DO WE STILL NEED IMAGENET PRE-TRAINING IN REMOTE SENSING SCENE CLASSIFICATION?

Abstract. Due to the scarcity of labeled data, using supervised models pre-trained on ImageNet is a de facto standard in remote sensing scene classification. Recently, availability larger high resolution (HRRS) image datasets and progress self-supervised learning have brought up questions whether pre-training still necessary for classification would HRRS or achieve better results target tasks. To answer these questions, this paper we both train from scratch fine-tune several datasets. We also evaluate transferability learned representations tasks show that outperforms one, while performance similar slightly lower. Finally, propose an model combined with second round in-domain images, i.e. domain-adaptive pre-training. The experimental state-of-the-art benchmarks. source code are available at https://github.com/risojevicv/RSSC-transfer.