<p>Transformer together with convolutional neural network for semantic segmentation of remote sensing images has achieved better performance than the pure module-based methods. However, the advantages of both encoding styles are not well considered, and the designed fusion modules have not achieved good effect in remote sensing image semantic segmentation. In this paper, to exploit local and global pixel dependencies, improved gated recurrent units combined with fusion module, named feature selection and fusion module, are proposed. Concretely, to precisely incorporate local and global representations that are the outputs of encoders of ResNet and Swin Transformer, respectively, the ConvGRU with improved reset and update gates, which is treated as feature selection unit, is designed to select the features of advantageous segmentation task. To merge the outputs from ResNet, Swin Transformer and FSU, feature fusion unit based on stack and sequential convolutional block operations is constructed. On public Vaihingen, Potsdam and BLU datasets, experimental results show that FSFM is effective, which outperforms state-of-the-art methods in some famous remote image semantic segmentation tasks.</p>