Abstract
The growing demands for advanced space applications have intensified the interest in on-board satellite processing. High performance on-board computing is an essential capability to meet the high speed and throughput requirements of these applications. The main overheads of high performance computing are the large hardware resources and the high power consumption, which are major design challenges in power constrained embedded applications such as spacecraft. Investigating a high performance intelligent system-on-a-chip (SoC) based architecture for space applications is the main objective of this research. In the proposed architecture, a spectral decorrelation module for hyperspectral image compression is developed for FPGA-based System-on-a-Chip platforms for space applications. While various techniques have been developed for spectral decorrelation, the Karhunen-Loeve Transform (KLT) technique outperforms other techniques in term of compression performance. However, this algorithm consists of sequential processes, which are computationally intensive, such as covariance matrix computation, eigenvector evaluation and matrix factorisations and multiplications. These processes slow down the overall computation significantly and increase the latency. In the proposed spectral decorrelation system, the KLT is utilised for lossy compression and the reversible Integer KLT is utilised for lossless compression. The computation of the algorithm is deeply investigated from mathematical and hardware perspectives in order to achieve a feasible solution within limited hardware resources, and therefore, limited power budget. The novelty of this work lies in the new architecture for the acceleration of the Integer Karhunen-Loeve Transform computation on FPGA-based System-on-a-Chip platforms for lossless hyperspectral image compression. Moreover, the proposed KLT architecture for lossy hyperspectral image compression outperforms previously proposed hardware architecture, as the proposed architecture offers further level of parallelism, which is more significant for hyperspectral data with large number of spectral bands. The experiments of the proposed system on the AVIRIS and the Hyperion data showed an overall improvement to the level of parallelism of up to 4. 9%, 11. 8% and 18. 4 % for 8, 16 and 32 spectral bands, respectively. Furthermore, this work addresses the KLT computations for large number of spectral bands from a hardware perspective, which has not been addressed in other works. In addition, this work also proposes a novel eigenvalues / eigenvectors computing hardware algorithm for large symmetric matrices, which can reduce the number of required iterations for large symmetric matrices and can offer partial computations of the eigenvectors and eigenvalues. Consequently, this can improve the parallelism level not only for the KLT computations but also for other applications where the some of the eigenvectors / eigenvalues can be utilised in the next computation stage. Therefore, this work contributed toward improving the acceleration of hyperspectral image compression and other high performance applications, where the hardware and the power resources are limited, such as space applications.