2024, 49(2): 43-53.
doi: 10.12265/j.gnss.2023175
Abstract:
To facilitate the problems of low accuracy, fuzzy boundary, and difficulty in identifying small targets in building extraction using deep learning semantic segmentation networks, we propose an advanced attention gate U-Net (AA_U-Net) to improve the effect of building extraction. This network improves the structure of classic U-Net, using VGG16 as the backbone feature extraction network, attention-gated module participating in skip connection, and bilinear interpolation instead of deconvolution for upsampling. In the experiment, we use the Wuhan University building dataset (WHD) to compare the extraction effect of the proposed network and some classical semantic segmentation networks and explore the influence of each module of the network improvement on the extraction. The results show that the total accuracy, intersection of union, precision, recall rate, and F1 score of the network are 98.78%, 89.71%, 93.30%, 95.89%, and 94.58%, respectively. All evaluation indexes are better than the classical semantic segmentation network, and the improved modules can effectively improve the extraction accuracy. The problem of unclear outlines of buildings and fragmentation of small target buildings was improved, too. It can be used to accurately extract building information from high-resolution remote sensing images, which has guiding significance for urban planning, land use, production, life, and military reconnaissance.
