Replsk: A Lightweight Backbone for Remote Sensing Semantic Segmentation

Recent research in remote sensing image processing has focused on how to get a broader range of prior knowledge. The utilization of such prior knowledge in the decoder part of the image segmentation network can significantly improve the network's accuracy in image segmentation tasks. However, pursuing lightweight design often restricts the backbone from effectively combining prior knowledge with local features. In this paper, we propose a lightweight backbone network(RepLSK) that combines accurate local information obtained through structural re-parameterization and extensive global information provided by large convolutional kernels. Compared to the baseline model, our model has improved in remote sensing image semantic segmentation tasks. It achieved a mean Intersection over Union (mIoU) of 87.3 % and 84.2 % on the Potsdam and Vaihingen datasets, with improvements of 0.5 % and 1.5 % compared with the baseline model. Additionally, the model's parameter was reduced by 3.6 %, and the floating-point operations (FLOPs) were reduced by 33.2 %.