Meta-learning to address diverse Earth observation problems across resolutions

Earth scientists often study remote sensing problems in isolation, limiting information flow. This work presents METEOR, an adaptive deep meta-learning model addressing Earth observation problems across resolutions. METEOR handles variable spectral channels and class numbers, adapting to new problems with few training examples using global land cover knowledge. It outperforms competing self-supervised methods on five downstream tasks, demonstrating its effectiveness in solving novel geospatial problems with limited labeled data.

Marc Rußwurm, Sherrie Wang, Benjamin Kellenberger, Ribana Roscher, Devis Tuia