Lightweight TransUNet Informal Residential Area Segmentation

Remote sensing image informal residential area segmentation plays an important role in urban planning, environmental detection, and disaster assessment. With the development of deep learning, the performance of informal residential area extraction has been improved. However, trying to accurately extract the boundaries between different buildings in the same complex is still a challenging task, and at the same time, segmentation model training takes a lot of time. To address the above problems, we propose a deep learning-based segmentation method for remote sensing images of informal residential area instances, aiming to further improve the accuracy of informal residential area segmentation. This paper has the following three main contributions. First, this paper designs and implements a remote sensing image informal residential area instance segmentation method called TransUNet-IRA, which is a model based on TransUNet, and reduces the training file weights by reducing the number of encoder layers and changing the format of the dataset, which increases the computing speed and improves the performance of the model. Second, we combine DiceCE Loss with IoU Loss to form DiceCE-IoU Loss to further improve the model's performance in informal residential area segmentation. Finally, we collect a large amount of meter-level remote sensing image data from Guilin City, Guangxi Province, and Chenzhou City, Hunan Province, and construct a remote sensing image informal residential area instance segmentation dataset for these two regions based on these remote sensing images, which is annotated with three segmentations of the background, informal residential area, and single building in the formal residential area. After experimental validation, our method demonstrates performance advantages on the constructed dataset, not only outperforming classical models such as UNet, ResUNet, AttUNet, and SwinUNet but also improving on the original TransUNet.