CMID: A Unified Self-Supervised Learning Framework for Remote Sensing Image Understanding

Self-supervised learning (SSL) has gained widespread attention in the remote sensing (RS) and earth observation (EO) communities owing to its ability learn task-agnostic representations without human-annotated labels. Nevertheless, most existing RS SSL methods are limited either global semantic separable or local spatial perceptible representations. We argue that this strategy is suboptimal realm of RS, since required for different downstream tasks often varied complex. In study, we proposed a unified framework better suited images representation learning. The framework, Contrastive Mask Image Distillation (CMID), capable with both separability perceptibility by combining contrastive (CL) masked image modeling (MIM) self-distillation way. Furthermore, our CMID architecture-agnostic, which compatible convolutional neural networks (CNN) vision transformers (ViT), allowing be easily adapted variety deep (DL) applications understanding. Comprehensive experiments have been carried out on four (i.e. scene classification, segmentation, object-detection, change detection) results show models pre-trained using achieve performance than other state-of-the-art multiple tasks. code will made available at https://github.com/NJU-LHRS/official-CMID facilitate research speed up development DL applications.