not require manual annotation, can adaptively achieve matching for multiple matching relationships (1∶1, 1∶M, M∶N), and achieves a precision (P) of 94.98%, recall (R) of 94.22%, and F₁ score of 94.60%. Its effectiveness is verified, and it can provide strong support for the fusion and updating of polygonal geographic entity data.

Key words: polygonal geographic entity matching, pseudo-labels, self-supervised learning, attention mechanism

CLC Number:

P208

Zhuang SUN, Po LIU, Liang ZHAI, Yu HE, Zutao ZHANG. A self-supervised matching method for polygonal geographic entity based on a three-branch attention network[J]. Acta Geodaetica et Cartographica Sinica, 2026, 55(1): 169-180.

Figures/Tables 16

Fig. 1

Fig. 2

Fig. 3

Fig. 4

Fig. 5

Fig. 6

Fig. 7

Fig. 8

Fig. 9

Fig. 10

Tab. 1

Tab. 2

Fig. 11

Fig. 12

Tab. 3

Fig. 13

References 30

[1]	燕琴, 翟亮, 刘坡. 实景三维中国建设关键技术研究综述[J]. 测绘科学, 2023, 48(7): 1-9.
	YAN Qin, ZHAI Liang, LIU Po. Review on the key technologies of 3D real scene construction in China[J]. Science of Surveying and Mapping, 2023, 48(7): 1-9.
[2]	曹全龙, 王会娜, 杜萌, 等. 地理实体模型构建与表达研究[J]. 时空信息学报, 2023, 30(1): 64-69.
	CAO Quanlong, WANG Huina, DU Meng, et al. Research on construction and representation of geo-entity model[J]. Journal of Spatio-temporal Information, 2023, 30(1): 64-69.
[3]	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
[4]	张祖涛, 刘坡, 翟亮, 等. 黄山市地理实体统一标识体系设计与实现[J]. 测绘科学, 2025, 50(6): 157-167.
	ZHANG Zutao, LIU Po, ZHAI Liang, et al. Design and implementation of unified identification system of basic geographic entity in mount Huangshan city[J]. Journal of Spatio-temporal Information, 2025, 50(6): 157-167.
[5]	FU Z, SUN Y, FAN L, et al. Multiscale and multifeature segmentation of high-spatial resolution remote sensing images using superpixels with mutual optimal strategy[J]. Remote Sensing, 2018, 10(8): 1289.
[6]	郝燕玲, 唐文静, 赵玉新, 等. 基于空间相似性的面实体匹配算法研究[J]. 测绘学报, 2008, 37(4): 501-506.
	HAO Yanling, TANG Wenjing, ZHAO Yuxin, et al. Polygonal feature matching algorithm based on spatial similarity[J]. Acta Geodaetica et Cartographica Sinica, 2008, 37(4): 501-506.
[7]	刘坡, 张宇, 龚建华. 中误差和邻近关系的多尺度面实体匹配算法研究[J]. 测绘学报, 2014, 43(4): 419-425.
	LIU Po, ZHANG Yu, GONG Jianhua. Root mean square error and neighbouring relation matching approach for multi-scale polygonal feature[J]. Acta Geodaetica et Cartographica Sinica, 2014, 43(4): 419-425.
[8]	石善球. 基于几何与属性匹配的地理信息数据融合更新方法[J]. 自然资源遥感, 2023, 35(1): 251-257.
	SHI Shanqiu. A geographic data fusion and update method based on geometric and attribute matching[J]. Remote Sensing for Natural Resources, 2023, 35(1): 251-257.
[9]	赵东保, 盛业华, 张卡. 利用几何矩和叠置分析进行多尺度面要素自动匹配[J]. 武汉大学学报(信息科学版), 2011, 36(11): 1371-1375.
	ZHAO Dongbao, SHENG Yehua, ZHANG Ka. An algorithm for muti-scale one-to-many polygonal feature matching based on geometry moments and overly analysis[J]. Geomatics and Information Science of Wuhan University, 2011, 36(11): 1371-1375.
[10]	刘凌佳, 朱道也, 朱欣焰, 等. 基于MBR组合优化算法的多尺度面实体匹配方法[J]. 测绘学报, 2018, 47(5): 652-662. DOI: . doi: 10.11947/j.AGCS.2018.20160625
	LIU Lingjia, ZHU Daoye, ZHU Xinyan, et al. A multi-scale polygonal object matching method based on MBR combinatorial optimization algorithm[J]. Acta Geodaetica et Cartographica Sinica, 2018, 47(5): 652-662. DOI: . doi: 10.11947/j.AGCS.2018.20160625
[11]	刘贺, 郭黎, 李豪, 等. 面实体匹配的集成学习CatBoost方法[J]. 地球信息科学学报, 2022, 24(11): 2198-2211.
	LIU He, GUO Li, LI Hao, et al. Matching polygonal entities with CatBoost ensemble method[J]. Journal of Geo-information Science, 2022, 24(11): 2198-2211.
[12]	ZHU D, CHENG C, ZHAI W, et al. Multiscale spatial polygonal object granularity factor matching method based on BPNN[J]. ISPRS International Journal of Geo-Information, 2021, 10(2): 75. DOI: . doi: 10.3390/ijgi/0020075
[13]	FAN Dazhao, DONG Yang, ZHANG Yongsheng. Satellite image matching method based on deep convolutional neural network[J]. Journal of Geodesy and Geoinformation Science, 2019, 2(2): 90-100.
[14]	ALBELWI S. Survey on self-supervised learning: auxiliary pretext tasks and contrastive learning methods in imaging[J]. Entropy, 2022, 24(4): 551. DOI: . doi: 10.3390/e24040551
[15]	RANI V, NABI S T, KUMAR M, et al. Self-supervised learning: a succinct review[J]. Archives of Computational Methods in Engineering, 2023, 30(4): 2761-2775.
[16]	CARON M, BOJANOWSKI P, JOULIN A, et al. Deep clustering for unsupervised learning of visual features[C]//Proceedings of 2018 European conference on computer vision (ECCV). Munich: Springer, 2018: 132-149.
[17]	LEE D H. Pseudo-label: the simple and efficient semi-supervised learning method for deep neural networks[C]//Proceedings of 2013 Workshop on Challenges in Representation Learning. Atlanta: ICML.2013, 3(2): 896.
[18]	RADFORD A, KIM J W, HALLACY C, et al. Learning transferable visual models from natural language supervision[C]//Proceedings of 2021 International Conference on Machine Learning. [S.l.]: PmLR, 2021: 8748-8763.
[19]	ZOU Y, YU Z, KUMAR B V K, et al. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training[C]//Proceedings of 2018 European Conference on Computer Vision (ECCV). Munich: Springer, 2018: 289-305.
[20]	LI M, LI Q, WANG Y. Class balanced adaptive pseudo labeling for federated semi-supervised learning[C]//Proceedings of 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Vancouver: IEEE, 2023: 16292-16301.
[21]	TARVAINEN A, VALPOLA H. Mean teachers are better role models: weight-averaged consistency targets improve semi-supervised deep learning results[J]. Advances in Neural Information Processing Systems, 2017: 30. DOI: . doi: 10.48550/arXiv.1703.01780
[22]	XU M C, ZHOU Y, JIN C, et al. Bayesian pseudo labels: expectation maximization for robust and efficient semi-supervised segmentation[C]//Proceedings of 2022 International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2022: 580-590.
[23]	LI L, SHUM H P H, BRECKON T P. Less is more: reducing task and model complexity for 3D point cloud semantic segmentation[C]//Proceedings of 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Vancouver: IEEE, 2023: 9361-9371.
[24]	SZEGEDY C, LIU W, JIA Y, et al. Going deeper with convolutions[C]//Proceedings of 2015 IEEE Conference on Computer Vision and Pattern Recognition. Boston: IEEE, 2015: 1-9.
[25]	ASLANI S, DAYAN M, STORELLI L, et al. Multi-branch convolutional neural network for multiple sclerosis lesion segmentation[J]. NeuroImage, 2019, 196: 1-15.
[26]	WU J, POLOCZEK M, WILSON A G, et al. Bayesian optimization with gradients[J]. Advances in Neural Information Processing Systems, 2017: 30. DOI: . doi: 10.48550/arXiv.1703.04389
[27]	FEDUS W, ZOPH B, SHAZEER N. Switch transformers: scaling to trillion parameter models with simple and efficient sparsity[J]. Journal of Machine Learning Research, 2022, 23(120): 1-39.
[28]	WEI R, FAN B, WANG Y, et al. MBNet: multi-branch network for extraction of rural homesteads based on aerial images[J]. Remote Sensing, 2022, 14(10): 2443. DOI: . doi: 10.3390/rs14102443
[29]	KANUNGO T, MOUNT D M, NETANYAHU N S, et al. An efficient K-means clustering algorithm: analysis and implementation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2002, 24(7): 881-892.
[30]	SELVARAJU R R, COGSWELL M, DAS A, et al. Grad-cam: visual explanations from deep networks via gradient-based localization[C]//Proceedings of 2017 IEEE International Conference on Computer Vision. Venice: IEEE, 2017: 618-626.

数据类别	训练数据	测试数据
自然幢实体数量	9094	1377
建筑用地实体数量	2343	612
1∶1数量	482	344
1∶M数量	1325	166
M∶N数量	24	17

方法	匹配关系	TP	FP	FN	P/（%）	R/（%）	F₁值/（%）
经验阈值	全部	450	40	37	91.84	92.40	92.12
	1∶1	302	29	13	91.24	95.87	93.50
	1∶M	136	8	22	94.44	86.08	90.07
	M∶N	12	3	2	80.00	85.71	82.76
CatBoost	全部	459	23	45	94.92	93.03	93.96
	1∶1	307	11	31	95.91	92.15	93.99
	1∶M	149	10	12	93.71	95.51	94.60
	M∶N	13	2	2	86.67	86.67	86.67
BPNN	全部	448	46	33	90.69	93.14	91.90
	1∶1	293	26	25	91.85	92.14	91.99
	1∶M	143	17	6	89.38	95.97	92.56
	M∶N	12	3	2	80.00	85.71	82.76
本文方法	全部	473	25	29	94.98	94.22	94.60
	1∶1	303	16	25	94.98	92.38	93.66
	1∶M	156	8	2	95.12	98.73	96.89
	M∶N	14	1	2	93.33	87.50	90.32

标签	匹配关系	TP	FP	FN	P/（%）	R/（%）	F₁值/（%）
真实标签	全部	488	22	17	95.69	96.63	96.16
	1∶1	313	16	15	95.14	95.43	95.28
	1∶M	160	5	1	96.97	99.38	98.16
	M∶N	15	1	1	93.75	93.75	93.75
伪标签	全部	473	25	29	94.98	94.22	94.60
	1∶1	303	16	25	94.98	92.38	93.66
	1∶M	156	8	2	95.12	98.73	96.89
	M∶N	14	1	2	93.33	87.50	90.32