Spatial and Semantic Consistency Contrastive Learning for Self-Supervised Semantic Segmentation of Remote Sensing Images

A critical requirement for the success of supervised deep learning lies in having numerous annotated images, which is often challenging to fulfill in remote sensing semantic segmentation tasks. Self-supervised contrastive learning (CL) offers a strategy for learning general feature representations by pretraining neural networks on vast amounts of unlabeled data and subsequently fine-tuning them on downstream tasks with limited annotations. However, the vast majority of CL methods are designed based on instance discriminative pretext tasks, focusing solely on learning the global representation of the entire image while disregarding the essential spatial and semantic correlations crucial for semantic segmentation tasks. To address the above issues, in this article, we propose a spatial and semantic consistency CL (SSCCL) framework for the semantic segmentation task of remote sensing images. Specifically, a consistency branch in SSCCL is designed to learn feature representations with spatial and semantic consistency by maximizing the similarity of the overlapping regions of the two augmented views. In addition, an instance branch is introduced to learn global representations by enforcing the similarity of two augmented views from one image. Through the integration of the consistency branch and instance branch, the proposed SSCCL framework can learn robust and informative feature representations for semantic segmentation in remote sensing scenarios. The proposed method was evaluated on three publicly available remote sensing semantic segmentation datasets, and the experiment results show that our method achieves superior segmentation performance with limited annotations compared to state-of-the-art CL methods as well as the ImageNet pretraining method.