Open Access
Gerg, Isaac
Graduate Program:
Electrical Engineering
Master of Science
Document Type:
Master Thesis
Date of Defense:
March 27, 2008
Committee Members:
  • Richard Laurence Tutwiler, Thesis Advisor
  • material abundance map
  • FastICA
  • independent component analysis (ICA)
  • orthogonal subspace projection
  • hyperspectral unmixing
  • atgp
  • least squares with constraints
  • Hyperspectral imagery
  • SNR
  • reflectance
  • vertex component analysis (VCA)
  • VCA
  • orthorectifciation
Hyperspectral unmixing is necessary for material abundance map (MAM) creation. Most unmixing algorithms search for a set of pure pixels, called endmembers, of which all other pixels in the image are linear combinations of this set. Hence, endmember extraction is an important process in the creation of useful material abundance maps. Material abundance maps are created using a three step process: estimation of the number of materials in an image, unmixing of the image to determine the spectral signatures (endmembers) of these fundamental materials, and finally, some type of constrained least squares algorithm using the recovered endmember spectra to generate the abundance maps. This thesis formally evaluates the first two steps of this process using simulated and real data. First, a common material count estimation algorithm, known as virtual dimensionality (VD), is examined. Second, three endmember extraction algorithms are evaluated: Automatic Target Generation Process (ATGP), ICA-Based Endmember Extraction Algorithm (ICA-EEA) and Vertex Component Analysis (VCA). Finally, the derivation of a constrained least squares technique is given from which the results of the first two steps are used as input to create abundance maps on real data from the AVIRIS sensor. This process serves as a means of qualitatively evaluating the efficacy of these algorithms on live data from which ground truth information could not be realized. The three unmixing algorithms arise from different schools of thought. The ICA-EEA uses independent component analysis (ICA) to isolate endmembers present in a scene. Conversely, ATGP and VCA work on the principle of orthogonality in that endmembers are extracted by iteratively projecting the data orthogonally to the current span of detected endmembers. The two methods differ in how they select the next orthogonal direction. Additionally, the authors of VCA state ICA based methods cannot perform well due to the sum-to-one constraint in linear endmember mixing. The results of this thesis prove otherwise. All methods require that pure pixels are present in the scene.