SatViT: Pretraining Transformers for Earth Observation

Abstract

Despite the enormous success of the ’pre-training and fine-tuning’ paradigm, widespread across machine learning, it has yet to pervade remote sensing (RS). To help rectify this, we pre-train a vision transformer (ViT) on 1.3 million satellite-derived RS images. We pre-train SatViT using a state-of-the-art self-supervised learning algorithm called masked autoencoding (MAE), which learns general representations by reconstructing held-out image patches. Crucially, this approach does not require annotated data, allowing us to pre-train on unlabeled images acquired from Sentinel-1 & 2. After fine-tuning, SatViT outperforms state-of-the-art ImageNet and RS-specific pre-trained models on both of our downstream tasks. We further improve overall accuracy (by 3.2% and 0.21%) by continuing to pre-train SatViT—still using MAE—on the unlabelled target datasets. Most importantly, we release our code, pre-trained model weights, and tutorials aimed at helping researchers fine-tune our models. (https://github.com/antofuller/SatViT).