RepreSent: non-supervised representation learning for Sentinel images

The past decade has seen exponential growth in the quantity of satellite-based remote sensing data being generated. This explosion of data and the advent and maturation of deep learning technologies has led to a new era of data-driven science, which has revolutionised the field of remote sensing. Through these technological advancements, we can monitor, analyse and understand our Earth and its ecosystems like never before.

Deep learning has been successfully applied across a range of Earth observation tasks, such as land cover classification, change detection, deforestation mapping, and object segmentation. However, in the vast majority of these cases supervised learning methodologies, which require a large number of labelled data, were applied. The generation of these labels is a time-consuming and expensive task which is fraught with complexities.

The ability to exploit these vast archives of remote sensing imagery without the need for labels can drastically reduce the overheads of developing new deep learning models.

We develop non-supervised and meta-learning approaches capable of learning generalisable representations from unlabelled, or self-labelled data. Furthermore, we study how to fine-tune and transfer these large-scale generic representations for application across a wide range of geographic problems.



European Space Agency