Pre-Training Applicability in Earth Observation
Deep Learning approaches nowadays achieve very good performance and accuracy. However, they often struggle to generalise well on unknown use cases if only a small amount of data is available. As a result, the training effort and time intensive dataset acquisition must be repeated for each new use case, leading to high costs in application development. Unfortunately - in contrast to already established pre-trained models such as ResNet or VGG, which are based on the generic image dataset ImageNet – there are currently no pre-trained models for satellite data or earth observation applications available. The PreTrainAppEO project aims to make the use of AI in the field of Earth observation and remote sensing more attractive and efficient by developing a methodology that uses the approach of pre-trained AI models in order to achieve better generalisation and be able to adapt those models to standard use cases in the respective field. In that sense, we investigate and compare the performances of various Meta-Learning approaches with the one of Transfer-Learning. Meta-Learning uses the idea of using plentiful data from other previously seen, related tasks in order to learn how to adapt to unseen scenarios, and then learn the new task more efficiently from just a small amount of data. The project is funded by the German Space Agency at DLR on behalf of the Federal Ministry for Economic Affairs and Climate Action (BMWK).