Towards On-Board Image Compression using Vector-Quantized Auto Encoders
Bart Beusen, Xenia Ivashkovych, Stefan Livens, Dirk Nuyts, Tanja Van Achteren · 2023
In the CORSA project [1] we demonstrated an AI method for lossy image compression for Sentinel-2 data using the concept of vector quantized auto-encoders. We now extend this method to hyperspectral images and investigate the requirements to run the algorithm on-board. Trade-offs include the model size, number of bands that are compressed simultaneously, the use of separable convolutions and model quantization techniques.