Fourier-Swin-Unet: A Lightweight Network for Remote Sensing Image Semantic Segmentation

Remote sensing semantic segmentation is a crucial branch of numerous semantic segmentation tasks, and it is extensively utilized in urban planning, geographic information system construction, and other applications. Due to the improvement of image resolution and the increase in image data size, there are many model parameters in the current mainstream remote sensing semantic segmentation model. Meanwhile, with the increasingly complex urban structure, the difficulty of semantic segmentation tasks is gradually increasing. To address the aforementioned challenges, this article proposed Fourier-Swin-Unet, which leverages the Fourier transform to provide enhanced features and noise suppression capabilities. Combined with Unet using an encoding-decoding structure, it can capture more detailed information about images and achieve more fine semantic segmentation. Furthermore, Swin-Transformer can better capture long information through a self-attention mechanism, fuse local information of each image region, and improve semantic segmentation performance. The experimental results demonstrate that the model achieves an F1 score of 91.9%, an overall accuracy (OA) of 90.3%, and a mean mIoU of 90.9%. Compared to previous models, the relevant indicators on the dataset have improved by more than 5% on average, showcasing the superior performance of the proposed model.