MSFENet: urban remote sensing imagery segmentation model based on multiscale feature enhancement and coordinate attention

The classification and segmentation of remote sensing imagery are key technologies for tasks related to the recognition and application of remote sensing images. However, high-definition satellite imagery segmentation faces challenges such as large differences in landform scales and severe segmentation loss of small targets with unclear features. To improve the integrity of ground object target segmentation by the model and optimize the segmentation effect of multi-class and multiscale targets, this paper proposes the Multi-scale feature enhancement network(MSFENet) model with the U-Net as the basis. This model adopts a Multi-scale feature enhancement technique to better extract multi-scale feature information of the model, and uses a coordinate spatial attention algorithm to analyses the coordinate spatial and channel dependency relationships of features. In this paper, the model is tested on the WHDLD public dataset released by Wuhan University for the segmentation integrity of different categories of ground objects. A comparative analysis is conducted with models such as UNet++ and SwinUnet. The MSFENet model, with an mIOU value of 73.89% and an mPA value of 86.37%, demonstrates that its segmentation performance is superior to other commonly used semantic segmentation networks.