Improved U-Net++ Semantic Segmentation Method for Remote Sensing Images

Abstract

Remote sensing image semantic segmentation has extensive applications in land resource planning and smart cities. Due to the problems of unclear boundary segmentation and insufficient Semantic information of small targets in high-resolution remote sensing images, an improved network TU net based on U-net++ is proposed. Secondly, the attention aggregation module of the base Transformer is introduced to capture global contextual information, replacing the original multi-level skip connections of U-net++. A cross-window interaction module is designed, which significantly reduces computational complexity and achieves a lightweight model. Finally, a dynamic feature fusion block is designed at the end of the decoder to obtain multi-class and multi-scale Semantic information and enhance the final segmentation effect. TU-net conducted experiments on two datasets, where OA, mIoU, and mF1 scores were higher than mainstream models. The IoU and F1 scores of small-sized target cars in the Vaihingen dataset were 0.896 and 0.962, respectively, which were 5% and 15.8% higher than the suboptimal model; The IoU and F1 scores of the trees in the Potsdam dataset are 0.913 and 0.936, respectively, which are 6.3% and 4.3% higher than the suboptimal model. The experimental results show that the model can more accurately segment small-sized targets and target boundaries.