Please use this identifier to cite or link to this item: http://hdl.handle.net/1946/20837
Hyperspectral imaging systems have gained a great attention from researchers in the past few years. These systems use sensors, which acquire data mostly from the visible through the middle infrared wavelength ranges and can simultaneously capture hundreds of (narrow) spectral channels from the same area on the surface of the Earth. Thanks to the detailed spectral information provided by hyperspectral sensors, the possibility of accurately discriminating materials of interest with an increased classification accuracy is increased. Furthermore, with respect to advances in hyperspectral imaging systems, the spatial resolution of recently operated sensors is getting finer, which enables analysis of small spatial structures in images.
Without any doubt, classification (or mapping) can be considered as the backbone of most image interpretation in remote sensing. In general, supervised classification approaches classify input data by considering the spectral information (e.g., intensity value of each pixel for grayscale images or intensity vector for RGB or high-dimensional images) of the data to produce a classification map in order to discriminate different classes of interest, by using a set of representative samples for each class, referred to as training samples. This way, by using a combination of training followed by classification, maps are produced from imagery. However, most of the existing classification techniques have been developed for the analysis of multispectral images, and consequently, they are not usually efficient for the classification of hyperspectral images, which can provide a detailed spectral information. This brings up the question whether the currently available classification techniques will be able to handle high-dimensional data.
The main objective of this thesis is the development of efficient spectral-spatial classification approaches in terms of classification accuracies. Beside the importance of classification accuracies, another critical issue for the purpose of hyperspectral image classification is simplicity and speed of the applied approaches. Therefore, in this thesis, a special emphasis is given on proposing robust techniques in terms of classification accuracies as well as being fast. In order to increase the efficiency of the existing techniques and reduce the laborious task of user interaction, a further development of automatic techniques plays a key role in remote sensing data analysis. Such techniques can be used for handling real-time applications such as hazard monitoring and risk management. Three different strategies are considered in the thesis as described below.
In the first strategy, a spectral-spatial classification approach, which is automatic and provides good classification accuracies is proposed. This method is based on integrating a Support Vector Machine (SVM) with Hidden Markov Random Field (HMRF). SVM and HMRF are two powerful approaches for high-dimensional data classification and spatial information extraction, respectively.
In the second strategy, we propose to use adaptive neighborhood systems by considering different approaches based on image segmentation and attribute profiles. These techniques are considered in order to extract spatial information for the purpose of spectral-spatial classification. In order to extract spectral information, SVM and Random Forests are applied due to their good performance in handling high dimensional data with limited number of training samples.
Finally, due to the fact that hyperspectral remote sensors acquire a massive amount of data and obtain many measurements, not knowing which data are relevant for a given problem, the third strategy is using novel feature selection approaches in order to address the curse of dimensionality and reduce the redundancy of high dimensional data.