Development of a spectral unmixing procedure using a genetic algorithm and spectral shape

Thumbnail Image
Date
2012
Authors
Chowdhury, Subir
Journal Title
Journal ISSN
Volume Title
Publisher
Lethbridge, Alta. : University of Lethbridge, Dept. of Geography, c2012
Abstract
Spectral unmixing produces spatial abundance maps of endmembers or ‘pure’ materials using sub-pixel scale decomposition. It is particularly well suited to extracting a greater portion of the rich information content in hyperspectral data in support of real-world issues such as mineral exploration, resource management, agriculture and food security, pollution detection, and climate change. However, illumination or shading effects, signature variability, and the noise are problematic. The Least Square (LS) based spectral unmixing technique such as Non-Negative Sum Less or Equal to One (NNSLO) depends on “shade” endmembers to deal with the amplitude errors. Furthermore, the LS-based method does not consider amplitude errors in abundance constraint calculations, thus, often leads to abundance errors. The Spectral Angle Constraint (SAC) reduces the amplitude errors, but the abundance errors remain because of using fully constrained condition. In this study, a Genetic Algorithm (GA) was adapted to resolve these issues using a series of iterative computations based on the Darwinian strategy of ‘survival of the fittest’ to improve the accuracy of abundance estimates. The developed GA uses a Spectral Angle Mapper (SAM) based fitness function to calculate abundances by satisfying a SAC-based weakly constrained condition. This was validated using two hyperspectral data sets: (i) a simulated hyperspectral dataset with embedded noise and illumination effects and (ii) AVIRIS data acquired over Cuprite, Nevada, USA. Results showed that the new GA-based unmixing method improved the abundance estimation accuracies and was less sensitive to illumination effects and noise compared to existing spectral unmixing methods, such as the SAC and NNSLO. In case of synthetic data, the GA increased the average index of agreement between true and estimated abundances by 19.83% and 30.10% compared to the SAC and the NNSLO, respectively. Furthermore, in case of real data, GA improved the overall accuracy by 43.1% and 9.4% compared to the SAC and NNSLO, respectively.
Description
xvi, 85 leaves : ill. (chiefly col.) ; 29 cm
Keywords
Genetic algorithms , Spectrum analysis , Image analysis -- Mathematical models , Dissertations, Academic
Citation