An improved SegFormer-based network for semantic segmentation of rural remote sensing images

Abstract

Semantic segmentation of rural remote sensing images is a challenging task due to complex land-cover distributions, large intra-class variations, and imbalanced category distributions. Although Transformer-based segmentation networks have achieved promising performance, improving the segmentation accuracy of lightweight models remains important for practical remote sensing applications. In this paper, we propose a lightweight attention-enhanced SegFormer network for rural remote sensing image semantic segmentation. The proposed method adopts SegFormer as the baseline architecture and introduces an efficient channel attention refinement module to enhance discriminative feature responses. In addition, a hybrid loss function combining cross-entropy loss and Dice loss is employed to alleviate the influence of class imbalance and improve the learning of minority categories. The experimental dataset consists of the rural subset of the LoveDA dataset and some non-public datasets. The results show that compared with the baseline SegFormer model, the proposed method maintains a lightweight network structure, achieves superior segmentation performance in terms of IoU and accuracy for specific land cover categories such as farmland, roads, and water areas. Meanwhile, it obtains slight improvements in mIoU and overall accuracy. The results indicate that the proposed method is suitable for rural land-cover mapping from high-resolution remote sensing images.