Development of hyperspectral image classification algorithms for unmanned aerial vehicles

Hyperspectral imaging is the measurement of light spectrum reflected from objects in many narrow wavelengths. High spectral resolution in hyperspectral images (HSIs) allows identification and discrimination of the land-cover materials. Therefore, hyperspectral imaging is used in various fields such as military, surveillance, mineralogy and agriculture. In these fields, classification of the pixels of an HSI is studied extensively. Sparse representation based classifiers have been a powerful tool for the classification purposes. These classifiers use the idea that the spectral pixels can be represented by only a few samples with same class label in a training dictionary. Recent studies have shown that use of spatial information in HSI in addition to spectral information increases the classification performance. In the scope of this thesis, two new sparse representation based classification methods which use both spectral and spatial information have been developed in order to increase the success of classification process. In the first study, multiscale superpixels (MSSs) are utilized to acquire spatial information in a local area using different region scales. Pixels in these areas are jointly classified by sparse representation classifier and then classification maps are formed. Guided filter (GF) is applied on these classification maps to improve the misclassifications near the edges. In the second study, the neighbor pixels having similar spectral characteristics with the test pixel are selected by spectral matching methods and others are ignored. To verify the feasibility of the proposed methods, the performance are evaluated over two widely used hyperspectral data sets. Experimental results demonstrate that the proposed algorithms exhibits good performance compared with other related methods in the literature.