Roof Segmentation of Remote Sensing Images Based on Improved UNET

Accurately extracting building roofs from high-resolution remote sensing images has important values in urban planning and monitoring. However, due to the dense buildings and complex backgrounds in urban areas, achieving high-precision segmentation of building roofs remains a challenging task. The UNet model is an effective technique for semantic segmentation (SS) tasks. However, the UNet encoder is relatively simple, and its feature extraction capability needs to be improved. To address these issues, this paper proposes an improved UNet model, which greatly improves its segmentation performane by introducing the MSCAN module and redesigning the decoder. Experimental results exhibit that compared with FCN, Deeplap-v3, and PSPNet, the improved UNet model achieves higher performance on the Cityscape val dataset and roof segmentation tasks.