Seyrek Hiperspektral Karışım Giderimi İçin Çizge Düzenli Bolluk Tahmini ve Sözlük Budama Yaklaşımı

Abstract

The main focus of this dissertation is semi-blind hyperspectral unmixing. Semi-blind hyperspectral unmixing, also called sparse unmixing, has opened a new direction to the field of spectral unmixing and has become a very popular technique with promising results over the last decade. In sparse unmixing, the spectral library is given beforehand as prior information to estimate unknown abundances in the scene by sparse regression techniques. In this dissertation, a graph Laplacian regularized unmixing method is proposed that can manipulate the latent structure of hyperspectral images to improve the estimation of abundances. Affinity matrices are constructed by using pointwise mutual information to exploit the fact that pixels with spatial-contextual proximity and spectral similarity exhibit high statistical dependencies. Furthermore, a double reweighted sparse regularizer is used to impose sparsity on the estimated abundances.

A spectral library of materials can be huge in size (e.g. can contain hundreds of material signatures) and this limits the performance of sparse unmixing methods. To boost the performance of sparse unmixing methods, making them more accurate and time-efficient, a dictionary pruning framework is also proposed within the scope of this dissertation. The proposed pruning approach is based on quantitative assessment of the dictionary elements. The total utility metric, which has a negligible computational cost, is used to figure out the relative influence of each dictionary member.

The effectiveness of the proposed methods is demonstrated by providing extensive comparisons with well-known approaches in the literature using different types of data sets.