Available at: https://digitalcommons.calpoly.edu/theses/85
Date of Award
MS in Electrical Engineering
A computationally efficient, lossy, bandwidth compression scheme for hyperspectral imagery is presented. In certain cases, the direct application of the standard KLT algorithm over an entire data set is not practical. The data may be too large for efficient processing by one encoder and so portions of the entire data set may necessarily be sampled by different encoders. The algorithm proposed handles such a limitation by leveraging the standard JPEG 2000 technology. The component decorrelation of the JPEG 2000 (extension 2) is replaced with a two-stage Karhunen-Loeve Transformation (KLT) operation resulting in a reduction in the computational complexity. At the first stage the spectral data is partitioned into sets of adjacent spectral images, which are decorrelated independently of one another. At the second-stage inter-partition correlation is exploited by combining principle component images and performing a second set of KLT operations. It is shown that reconstructed image quality is improved by minimizing partition size at the first stage. The method is evaluated using the criteria of root-mean-square error as well as the performance of a terrain-classification algorithm. A relationship between partition size and minimal inter-partition decorrelation is also proposed and may act as a guide in the tuning of compression settings.