FEST: Feature Enhancement Swin Transformer for Remote Sensing Image Semantic Segmentation

The global context is crucial for the precise segmentation of remote sensing images. However, the large volumes and high spatial resolutions of remote sensing images make efficient analysis of the entire scene challenging for most convolutional neural network (CNN)-based methods. To address this issue, we propose to design an innovative framework for semantic segmentation of remote sensing images called Feature Enhancement Swin Transformer (FEST). Firstly, we utilize the Swin Transformer as the encoder and incorporates a Global Information Enhancement Model (GIEM) within each Swin Transformer block to reduce information loss and enable encoding of more accurate spatial information. Secondly, we introduce an enhanced decoding structure called Enhanced Feature Fusion Module (EFFM) with added enhanced channel and spatial attention modules to retain localized information while obtaining extensive contextual information. Finally, for loss calculation, we utilize the dice and cross-entropy loss to jointly supervise the model, aiming to achieve a competitive performance. We comprehensively evaluated FEST on the ISPRS-Vaihingen and Potsdam datasets. The results indicate that our approach has achieved significant improvements in semantic segmentation tasks compared to existing methods.