Self-Supervised Representation Learning for Remote Sensing

Self-supervised representation learning has become a popular and powerful pre-training step for large vision datasets without ample availability of label data in recent years. Few other fields have as much available unlabeled data as remote sensing, making it a perfect fit for self-supervised representation learning. However, remote sensing data are unique in their variety, with different sensors, different spatial and temporal resolutions, and variable available spectra. To make full use of the available data, an architecture must be able to learn under these varying conditions. In this thesis, self-supervised learning is explored in the context of remote sensing, given the stated goal. A new architecture, called CIMAE, is proposed that allows for full flexibility of the channels used. In turn, CIMAE can be used to train on multiple datasets with different channels solving one part of the unique problem of remote sensing data