Date on Master's Thesis/Doctoral Dissertation


Document Type

Doctoral Dissertation

Degree Name

Ph. D.


Computer Engineering and Computer Science

Committee Chair

Frigui, Hichem

Committee Member

Gader, Paul

Committee Member

Nasraoui, Olfa

Committee Member

Yampolskiy, Roman V.

Committee Member

Amini, Amir


Multispectral imaging; Algorithms--Data processing


A hyperspectral unmixing algorithm that finds multiple sets of endmembers is proposed. The algorithm, called Context Dependent Spectral Unmixing (CDSU), is a local approach that adapts the unmixing to different regions of the spectral space. It is based on a novel function that combines context identification and unmixing. This joint objective function models contexts as compact clusters and uses the linear mixing model as the basis for unmixing. Several variations of the CDSU, that provide additional desirable features, are also proposed. First, the Context Dependent Spectral unmixing using the Mahalanobis Distance (CDSUM) offers the advantage of identifying non-spherical clusters in the high dimensional spectral space. Second, the Cluster and Proportion Constrained Multi-Model Unmixing (CC-MMU and PC-MMU) algorithms use partial supervision information, in the form of cluster or proportion constraints, to guide the search process and narrow the space of possible solutions. The supervision information could be provided by an expert, generated by analyzing the consensus of multiple unmixing algorithms, or extracted from co-located data from a different sensor. Third, the Robust Context Dependent Spectral Unmixing (RCDSU) introduces possibilistic memberships into the objective function to reduce the effect of noise and outliers in the data. Finally, the Unsupervised Robust Context Dependent Spectral Unmixing (U-RCDSU) algorithm learns the optimal number of contexts in an unsupervised way. The performance of each algorithm is evaluated using synthetic and real data. We show that the proposed methods can identify meaningful and coherent contexts, and appropriate endmembers within each context. The second main contribution of this thesis is consensus unmixing. This approach exploits the diversity and similarity of the large number of existing unmixing algorithms to identify an accurate and consistent set of endmembers in the data. We run multiple unmixing algorithms using different parameters, and combine the resulting unmixing ensemble using consensus analysis. The extracted endmembers will be the ones that have a consensus among the multiple runs. The third main contribution consists of developing subpixel target detectors that rely on the proposed CDSU algorithms to adapt target detection algorithms to different contexts. A local detection statistic is computed for each context and then all scores are combined to yield a final detection score. The context dependent unmixing provides a better background description and limits target leakage, which are two essential properties for target detection algorithms.