On-Board Hyperspectral Image Compression Using Vector-Quantized Auto Encoders

In the CORSA project [1] we demonstrated an AI method for near-lossless image compression for Sentinel-2 data using the concept of vector quantized auto-encoders. As part of the MOVIQ project, this compression model was adapted to the domain of hyperspectral data [2] and optimized to run on-board. In these previous works, the train dataset and test dataset belonged to the same overall dataset, being BigEarthNet (Sentinel-2) for [1] and HyspecNet-11k (EnMAP ) for [2]. We now test different variants of the model, including a quantized int8 version optimized for on-board processing. Furthermore, we investigate the transferability of a trained model for compression of EnMAP data to be used directly on PRISMA data that was not part of the training data set.