Improved remote sensing image building extraction network based on U-net network

Deep learning method is used to extract buildings from remote sensing images. Due to the diversity and complex spectral information of buildings, there are some problems in semantic segmentation of buildings, such as poor recognition effect of misclassification and omission, and holes in buildings. In view of the above problems, this paper proposes an improved unet model, which takes the traditional U-net network model as the basic framework, and improves the network model by introducing improved residual module, Coordconv unit and multi-scale enhancement module. In order to test the reliability of the network model, experiments were performed on two building datasets, Building New Zealand and Building New East Asia, with image resolutions of 0.075 meters and 2.77 meters, respectively. The accuracy evaluation index intersection over union (IOU) reached 90.60 % and 72.18 %, respectively, which were 0.27 % and 1.70 % higher than the original Unet network. The results show that the improved Unet network shows high extraction accuracy and strong generalization ability in the face of building extraction tasks of different scales.