Please use this identifier to cite or link to this item: http://hdl.handle.net/1946/20145
The first step in signal processing is to find an appropriate model for the observed signal. This is a very crucial task especially in remote sensing. From one hand, the model should be close to the reality and take into account all the facts and interferences along the signal acquisition. On the other hand, the model should be simple to work with. Therefore, there always is a trad-off between these two conflict facts.
In some applications such as remote sensing, the captured signal is usually degraded by different factors such as atmospheric haze and instrumental noises. Hence, signal restoration plays a significant role in remote sensing including hyperspectral imaging. Hyperspectral imaging is an imaging technique which provides contiguous spectra from a scene. Hyperspectral images have several applications in remote sensing such as unmixing, change detection and classification. Therefore, the main goal of the thesis is to find suitable models and estimating approaches for hyperspectral images to improve the result of the aforementioned applications.
The first major contribution of the thesis is to introduce Stein's unbiased risk estimator (SURE) as an automatic technique for model and tuning parameter selection (including rank selection) for hyperspectral image analysis. As a result, this thesis proposes two automatic SURE-based techniques for both hyperspectral rank selection and hyperspectral image restoration (called SPAMARS and SPAWMARS).
Proposing a new and fast restoration method (called FORPDN) based on classical linear model and inspired by the specific spectral characteristic of hyperspectral image is the next contribution of the thesis. That technique is improved by incorporating the sparsity property (called GLASSORPDN).
Also, several sparse reduced-rank restoration techniques are proposed for hyperspectral feature extraction and restoration and compared based on classification accuracies.
An important contribution of the thesis is to propose a novel reduced-rank wavelet-based model for hyperspectral images. Based on the proposed model, wavelet-based sparse reduced-rank regression (WSRRR) is suggested for hyperspectral restoration and feature extraction.
The restoration techniques proposed in this dissertation are validated based on image quality metrics such as signal to noise ratio and are also compared to similar techniques from the literature. Additionally, the consequences of those techniques on hyperspectral image classification are demonstrated. It is shown that the proposed methods are very competitive compared to other similar methods from the literature. Due to the very large volume of hyperspectral images, the memory consumption and computational time are considered for the derivation of the algorithms. Therefore, the methods given in this dissertation can be applied on real hyperspectral data.