Abstract

The development of semi-supervised learning methods is essential to Earth Observation applications. Indeed, labeled remote sensing data are scarce and likely insufficient to train fully supervised models with good generalization capacities. Conversely, raw data are abundant and therefore it is crucial to leverage unlabeled inputs to build better deep learning models. This work addresses the problem of semi-supervised semantic segmentation from a multi-task learning perspective. In this context, we explore several auxiliary tasks (reconstruction, unsupervised segmentation or self-supervision), and corresponding unsupervised losses, to perform along with semantic segmentation. Our experiments show the potential of semi-supervised learning approaches in a life-like scenario, outperforming a classical supervised setting.