Primary Endmember Analysis for Compression of Hypercubes

This proposal addresses NASA?s need for methods to organize observed data for storage or processing using intelligent, goal-directed data compression. Spectral Sciences, Inc. proposes to develop an innovative, low loss data compression algorithm that provides significant data compression on many types of multidimensional remote sensing data, in particular for multispectral and hyperspectral imaging (MSI/HSI) data. The proposed approach is a novel application of convex matrix factorization. It utilizes proven MSI/HSI spectral analysis algorithms to compress datasets to their most information rich components, called endmembers, and endmember abundances. Compression factors as great as 50:1 are expected. Significantly, the endmembers provide a physically meaningful basis that can be used, even in the compressed state, to perform data analysis functions such as material classification. In Phase I, the existing algorithm will be modified for use as a tool for compression and demonstrated on MSI/HSI data. Tradeoffs between number of endmembers retained, speed and reconstruction accuracy will be assessed. The adaptive capability of the algorithm will be investigated. In Phase II the adaptive algorithm will be implemented and strategies for the compression, storage and reconstruction of the matrix products will be selected and implemented into a commercial compression product.