Deep learning has achieved great success in learning features from massive remote sensing images (RSIs). To better understand the connection between feature learning paradigms (e.g., unsupervised feature learning (USFL), supervised feature learning (SFL), and self-supervised feature learning (SSFL)), this paper analyzes and compares them from the perspective of feature learning signals, and gives a unified feature learning framework. Under this unified framework, we analyze the advantages of SSFL over the other two learning paradigms in RSIs understanding tasks and give a comprehensive review of the existing SSFL work in RS, including the pre-training dataset, self-supervised feature learning signals, and the evaluation methods. We further analyze the effect of SSFL signals and pre-training data on the learned features to provide insights for improving the RSI feature learning. Finally, we briefly discuss some open problems and possible research directions.