边缘增强的EDU-Net遥感影像建筑物提取

Edge-enhanced EDU-Net for Remote Sensing Image Building Extraction

针对高空间分辨率遥感影像背景信息复杂,现有语义分割模型提取建筑物轮廓易出现边缘缺失、边界划分不清晰等问题,提出一种边缘增强型EDU-Net深度学习网络。在EDU-Net结构设计中,通过构建边缘特征约束模块,结合Sobel边缘检测图细化建筑物边缘特征;同时,基于二次强化策略提升模型对建筑物边缘信息的表征学习能力。在WHU数据集上,EDU-Net语义分割指标MIoU和F1分别为91.99%和92.37%,相较DoubleU-Net提升0.99%和1.05%;在中国典型城市建筑物数据集上,MIoU达83.12%,同时边缘与边界分割效果更佳,证明了所提出模型具有较好的分割性能和普适性。

In view of the complex background information of high spatial resolution remote sensing images,the existing semantic segmentation models are prone to edge missing and unclear boundary delineation in extracting building outlines.In this paper,an edge-enhanced EDU-Net deep learning network is proposed.In the design of the EDU-Net structure,the building edge features are refined by constructing an edge feature constraint module combined with a Sobel edge detection map;at the same time,the model’s ability to learn the representation of building edge information is enhanced based on a secondary reinforcement strategy.In the WHU dataset,the EDU-Net semantic segmentation metrics MIoU and F1 are 91.99%and 92.37%respectively,which are 0.99%and 1.05%better than that of DoubleU-Net.In the typical Chinese urban building dataset,MIoU reaches 83.12%and the edge and boundary segmentation is better,which further confirms the segmentation performance and universality of the proposed model.