空间关系增强与异构特征融合相结合的道路信息提取方法

doi:10.11947/j.AGCS.2025.20250250

摘要/Abstract

摘要：

针对高原特殊环境下遥感影像中道路材质与背景混淆、结构细长易断裂等问题，本文提出了一种结合空间关系增强器（spatial relationship enhancer，SRE）和连通性约束损失（connectivity loss，Cnt_Loss）的改进道路提取模型SRENet。核心贡献包括：①设计空间关系增强器，通过关键点图卷积显式建模道路拓扑结构，显著提升弯曲与遮挡区域的连通性检测能力；②构建双分支架构并设计异构特征融合模块，实现语义特征与空间细节的互补增强，增强对材质与环境类似的低对比度道路的提取能力；③提出连通性约束损失函数，通过几何驱动优化抑制狭窄断裂区域的误分割。本文方法以双分支深度神经网络为基础，通过异构特征融合模块实现多尺度特征互补，并结合连通性约束损失函数Cnt_Loss对道路几何特征进行优化。研究表明：SRENet在JL1与DGRD数据集上的IoU分别达到0.700 2和0.660 4，较现有模型分别提升了0.011 6和0.025 2；在道路连接性优化方面表现突出，显著减少了在弯曲路段与行道树遮挡区域的断裂数量；提出的Cnt_Loss函数通过几何约束机制，有效解决了弱边界道路的漏检问题。

关键词: 青藏高原, 道路提取, 异构特征融合, 空间关系增强, 图卷积

Abstract:

This study addresses the issues of material confusion and the susceptibility of thin and elongated structures to fragmentation in road extraction from remote sensing images under complex plateau environments. We propose an improved road extraction model, SRENet, which incorporates a spatial relationship enhancer (SRE) and a connectivity loss (Cnt_Loss). The core contributions of this work are as follows: ① The spatial relationship enhancer is designed to explicitly model the topological structure of roads through key point graph convolution, significantly improving the connectivity detection capability in curved and occluded areas. ② A dual-branch heterogeneous architecture was constructed with a specially designed heterogeneous feature fusion module to achieve complementary enhancement between semantic features and spatial details, thereby improving extraction capability for low-contrast roads with material and environmental similarities. ③ A connectivity constraint loss function is proposed to suppress mis-segmentation in narrow and fragmented regions through geometry-driven optimization. Based on a dual-branch deep neural network, this method achieves multi-scale feature complementarity through the heterogeneous feature fusion module and optimizes road geometric features using the Cnt_Loss. The research results demonstrate that SRENet achieves IoU scores of 0.700 2 and 0.660 4 on the JL1 and DGRD datasets, respectively, representing improvements of 0.011 6 and 0.025 2 over existing models. The model also demonstrates outstanding performance in optimizing road connectivity, such as significantly reducing the number of fractures in curved sections and areas occluded by roadside trees. The proposed Cnt_Loss function effectively addresses the problem of missing detections in roads with weak boundaries through geometric constraint mechanisms. This method provides a new solution for road extraction from high-resolution remote sensing images.

Key words: Qinghai-Xizang Plateau, road extraction, heterogeneous feature fusion, spatial relationship enhancement, graph convolution

中图分类号:

P237

曹云刚, 杨鹏, 龚江波, 朱高, 沈星宇. 空间关系增强与异构特征融合相结合的道路信息提取方法[J]. 测绘学报, 2025, 54(12): 2219-2232.

Yungang CAO, Peng YANG, Jiangbo GONG, Gao ZHU, Xingyu SHEN. A road extraction method integrating spatial-relation enhancement and heterogeneous feature fusion[J]. Acta Geodaetica et Cartographica Sinica, 2025, 54(12): 2219-2232.

图/表 16

图1

图2

图3

图4

表1

图5

图6

图7

表2

图8

图9

图10

图11

表3

图12

表4

参考文献 35

[1]	MA Yaoming, ZHONG Lei, MA Weiqiang, et al. Determination of land surface heat fluxes at different temporal scales over the Tibetan Plateau[J]. Journal of Geodesy and Geoinformation Science, 2021(1): 144-152.
[2]	李程. 遥感技术在农村公路核查中的应用研究[J]. 测绘与空间地理信息, 2018, 41(5): 48-49, 52.
	LI Cheng. Application of remote sensing technology in rural road verification[J]. Geomatics & Spatial Information Technology, 2018, 41(5): 48-49, 52.
[3]	戴激光, 杜阳, 方鑫鑫, 等. 多特征约束的高分辨率光学遥感影像道路提取[J]. 遥感学报, 2018, 22(5): 777-791.
	DAI Jiguang, DU Yang, FANG Xinxin, et al. Road extraction method for high resolution optical remote sensing images with multiple feature constraints[J]. Journal of Remote Sensing, 2018, 22(5): 777-791.
[4]	曹云刚, 王志盼, 慎利, 等. 像元与对象特征融合的高分辨率遥感影像道路中心线提取[J]. 测绘学报, 2016, 45(10): 1231-1240, 1249. DOI: . doi: 10.11947/j.AGCS.2016.20160158
	CAO Yungang, WANG Zhipan, SHEN Li, et al. Fusion of pixel-based and object-based features for road centerline extraction from high-resolution satellite imagery[J]. Acta Geodaetica et Cartographica Sinica, 2016, 45(10): 1231-1240, 1249. DOI: . doi: 10.11947/j.AGCS.2016.20160158
[5]	张刚. 基于深度学习的遥感图像语义分割关键技术研究[D]. 成都: 中国科学院大学(中国科学院光电技术研究所), 2020.
	ZHANG Gang. Research on key technologies of remote sensing image semantic segmentation based on deep learning[D]. Chengdu: Institute of Optics and Electronics, Chinese Academy of Sciences, 2020.
[6]	贺浩, 王仕成, 杨东方, 等. 基于Encoder-Decoder网络的遥感影像道路提取方法[J]. 测绘学报, 2019, 48(3): 330-338. DOI: . doi: 10.11947/j.AGCS.2019.20180005
	HE Hao, WANG Shicheng, YANG Dongfang, et al. An road extraction method for remote sensing image based on Encoder-Decoder network[J]. Acta Geodaetica et Cartographica Sinica, 2019, 48(3): 330-338. DOI: . doi: 10.11947/j.AGCS.2019.20180005
[7]	王艳军, 唐徐超, 王成, 等. 基于道路拓扑关联特征的城乡道路面精细提取网络[J]. 测绘学报, 2025, 54(1): 75-89. DOI: . doi: 10.11947/j.AGCS.2025.20240124
	WANG Yanjun, TANG Xuchao, WANG Cheng, et al. Urban and rural road surface extraction network based on road topological correlation features[J]. Acta Geodaetica et Cartographica Sinica, 2025, 54(1): 75-89. DOI: . doi: 10.11947/j.AGCS.2025.20240124
[8]	MNIH V, HINTON G E. Learning to detect roads in high-resolution aerial images[C]//Proceedings of Computer Vision-ECCV 2010. Berlin: Springer, 2010: 210-223.
[9]	LI Peikang, ZANG Yu, WANG Cheng, et al. Road network extraction via deep learning and line integral convolution[C]//Proceedings of 2016 IEEE International Geoscience and Remote Sensing Symposium. Beijing: IEEE, 2016: 1599-1602.
[10]	CHENG Guangliang, ZHU Feiyun, XIANG Shiming, et al. Road centerline extraction via semisupervised segmentation and multidirection nonmaximum suppression[J]. IEEE Geoscience and Remote Sensing Letters, 2016, 13(4): 545-549.
[11]	WEI Yanan, WANG Zulin, XU Mai. Road structure refined CNN for road extraction in aerial image[J]. IEEE Geoscience and Remote Sensing Letters, 2017, 14(5): 709-713.
[12]	ALSHEHHI R, MARPU P R. Hierarchical graph-based segmentation for extracting road networks from high-resolution satellite images[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2017, 126: 245-260.
[13]	ZHOU Lichen, ZHANG Chuang, WU Ming. D-LinkNet: LinkNet with pretrained encoder and dilated convolution for high resolution satellite imagery road extraction[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Salt Lake City: IEEE, 2018: 182-186.
[14]	LI Jing, LIU Yong, ZHANG Yindan, et al. Cascaded attention DenseUNet (CADUNet) for road extraction from very-high-resolution images[J]. ISPRS International Journal of Geo-Information, 2021, 10(5): 329.
[15]	WU Qiangqiang, LUO Feng, WU Penghai, et al. Automatic road extraction from high-resolution remote sensing images using a method based on densely connected spatial feature-enhanced pyramid[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020, 14: 3-17.
[16]	JIE Yongshi, HE Hongyan, XING Kun, et al. MECA-net: a multiscale feature encoding and long-range context-aware network for road extraction from remote sensing images[J]. Remote Sensing, 2022, 14(21): 5342.
[17]	ZHANG Zixuan, SUN Xuan, LIU Yuxi. GMR-net: road-extraction network based on fusion of local and global information[J]. Remote Sensing, 2022, 14(21): 5476.
[18]	ZHANG Yang, XIONG Zhangyue, ZANG Yu, et al. Topology-aware road network extraction via multi-supervised generative adversarial networks[J]. Remote Sensing, 2019, 11(9): 1017.
[19]	LI Xingang, WANG Yuebin, ZHANG Liqiang, et al. Topology-enhanced urban road extraction via a geographic feature-enhanced network[J]. IEEE Transactions on Geoscience and Remote Sensing, 2020, 58(12): 8819-8830.
[20]	HUANG Zilong, WANG Xinggang, HUANG Lichao, et al. CCNet: criss-cross attention for semantic segmentation[C]//Proceedings of 2019 IEEE/CVF International Conference on Computer Vision. Seoul: IEEE, 2019: 603-612.
[21]	CUI Fumin, SHI Yichang, FENG Ruyi, et al. A graph-based dual convolutional network for automatic road extraction from high resolution remote sensing images[C]//Proceedings of IGARSS 2022—2022 IEEE International Geoscience and Remote Sensing Symposium. Kuala Lumpur: IEEE, 2022: 3015-3018.
[22]	LU Xiaoyan, ZHONG Yanfei, ZHENG Zhuo, et al. Multi-scale and multi-task deep learning framework for automatic road extraction[J]. IEEE Transactions on Geoscience and Remote Sensing, 2019, 57(11): 9362-9377.
[23]	HE Wei, XUE Ruoyao, LU Fangxiao, et al. KDGraph: a keypoint detection method for road graph extraction from remote sensing images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 63: 5604814.
[24]	HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]//Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 770-778.
[25]	WANG Jingdong, SUN Ke, CHENG Tianheng, et al. Deep high-resolution representation learning for visual recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 43(10): 3349-3364.
[26]	VELICKOVIC P, CUCURULL G, CASANOVA A, et al. Graph attention networks[C]//Proceedings of 2018 International Conference on Learning Representations. Vancouver: [s. n.], 2018.
[27]	CHEN C R, FAN Quanfu, PANDA R. CrossViT: cross-attention multi-scale vision Transformer for image classification[C]//Proceedings of 2021 IEEE/CVF International Conference on Computer Vision. Montreal: IEEE, 2021: 347-356.
[28]	ZHAO Hengshuang, SHI Jianping, QI Xiaojuan, et al. Pyramid scene parsing network[C]//Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2017: 6230-6239.
[29]	DEMIR I, KOPERSKI K, LINDENBAUM D, et al. DeepGlobe 2018: a challenge to parse the Earth through satellite images[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Salt Lake City: IEEE, 2018: 172-17209.
[30]	CHEN L C, ZHU Yukun, PAPANDREOU G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation[C]//Proceedings of 2018 European Conference on Computer Vision. Cham: Springer, 2018: 833-851.
[31]	CHENG Bowen, MISRA I, SCHWING A G, et al. Masked-attention mask Transformer for universal image segmentation[C]//Proceedings of 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Orleans: IEEE, 2022: 1280-1289.
[32]	LIU Ze, LIN Yutong, CAO Yue, et al. Swin Transformer: hierarchical vision Transformer using shifted windows[C]//Proceedings of 2021 IEEE/CVF International Conference on Computer Vision. Montreal: IEEE, 2021: 9992-10002.
[33]	YANG Zhigang, ZHOU Daoxiang, YANG Ying, et al. TransRoadNet: a novel road extraction method for remote sensing images via combining high-level semantic feature and context[J]. IEEE Geoscience and Remote Sensing Letters, 2022, 19: 6509505.
[34]	YU Zhengbo, CHEN Zhe, XIAO Keyan, et al. SegRoadv2: a hybrid deformable self-attention and convolutional network for road extraction with connectivity structure[J]. International Journal of Digital Earth, 2025, 18(1): 2480267.
[35]	ZHOU Bolei, KHOSLA A, LAPEDRIZA A, et al. Learning deep features for discriminative localization[C]//Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 2921-2929.

数据集	数量/组	尺寸/像素	分辨率/m	训练集/组	验证集/组	描述
扩充数据集	600	512×512	0.75	540	40	使用MakeSense基于像素级进行标注
吉林一号道路提取数据集	3800	512×512	0.75	3400	400	取原数据集训练集与扩充数据集构建新数据集（以下称“JL1-P”），共3800组
DGRD	6800	1024×1024	0.5	6100	700	取原数据集训练集与扩充数据集构建新数据集（以下称“DGRD-P”），共6800组

数据集	网络	IoU	Precision	Recall	F₁值
JL1-P	DeepLabV3+	0.558 9	0.745 2	0.682 1	0.692 0
	Mask2former	0.612 1	0.760 9	0.754 4	0.740 6
	SwinTransformer	0.623 4	0.796 2	0.732 6	0.748 6
	DLinkNet	0.660 1	0.839 9	0.678 7	0.741 6
	TransRoadNet	0.684 0	0.845 0	0.785 8	0.802 7
	SegRoadv2	0.688 6	0.850 1	0.787 0	0.803 6
	SRENet（本文方法）	0.700 2	0.840 1	0.791 3	0.805 4
DGRD-P	DeepLabV3+	0.479 5	0.758 5	0.566 9	0.639 4
	Mask2former	0.602 1	0.742 9	0.7629	0.742 6
	SwinTransformer	0.611 0	0.760 7	0.758 8	0.748 9
	DLinkNet	0.620 2	0.766 3	0.766 6	0.756 8
	TransRoadNet	0.632 7	0.771 6	0.782 0	0.769 8
	SegRoadv2	0.635 2	0.760 3	0.797 2	0.767 7
	SRENet（本文方法）	0.660 4	0.775 5	0.825 7	0.794 0

Net	HRNet	CA-HFM	SRE	Cnt_Loss	IoU	Precision	Recall	F₁值
Baseline0					0.441 1	0.754 8	0.481 0	0.558 3
Baseline1	√				0.481 7	0.826 1	0.529 9	0.627 3
Baseline2	√	√			0.539 3	0.759 4	0.649 8	0.693 3
Baseline3	√	√	√		0.623 4	0.796 2	0.732 6	0.748 6
Baseline4	√	√	√	√	0.700 2	0.840 1	0.791 3	0.805 4

网络	数据集	IoU	Precision	Recall	F₁值
SRENet	JL1	0.673 0	0.805 4	0.751 8	0.766 1
	JL1-P	0.700 2	0.840 1	0.791 3	0.805 4
	DGRD	0.623	0.749 4	0.775 0	0.751 3
	DGRD-P	0.660 4	0.775 5	0.825 7	0.794 0
DLinkNet	JL1	0.621 3	0.782 9	0.709 6	0.729 4
	JL1-P	0.660 1	0.83 99	0.678 7	0.741 6
	DGRD	0.613 5	0.745 4	0.769 4	0.743 8
	DGRD-P	0.620 2	0.766 3	0.766 6	0.756 8
TransRoadNet	JL1	0.644 2	0.799 1	0.730 0	0.748 0
	JL1-P	0.684 0	0.845 0	0.785 8	0.802 7
	DGRD	0.621 0	0.753 8	0.771 4	0.751 0
	DGRD-P	0.632 7	0.771 6	0.782 0	0.769 8
SegRoadv2	JL1	0.662 0	0.805 3	0.748 9	0.764 2
	JL1-P	0.688 6	0.850 1	0.7870	0.803 6
	DGRD	0.621 1	0.750 2	0.777 0	0.751 4
	DGRD-P	0.635 2	0.760 3	0.797 2	0.767 7