Classification of hyperspectral images with ensemble learning methods

Abstract

Hyperspectral imaging is a remote sensing technology that enables the acquisition of hundreds of consecutive bands in high frequencies. Hyperspectral sensors allow to capture images between 10-20nm wavelength intervals by operating in an area called the optic region of the electromagnetic spectrum. The application of this technology is increasing day by day in a number of disciplines including the defense industry, chemistry, forestry, agriculture, urban planning, and medicine. Developing hyperspectral imaging increases the need for advanced analysis of these images. For this reason, hyperspectral image processing subjects are being processed frequently in the fields of pattern recognition and machine learning. In this thesis, multiple instance learning, multiple classifier systems and kernel methods are emphasized in order to increase classification performance and to perform advanced image analysis. The use of spatial information in the proposed methods is also emphasized. In the proposed multiple instance ensemble learning approach, the use of unlabeled areas on hyperspectral images was provided. Methods such as bagging and random feature subspace selection have been used to increase the classification performance. In this approach, base classifiers such as decision trees, support vector machines, and k-nearest neighbors are used. Combining more than one kernel methods provides an efficient way to manage data with a compound distribution, such as hyperspectral images. However, the proposed multiple kernel learning methods often require complex optimization procedures. In order to address this issue, a boosting-based ensemble learning method is presented. Hybrid kernels are taken into consideration together with composite kernels which allow the use of spatial information, and it is aimed to perform advanced hyperspectral image analysis. Although this proposed method has shown high performance, the ratio of hybrid and composite kernels should be determined manually. For this reason, another method called multiple composite kernel extreme learning machine is proposed. In this method, hybrid and composite kernels are presented as an aggregated input, and the weight value of each kernel is determined automatically with an extreme learning machine based optimization algorithm. Since the extreme learning machine allows for multiple classification, the overloading calculation time is avoided and the result is achieved in a less complicated way. The proposed methods have been tested on hyperspectral images with ground-truth information. Obtained results are compared with state-of-the-art methods in the literature. Both numerical and statistical methods are used in these comparisons. In addition to that, the obtained classification maps are presented together with the experimental results for comparison purposes.