Multiscale feature U-Net for remote sensing image segmentation

The segmentation and extraction of buildings in high-resolution remote sensing images has good application prospects in military, civil, and other fields. With a depth encoder–decoder structure, U-Net is a frequently used model for high-precision image segmentation. However, the design of U-Net makes it hard to retain the detailed information of edges when processing the building segmentation. Specifically, the low-level features extracted from the shallow layer and the abstract features extracted from the deep layer cannot be completely merged, resulting in inaccurate segmentation. In response to this problem, we design a new multiscale feature extraction module that extracts target information through three convolution kernels of different scales. Taking U-Net as the baseline, by replacing skip connections with this module, we propose a multiscale feature extraction U-Net. This method can perform secondary feature extraction on the shallow feature information in the skip connection, refine the detailed information, and narrow the semantic gap between the low-level features and high-level features. It can not only improve the ability of the network to extract multiscale feature information, from a larger range to more layers to extract the edge detail information of the building in the remote sensing image, but also increase the number of skip connections to reduce network overfitting. Experimental results on Massachusetts remote sensing data and Massachusetts building data show that the method proposed offers significant improvement in terms of precision and accuracy compared with the methods full convolutional network, U-Net, SegNet, and high-resolution network, with an F1 score of 88.73%, mean IoU of 91.15%, precision of 89.74%, accuracy of 97.36%, and recall of 87.74%.