| Subpixel information on the components within a pixel is usually desired for remote sensing imagery produced by the medium resolution (30m) sensors. This information includes the number of endmembers in the image scene, endmember spectra, and the endmember fractions for each of the image pixels. With its property of high spectral and relatively low spatial resolution, hyperspectral imagery provides us a perfect data source to explore subpixel information with linear spectral unmixing techniques. In this thesis, linear spectral unmixing algorithms are classified according to whether one is considering the unmixing constraints or requiring all endmembers as a priori. Two algorithms are analyzed and extended for endmember extraction and unmixing. The endmember extraction algorithm is developed based on the unconstrained least squares regression, while the unmixing algorithms are derived from the constrained energy minimization and constrained least squares regression, respectively. These algorithms are implemented in the IDL language and tested with both simulated and real hyperspectral data. The computed result shows a very good agreement with the truth data. The accuracy is within 2% for simulated data and reasonable for the real data. |
