One to Two, Two to All: Towards Multimodal Self-supervised Learning for Earth Observation

Abstract

This dissertation investigates self-supervised techniques to learn generic representations from large-scale, unlabeled Earth observation (EO) data. With a focus on the multimodal nature of EO sensors, it curates a large-scale dataset and benchmark for pretraining and develops both modality-specific and joint multimodal self-supervised approaches for EO representation learning.