Abstract
High-precision building information extraction is essential in surveying and mapping. Deep learning methods have proven effective in extracting buildings from satellite optical images. This paper introduces a Boundary-Enhanced Semantic Segmentation Model (BE-SegModel) to address the issue of unclear building boundary segmentation in high-resolution satellite images. The model simultaneously extracts building masks and boundary information, ultimately enhancing extraction accuracy. Deep-level heatmaps and feature maps are visualized to analyze the knowledge acquired by the proposed model. The model's performance is evaluated using the xBD, WHU Global, and WHU East Asia Building datasets. Experimental results indicate that, with the inclusion of semantic boundaries, the proposed BE-SegModel outperforms seven mainstream models, including ResNet34, HRNet, Deeplabv3+, CBRNet, FCN, FCN-UNet, and ResNeXt50. In summary, BE-SegModel significantly improves segmentation accuracy, achieving over 6.9% improvement relative to the baseline model.
Funding Information
  • Basic Frontier Project of Xinjiang Institute of Ecology and Geography
  • Chinese Academy of Sciences (E3500201)
  • State Key Project of National Natural Science Foundation of China-University of Chinese Academy of Sciences
  • Beijing, China
  • Key Projects of Joint Fund for Regional Innovation and Development (U21A20108, U22A20566)

This publication has 55 references indexed in Scilit: