Robust semantic segmentation algorithm of remote sensing images is essential for flood 8 detection, land use, land cover and mapping applications. However, high-resolution remote sensing 9 images contain a large amount of ground object information, showing diversity and complexity 10 with large intra-class variance, small inter-class variance and low-class discrimination, which makes 11 semantic segmentation more difficult. Especially when segmenting the edge of objects in remote 12 sensing images, it is easy to have irregular shapes and segmentation errors inside the objects. Re- 13 cently, Unet series of semantic segmentation networks have become popular not only in the field of 14 medical image segmentation, but also in the field of general semantic segmentation. Unet is based 15 on the symmetrical encoder-decoder structure. As an improvement of Unet, Unet++ is composed of 16 Unets of different depths. The skip connections in Unet++ are redesigned to achieve flexible feature 17 fusion in the decoder. However, it still does not make full use of multi-scale feature information and 18 there is room for improvement. In this paper, we propose a new Unet called RSUnet for semantic 19 segmentation of remote sensing images and Adaptive Feature Selection Module, which fuses deep 20 feature and shallow feature to the extreme and adaptively selects useful feature during feature fu- 21 sion to solve the problem of poor segmentation caused by the migration of Unet series of networks 22 to the field of remote sensing image segmentation, especially the poor edge segmentation of objects. 23 Experiments show RSUnet with adaptive feature selection module performs better than Unet series 24 of networks and other mainstream semantic segmentation models in five public remote sensing 25 image segmentation datasets.