From natural images to spaceborne imagery: an empirical study of self-supervised learning for Earth observation

In this work, we provide an empirical study on the performance of self-supervised learning for spaceborne imagery. Specifically, we conduct extensive experiments on three well-known remote sensing datasets BigEarthNet, SEN12MS and LCZ42 using four representative state-of-the-art SSL algorithms MoCo, SwAV, SimSiam and Barlow Twins. We analyze the performance of SSL algorithms under different data regimes and compare them to vanilla supervised learning. In addition, we explore the impact of data augmentation, which is known to be a key component in the design and tuning of modern SSL methods