Article Details

Abstract

To address the problems of unclear building boundaries and easy loss of tiny buildings in existing remote sensing image building extraction models, an improved multi-scale gated Self-Attention network method is proposed. A multi-scale feature aggregation strategy is employed to enhance feature extraction effectiveness. A gated Self-Attention mechanism is combined to optimize feature correlation, integrate global and local features, and reduce the loss of tiny building detail information. An edge loss enhancement mechanism is introduced to refine building boundary extraction, compensate for information loss, and improve segmentation accuracy. Comparative experiments and model applicability validation are also conducted. Experimental results on the Inria dataset demonstrate that the proposed method achieves significantly higher Dice coefficient and mIoU values than other methods, and highlights its good robustness and broad generalization capability.

Keywords

Self-Attention Mechanism; multi-scale feature aggregation; global and local feature fusion; edge loss; remote sensing image segmentation; building extraction

References

1. [1] He Zhimeng, Ding Haiyong, An Bingqi Hollow Convolutional E-Unet Algorithm for Building Extraction from High Resolution Remote Sensing Images [J]. Journal of Surveying and Mapping, 2022, 51 (3): 457-467
2. [2] Liu Erhu, Li Haowen, Hu Yu, etc Dual branch remote sensing image building extraction network integrating local global features [J]. Computer Science and Exploration, 2025, 19 (9): 2430-2444
3. [3] AVUDAIAMMAL R，ELAVENI P，SELVAN S，et al. Extraction of Buildings in Urban Area for Surface Area Assessment from Satellite Imagery based on Morphological Building Index using SVM Classifier [J].Journal of the Indian Society of Remote Sensing，2020，48：1325-1344.
4. [4] YANG K，YI J Z，CHEN A B，et al.ConDinet++: Full-Scale Fusion Network Based on Conditional Dilated Convolution to Extract Roads From Remote Sensing Images [J]. IEEE Geoscience and Remote Sensing Letters，2021，19：1-5.
5. [5] HE D，SHI Q，LIU X P，et al. Generating 2 m Fine-scale Urban Tree Cover Product Over 34 Metropolises in China Based on Deep Context-aware Sub-pixel Mapping Network [J/OL].International Journal of Applied Earth Observation and Geoinformation，2022，106：102667[2025-11-06].https://doi.org/10.1016/j.jag.2021.102667.
6. [6] ZHANG Z X，LIU Q J，WANG Y H.Road Extraction by Deep Residual U-Net [J].IEEE Geoscience and Remote Sensing Letters，2018，15（5）：749-753.
7. [7] DING W W，HUANG H M，WANG Y. Semantic Segmentation of Remote Sensing Architectural Images Based on GAN and UNet3+ Model [C]// PRICAI 2023: Trends in Artificial Intelligence.Jakarta：Springer，2023：262-273.
8. [8] CHEN L C，ZHU Y K，PAPANDREOU G，et al. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation [C]//Computer Vision - ECCV 2018.Munich：Springer，2018：833-851.
9. [9] Ming Xingtao, Yang Dehong Multi module based remote sensing image building extraction method [J]. Advances in Laser and Optoelectronics, 2024, 61 (4): 385-393
10. [10] Lei Lankun, Zhong Hao, Yang Rui, etc Building extraction method based on Res UNet high-resolution remote sensing images [J]. Computer Science and Applications, 2024, 14 (9): 103-110
11. [11] CAO H，WANG Y Y，CHEN J，et al. Swin-unet: Unet-like Pure Transformer for Medical Image Segmentation [C]//Computer Vision – ECCV 2022 Workshops.Tel Aviv：Springer，2022：205-218.

Funding

Key Research Project of Higher Education Institutions in Henan Province (24B510016); Zhengzhou Sias College 2025 Research Funding Project (2025XKB011)

About the authors

Xu Xinhua (1986-), female, Han ethnicity, from Hebi, Henan, associate professor, master's student, research direction: image processing and applications.