A UNet-Like Hybrid Transformer for Efficient Semantic Segmentation of Remote Sensing Images

In recent years, high-resolution remote sensing images have played a key role in areas such as precision agriculture, environmental protection, and economic evaluation. The rapid development of deep learning techniques has made Convolutional Neural Networks (CNN) and Transformers the dominant approaches in semantic segmentation. However, the traditional Transformer architecture suffers from computational and memory burdens when processing large-scale high-resolution remote-sensing images. At the same time, the local nature of CNN convolutional layers limits the network to capture the global context. To address these challenges, this paper proposes a hybrid CNN- Transformer network based on the U-Net architecture, which achieves efficient semantic segmentation of remote sensing images by employing resnet18 in the backbone network and introducing the BiFormer attention module in the decoder. The experimental results show that the proposed network achieves 92.5 % F1 and 86.68% mIoU on the Potsdam dataset and 6.59G less in terms of model computational power (FLOPs) than the best existing method, UNetFormer, for the same input case. As a result, our approach is not only lightweight and efficient but also exhibits excellent performance in semantic segmentation.