Self-Supervised Learning for Hyperspectral Remote Sensing: Opportunities and Limitations

Kurzfassung

The buzz of "Foundation Models" reached the remote sensing world. Hyperspectral satellite missions such as DLR's EnMAP, the Italian PRISMA mission, and NASA's EMIT sensor make available large amounts of hyperspectral imagery. To comprehend the rich amounts of information contained in spatial-spectral data cubes, we utilize self-supervised learning techniques to generate semantically informed "embeddings".At the same time, the hunt is on for applications that make a difference when compared to multi-spectral data such as from the EU Sentinel-2 or the US Landsat-8 missions. I will explore the value of EnMAP for applications such as land cover classification, mineral mapping, tree species identification, and trace gas detection. With boots on the ground I outline challenges hoping to spark a lively discussion with the audience on promising future direction for hyperspectral remote sensing with artificial intelligence.