This research is part of ScOSA (Scalable On-Board Computing for Space Avionics), the DLR (German Aerospace Center) research project dealing with on-board computers as a distributed system, which will be part of a DLR CubeSat mission. The research evaluates the efficiency of various autoencoder neural networks for image compression regarding satellite imagery. The results highlight the evaluation and implementation of autoencoder architectures and the procedures required to deploy neural networks to reliable embedded devices. The developed autoencoders evaluated, targeting a ZYNQ 7020 FPGA and a ZU7EV FPGA. The outcome appears in the successful implementation of an autoencoder for compressing infrared and hyperspectral images, with 64 and 21 compression factors, respectively, compared to classical compression techniques. The overall outcome showed promising results for the integration in the ScOSA project.