Satellite imagery provides immense potential for Earth observation applications. By training AI to analyse these images, we can monitor huge areas with minimal effort. However, there are many technical hurdles that need to be overcome to successfully use these images for application areas such as agriculture and ecology. Satellite imagery has much lower spatial resolution than ground-based images, and the higher resolutions satellite images (30cm) must be taken selectively as it is infeasible to capture the whole planet at that resolution. This leads to sporadic images being available for any site with only one image taken every few months, or several on the same day. To train an AI model on these images, we need to construct training sets of ground-based observations associated with positions in satellite images. The ground truth observations are also sporadic, and these combine to make training an AI algorithm on satellite images challenging due a lack of observations that are temporally aligned with images.

This project aims to develop sample efficient AI algorithms which require less ground truth observations to train accurate models. Both semi-supervised and unsupervised training methods will be adopted to train highly effective feature extractors using a minimum of labelled data. Highly successful semi-supervised learning algorithms, and unsupervised/self-supervised methods (which have achieved very high accuracy with just a few labelled examples or no labels at all (respectively) in generic image classification) will be customized and these methods extended for satellite images.