The conceptual division of remote sensing pictures is essential to remote sensing technology. However, predictions are hard to make because the main groups of these remote-sensing pictures are very complicated. Also, the things shown in shots from space are more involved, and many things in different groups are mixed. Because of this, it is hard to optimize based on the feature area. This study introduces a new non-supervised semantic segmentation method based on Mean Teacher (MT). This method is meant to make models more stable and feature-based class naming better. We also change things at the feature level. When we learn about features, we also use contrastive learning to ensure that things do not change when features change. The ISPRS Potsdam dataset and the challenging iSAID dataset have been used in many tests.