一种面向地图综合建筑多边形化简的Transformer模型

doi:10.11947/j.AGCS.2026.20250262

测绘学报 ›› 2026, Vol. 55 ›› Issue (1): 124-137.doi: 10.11947/j.AGCS.2026.20250262

• 地图学与地理信息 • 上一篇

一种面向地图综合建筑多边形化简的Transformer模型

刘鹏程¹^,²(), 成晓强³, 肖天元⁴, 杨敏⁴, 艾廷华⁴

^1.华中师范大学地理过程分析与模拟湖北省重点实验室，湖北　武汉　430079
^2.华中师范大学城市与环境科学学院，湖北　武汉　430079
^3.湖北大学资源环境学院，湖北　武汉　430062
^4.武汉大学资源与环境科学学院，湖北　武汉　430079

收稿日期:2025-06-30 修回日期:2026-01-05 发布日期:2026-02-13
作者简介:刘鹏程（1968—），男，博士，教授，研究方向为地图综合、空间模式识别和空间智能分析。E-mail：liupc@ccnu.edu.cn
基金资助:
国家自然科学基金(42471486; 42071455);中央高校基本科研业务费专项资金(CCNU25JC043)

A Transformer model for building polygon simplification in map generalization

Pengcheng LIU¹^,²(), Xiaoqiang CHENG³, Tianyuan XIAO⁴, Min YANG⁴, Tinghua AI⁴

^1.Key Laboratory for Geographical Process Analysis & Simulation of Hubei Province, Central China Normal University, Wuhan 430079, China
^2.School of Urban and Environmental Sciences, Central China Normal University, Wuhan 430079, China
^3.Faculty of Resources and Environmental Science, Hubei University, Wuhan 430062, China
^4.School of Resources and Environmental Sciences, Wuhan University, Wuhan 430079, China

Received:2025-06-30 Revised:2026-01-05 Published:2026-02-13
About author:LIU Pengcheng (1968—), male, PhD, professor, majors in map generalization, spatial pattern recognition and GeoAI. E-mail: liupc@ccnu.edu.cn
Supported by:
The National Natural Science Foundation of China(42471486; 42071455);The Fundamental Research Funds for the Central Universities(CCNU25JC043)

摘要/Abstract

摘要：

针对地图综合中建筑多边形化简方法依赖人工规则、自动化程度低且难以利用已有化简成果的问题，本文提出了一种基于Transformer机制的建筑多边形化简模型。该模型首先把建筑多边形映射至一定范围的网格空间，将建筑多边形的坐标串表达为网格序列，从而获取建筑多边形化简前后的Token序列，构建出建筑多边形化简样本对数据；随后采用Transformer架构建立模型，基于样本数据利用模型的掩码自注意力机制学习点序列之间的依赖关系，最终逐点生成新的简化多边形，从而实现建筑多边形的化简。在训练过程中，模型使用结构化的样本数据，设计了忽略特定索引的交叉熵损失函数以提升化简质量。试验设计包括主试验与泛化验证两部分。主试验基于洛杉矶1∶2000建筑数据集，分别采用0.2、0.3和0.5 mm 3种网格尺寸对多边形进行编码，实现了目标比例尺为1∶5000与1∶10 000的化简。试验结果表明，在0.3 mm的网格尺寸下模型性能最优，验证集上的化简结果与人工标注的一致率超过92.0%，且针对北京部分区域的建筑多边形数据的泛化试验验证了模型的迁移能力；与LSTM模型的对比分析显示，在参数规模相近的条件下，LSTM模型无法形成有效收敛，并生成可用结果。本文证实了Transformer在处理空间几何序列任务中的潜力，且能够有效复用已有化简样本，为智能建筑多边形化简提供了具有工程实用价值的途径。

关键词: 地图综合, 建筑多边形化简, Tokenization, Transformer模型, 上下文工程

Abstract:

Addressing the issues in building polygon simplification within map generalization—such as reliance on manual rules, low automation, and difficulty in reusing existing simplification results—this paper proposes a building polygon simplification model based on the Transformer mechanism. The model begins by mapping building polygons into a grid space of a certain scale, representing coordinate strings of the polygons as grid sequences. This process allows the acquisition of Token sequences before and after simplification, thereby constructing paired datasets of building polygon simplification samples. Utilizing the Transformer architecture, the model learns dependencies between point sequences through its masked self-attention mechanism, ultimately generating new simplified polygons point by point to achieve building polygon simplification. During training, the model employs structured sample data and incorporates a cross-entropy loss function that ignores specific indices to enhance simplification quality. The experimental design consists of two parts: a main experiment and a generalization validation. The main experiment, based on the Los Angeles 1∶2000 building dataset, encodes polygons using three grid sizes—0.2, 0.3, and 0.5 mm—to achieve simplifications at target scales of 1∶5000 and 1∶10 000. Results indicate that the model performs optimally with a grid size of 0.3 mm, achieving a consistency rate exceeding 92.0% with manual annotations on the validation set. A generalization experiment on building polygon data from parts of Beijing further verified the model's transferability. Comparative analysis with an LSTM model, under similar parameter scales, showed that the LSTM model failed to converge effectively and could not produce usable results. This study confirms the potential of Transformer in handling spatial geometric sequence tasks and demonstrates its ability to effectively reuse existing simplification samples. The proposed approach offers a new, engineering-practical pathway for intelligent building polygon simplification.

Key words: map generalization, buildings polygon simplification, Tokenization, Transformer model, context engineering

中图分类号:

P208

刘鹏程, 成晓强, 肖天元, 杨敏, 艾廷华. 一种面向地图综合建筑多边形化简的Transformer模型[J]. 测绘学报, 2026, 55(1): 124-137.

Pengcheng LIU, Xiaoqiang CHENG, Tianyuan XIAO, Min YANG, Tinghua AI. A Transformer model for building polygon simplification in map generalization[J]. Acta Geodaetica et Cartographica Sinica, 2026, 55(1): 124-137.

图/表 16

图1

图2

图3

图4

图5

表1

表2

图6

图7

表3

图8

图9

图10

图11

图12

表4

参考文献 27

[1]	王家耀, 孙群, 王光霞, 等. 地图学原理与方法[M]. 2版. 北京: 科学出版社, 2014.
	WANG Jiayao, SUN Qun, WANG Guangxia, et al. Principle and method of cartography[M]. 2nd ed. Beijing: Science Press, 2014.
[2]	WANG Zeshen, LEE D. Building simplification based on pattern recognition and shape analysis[C]//Proceedings of the 9th International Symposium on Spatial Data Handling. Beijing: [s.n.], 2000: 58-72.
[3]	尹烁, 闫小明, 晏雄锋. 基于特征边重构的建筑物化简方法[J]. 测绘学报, 2020, 49(6): 703-710. DOI: . doi: 10.11947/j.AGCS.2020.20190299
	YIN Shuo, YAN Xiaoming, YAN Xiongfeng. Simplification method of building polygon based on feature edges reconstruction[J]. Acta Geodaetica et Cartographica Sinica, 2020, 49(6): 703-710. DOI: . doi: 10.11947/j.AGCS.2020.20190299
[4]	SESTER M. Generalization based on least squares adjustment[C]//Proceedings of 2000 International Archives of Photogrammetry and Remote Sensing. Amsterdam: International Society for Photogrammetry & Remote Sensing, 2000: 931G938.
[5]	刘鹏程, 艾廷华, 邓吉芳. 基于最小二乘的建筑物多边形的化简与直角化[J]. 中国矿业大学学报, 2008, 37(5): 699-704.
	LIU Pengcheng, AI Tinghua, DENG Jifang. Simplification and rectangularity of building polygon based on least squares adjustment[J]. Journal of China University of Mining and Technology, 2008, 37(5): 699-704.
[6]	许文帅, 龙毅, 周侗, 等. 基于邻近四点法的建筑物多边形化简[J]. 测绘学报, 2013, 42(6): 929-936.
	XU Wenshuai, LONG Yi, ZHOU Tong, et al. Simplification of building polygon based on adjacent four point method[J]. Acta Geodaetica et Cartographica Sinica, 2013, 42(6): 929-936.
[7]	李俐俐, 李成名, 卢小平, 等. 顾及邻近五点的建筑物多边形化简方法[J]. 测绘通报, 2019(3): 41-45.
	LI Lili, LI Chengming, LU Xiaoping, et al. Simplification of building polygon based on adjacent five-point method[J]. Bulletin of surveying and mapping, 2019(3): 41-45.
[8]	RAINSFORD D, MACKANESS W. Template matching in support of generalization of rural buildings[M]//RICHARDSON D E, VAN OOSTEROM P. Advances in spatial data handling. Berlin: Springer, 2002: 137-151.
[9]	刘鹏程, 艾廷华, 胡晋山, 等. 基于原型模板形状匹配的建筑多边形化简[J]. 武汉大学学报(信息科学版), 2010, 35(11): 1369-1371.
	LIU Pengcheng, AI Tinghua, HU Jinshan, et al. Building-polygon simplification based on shape matching of prototype template[J]. Geomatics and Information Science of Wuhan University, 2010, 35(11): 1369-1372.
[10]	AI Tinghua, CHENG Xiaoqiang, LIU Pengcheng, et al. A shape analysis and template matching of building features by the Fourier transform method[J]. Computers, Environment and Urban Systems, 2013, 41: 219-233.
[11]	YAN X, AI T, ZHANG X. Template matching and simplification method for building features based on shape cognition[J]. International Journal of Geo-Information, 2017, 6(8): 250.
[12]	晏雄锋, 艾廷华, 杨敏. 居民地要素化简的形状识别与模板匹配方法[J]. 测绘学报, 2016, 45(7): 874-882. DOI: . doi: 10.11947/j.AGCS.2016.20150162
	YAN Xiongfeng, AI Tinghua, YANG Min. A simplification of residential feature by the shape cognition and template matching method[J]. Acta Geodaetica et Cartographica Sinica, 2016, 45(7): 874-882. DOI: . doi: 10.11947/j.AGCS.2016.20150162
[13]	郭仁忠, 艾廷华. 制图综合中建筑物多边形的合并与化简[J]. 武汉测绘科技大学学报, 2000, 25(1), 25-30.
	GUO Renzhong, AI Tinghua. Simplification and aggregation of building polygon automatic map generalization[J]. Journal of Wuhan Technical University of surveying and Mapping, 2000, 25(1): 25-30.
[14]	王辉连, 武芳, 张琳琳, 等. 数学形态学和模式识别在建筑多边形化简中的应用[J]. 测绘学报, 2005, 34(3): 269-276.
	WANG Huilian, WU Fang, ZHANG Linlin, et al. The application of mathematical morphology and pattern recognition to building polygon simplification[J]. Acta Geodaetica et Cartographica Sinica, 2005, 34(3): 269-276.
[15]	武芳, 杜佳威, 钱海忠, 等. 地图综合智能化研究的发展与思考[J]. 武汉大学学报(信息科学版), 2022, 47(10): 1675-1687.
	WU Fang, DU Jiawei, QIAN Haizhong, et al. Overview of research progress and reflections in intelligent map generalization[J]. Geomatics and Information Science of Wuhan University, 2022, 47(10), 1675-1687.
[16]	YAN X, YANG M, AI T. Deep learning in automatic map generalization: achievements and challenges[J]. Geo-Spatial Information Science, 2025: 1-22.
[17]	FENG Y, THIEMANN F, SESTER M. Learning cartographic building generalization with deep convolutional neural networks[J]. International Journal of Geo-Information, 2019, 8(6): 258.
[18]	YAN X, YANG M. A deep learning approach for polyline and building simplification based on graph autoencoder with flexible constraints[J]. Cartography and Geographic Information Science, 2024, 51(1), 79-96.
[19]	YANG M, YUAN T, YAN X, et al. A hybrid approach to building simplification with an evaluator from a backpropagation neural network[J]. International Journal of Geographical Information Science, 2022, 36(2): 280-309.
[20]	晏雄锋, 袁拓, 杨敏, 等. 建筑物形状特征分析表达与自适应化简方法[J]. 测绘学报, 2022, 51(2): 269-278. DOI: . doi: 10.11947/j.AGCS.2020.20190299
	YAN Xiongfeng, YUAN Tuo, YANG Min, et al. An adaptive building simplification approach based on shape analysis and representation[J]. Acta Geodaetica et Cartographica Sinica, 2022, 51(2): 269-278. DOI: . doi: 10.11947/j.AGCS.2020.20190299
[21]	HU Y, GOODCHILD M, YUAN M, et al. A five-year milestone: reflections on advances and limitations in GeoAI research[J]. Annals of GIS, 2024, 30(1): 1-14. DOI: . doi: 10.1080/19475683.2024.2309866
[22]	吴华意, 沈张骁, 侯树洋, 等. 大语言模型驱动的GIS分析：方法、应用与展望[J]. 测绘学报, 2025, 54(4): 621-635. DOI: . doi: 10.11947/j.AGCS.2025.20240468
	WU Huayi, SHEN Zhangxiao, HOU Shuyang, et al. Large language model-driven GIS analysis: methods, applications, and prospects[J]. Acta Geodaetica et Cartographica Sinica, 2025, 54(4): 621-635. DOI: . doi: 10.11947/j.AGCS.2025.20240468
[23]	VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach: Curran Associates Inc., 2017: 5998-6008.
[24]	ZHU G, TANG M, MA Y, et al. LandGPT: a multimodal large language model for parcel-level land use classification with multi-source data[J]. International Journal of Geographical Information Science, 2025: 1-24.
[25]	LI H, XU C, YANG W, et al. Unsupervised multiview UAV image geo-localization via iterative rendering[J]. IEEE Transactions on Geoscience and Remote Sensing, 2025, 63: 1-15.
[26]	DU J, WU F, YIN J, et al. Polyline simplification based on the artificial neural network with constraints of generalization knowledge[J]. Cartography and Geographic Information Science, 49(4): 313-337.
[27]	HUANG F, LÜ J R, LI G L, et al. SPOK: tokenizing geographic space for enhanced spatial reasoning in GeoAI[J]. International Journal of Geographical Information Science, 2025: 1-41. https://doi.org/10.1080/13658816.2025.2497810.

参数	网格尺寸
参数	0.2 mm	0.3 mm	0.5 mm
点位平均偏移/m	0.26	0.39	0.49
词汇表尺寸	250 002	110 891	40 002
模型参数量	2.9×10⁸	1.5×10⁸	8×10⁷

参数	目标比例尺	网格尺寸			目标比例尺	网格尺寸
参数	目标比例尺	0.2 mm	0.3 mm	0.5 mm	目标比例尺	0.2 mm	0.3 mm	0.5 mm
训练时间/min		2650	2130	1980		2730	2440	1802
训练损失值		0.43	0.035	0.039		0.54	0.13	0.13
验证损失值	1∶5000	0.45	0.038	0.043	1∶10 000	0.54	0.14	0.14
一致率/（%）	1∶5000	32.0	92.3	90.6	1∶10 000	34.3	94.3	92.1
面积增量/（%）		－0.14	0.14	0.23		－0.20	0.21	0.32
直角率/（%）		93.3	96.7	97.0		89.8	97.2	97.8

序号	一致性评价	序号	一致性评价
1	Y Y	11	N Y
2	Y Y	12	N Y
3	Y Y	13	N Y
4	Y Y	14	N Y
5	Y Y	15	N Y
6	Y Y	16	N N
7	Y Y	17	N N
8	Y Y	18	N N
9	Y Y	19	N N
10	Y Y	20	N N

序号	面积差百分比/（%）	序号	面积差百分比/（%）
1	－1.2	6	1.0
2	2.6	7	－3.0
3	－1.4	8	3.4
4	0.5	9	0.8
5	2.8	10	1.7

一种面向地图综合建筑多边形化简的Transformer模型

A Transformer model for building polygon simplification in map generalization

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 16

参考文献 27

相关文章 15

编辑推荐

Metrics

本文评价

序号	一致性评价	序号	一致性评价
1	Y Y	11	N Y
2	Y Y	12	N Y
3	Y Y	13	N Y
4	Y Y	14	N Y
5	Y Y	15	N Y
6	Y Y	16	N N
7	Y Y	17	N N
8	Y Y	18	N N
9	Y Y	19	N N
10	Y Y	20	N N

[1]	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.
[2]	刘鹏程, 马宏然, 周洋, 邵子芹. 一种曲线数据压缩的自编码器神经网络方法[J]. 测绘学报, 2024, 53(8): 1634-1643.
[3]	陈占龙, 鲁谢春, 徐永洋. 基于图顶点深度聚类的建筑物合并方法[J]. 测绘学报, 2024, 53(4): 736-749.
[4]	肖天元, 艾廷华, 余华飞, 杨敏, 刘鹏程. 地图综合图卷积神经网络点群简化方法[J]. 测绘学报, 2024, 53(1): 158-172.
[5]	安晓亚, 朱余德, 晏雄锋. 卷积神经网络支持下的建筑物选取方法[J]. 测绘学报, 2023, 52(9): 1574-1583.
[6]	张鑫港, 闫浩文. 河流要素比例尺变化与中位数Hausdorff距离的关系[J]. 测绘学报, 2023, 52(8): 1364-1374.
[7]	吴晨昊, 向隆刚, 张叶廷, 吴华意. 基于地理空间感知型表征学习的轨迹相似度计算[J]. 测绘学报, 2023, 52(4): 670-678.
[8]	江宝得, 许少芬, 巫勇, 王淼. 基于区域候选网络的矢量线要素自动化简方法[J]. 测绘学报, 2023, 52(12): 2209-2222.
[9]	张新长, 何显锦, 孙颖, 黄健锋, 张志强. 多尺度空间数据联动更新技术研究现状及展望[J]. 测绘学报, 2022, 51(7): 1520-1535.
[10]	武芳, 杜佳威, 吴芳华. 海底地貌数据综合研究进展[J]. 测绘学报, 2022, 51(7): 1588-1605.
[11]	艾廷华, 张翔. 地理信息科学中尺度概念的诠释与表达[J]. 测绘学报, 2022, 51(7): 1640-1652.
[12]	杜佳威, 武芳, 朱丽, 刘呈熠, 王安东. 图形、图像融合利用的集成学习智能化简方法及其在岛屿岸线化简中的应用[J]. 测绘学报, 2022, 51(3): 373-387.
[13]	高晓蓉, 闫浩文, 禄小敏. 多尺度地图空间居民地语义相似度计算方法[J]. 测绘学报, 2022, 51(1): 95-103.
[14]	刘远刚, 李少华, 蔡永香, 何贞铭, 马潇雅, 李鹏程, 郭庆胜, 何宗宜. 移位安全区约束下的建筑物群移位免疫遗传算法[J]. 测绘学报, 2021, 50(6): 812-822.
[15]	刘呈熠, 武芳, 巩现勇, 行瑞星, 杜佳威. 空间知识挖掘的自然面群聚集度聚类方法[J]. 测绘学报, 2021, 50(4): 544-555.

序号	一致性评价	序号	一致性评价
1	Y Y	11	N Y
2	Y Y	12	N Y
3	Y Y	13	N Y
4	Y Y	14	N Y
5	Y Y	15	N Y
6	Y Y	16	N N
7	Y Y	17	N N
8	Y Y	18	N N
9	Y Y	19	N N
10	Y Y	20	N N

序号	一致性评价	序号	一致性评价
1	Y Y	11	N Y
2	Y Y	12	N Y
3	Y Y	13	N Y
4	Y Y	14	N Y
5	Y Y	15	N Y
6	Y Y	16	N N
7	Y Y	17	N N
8	Y Y	18	N N
9	Y Y	19	N N
10	Y Y	20	N N