Abstract
Continual advancements in Earth Observation (EO) optical imager payloads has led to a significant increase in the volume of multispectral data generated onboard EO satellites. As a result, a growing onboard data bottleneck need to be alleviated. One technique commonly used is onboard image compression. However, the performance of traditional space qualified processors, such as radiation hardened FPGAs, are not able to meet current nor future onboard data processing requirements. Therefore, a new high capability hardware architecture is required. In previous work a new GPU accelerated scalable heterogeneous hardware architecture for onboard data processing was proposed. In this paper, two new CUDA GPU implementations of the state-of-the-art lossless multidimensional image compression algorithm CCSDS-123, are discussed. The first implementation is a generic CUDA implementation of the CCSDS-123 algorithm whilst the second is optimised specifically for multispectral EO imagery. Both implementations utilise image tiling to leverage an additional axis for algorithm parallelisation to increase processing throughput. The CUDA implementation and optimisation techniques deployed are discussed in the paper. In addition, compression ratio and throughput performance results are presented for each implementation. Further experimental studies into the relationships between algorithm user definable compression parameters, tile sizes, tile dimensions and the achieved compression ratio and throughput, were performed.