Semantic segmentation model for remote sensing images based on U-Net++ guided by dual attention

An essential component of the intelligent interpretation of remote sensing images is the use of semantic segmentation algorithms to assign feature class labels to individual pixels. Aiming at the problem of low segmentation accuracy of deep neural networks for small-scale objects caused by the large scale difference between different categories of objects in high-resolution remote sensing images, a U-Net++ guided by dual attention semantic segmentation model is proposed in this paper. In the encoding stage of the network, a dual parallel backbone network is constructed to extract features, and mutual attention is utilized to capture the dependencies between pixels of feature maps of different scales, adaptively fusing features of different scales with the same network depth to enhance the attention to small-scale objects. To address the issue of fine segmentation in complex scenes, a spatial and channel hybrid attention mechanism is introduced in the network’s decoding stage to reduce the semantic gap between the outputs of various depth sub-decoders while fusing the semantic information and spatial location representations at various levels therein. The proposed algorithm achieves notable performance metrics, with the mean intersection over union (mIoU) values of 86.77% and 82.73% on the Potsdam dataset and Vaihingen dataset, respectively, accompanied by the mean F₁-score of 92.32% and 90.79%. These results underscore the algorithm’s efficacy in delivering comprehensive segmentation of small scale objects, surpassing the performance of other state-of-the-art semantic segmentation algorithms such as U-Net++, FarSeg, DMAU-Net, and SAPNet.