SPM-MIM: A super-pixel based masked image modeling method for multitemporal remote sensing land cover classification

In this paper, we propose a super-pixel based multitemporal masked image modeling (SPM-MIM) method for multitemporal remote sensing land cover classification. Compared with the previous self-supervised masked image modeling methods, SPM-MIM masks and reconstructs images based on super-pixels, effectively avoiding the problem that tiny remote sensing scenes are completely covered, while better retaining the structural information of remote sensing scenes, so that the model can be trained with stronger feature extraction capabilities. In order to better combine multi-temporal information to achieve land cover classification, we propose a cross-temporal information complementary module (CTICM). CTICM interacts with shallow multi-temporal features in spatial dimension to promote cross-phase information complementarity in local scenes, and interacts with deep multi-temporal features in channel dimension to promote cross-phase information complementarity in global scenes. Experimental results on remote sensing images of Gaofen-1 Harbin area and Gaofen2 Dalian Lushun area show that our proposed SPM-MIM method has excellent land cover classification capability.