EA U2-Net: An Efficient Building Extraction Algorithm Based on Complex Background Information

Effective extraction of building edge information based on high-resolution remote sensing images is the basis for efficient urban 3D modeling. Existing building extraction methods still have some problems, such as an uncertain segmentation scale, effective feature selection, and sample selection. In this paper, we propose a practical building extraction method based on convolutional network edge-enhanced attention U2-Net (EA U2-Net) to accurately achieve multi-scale extraction of buildings from remote sensing imagery. First, the U2-Net is used as the backbone network for building extraction because each stage of the network is filled by residual U-block (RSU), and the network can better aggregate multi-scale features. Second, the building edge feature map is introduced into the generation network to compensate for the problems of insufficient extracted building edge features and loss of detail. Finally, the convolutional block attention module is used to achieve effective feature extraction of buildings. We performed the experiment on the WHU building dataset, and the experimental results showed that the EA U2-Net model has significantly improved the ability to extract buildings, with an accuracy of 96.30%, a recall rate of 94.91%, f1 of 95.26%, and iou of 91.57%. This proves that EA U 2 - Net can achieve better remote sensing image-building segmentation results. Finally, in view of the problem that the deep learning network relies on training samples, this study examined the influence of the number of building samples, sample purity, and sample resolution on the effect of building extraction. The results confirmed that reasonable sample parameter settings can improve the target extraction accuracy and the optimal sample parameter combination was verified in this experiment.