An experimental analysis of feature selection algorithms in hyperspectral image classification

Recently, hyperspectral images have been an attractive subject for many researches in remote sensing area since they provide abundant information due to their wide range of spectral bands. On the one hand, providing such a huge amount of data by hyperspectral images may lead to complexity and bring some redundancy due to high correlation among the hyperspectral bands. On the other hand, this redundancy often negatively effects the classification of hyperspectral data by imposing extra computational costs without providing any advantageous information to the performance of the classifier. Moreover, the redundancy or using more features may lead to a decrease in the classification accuracy, which is known also as Hughes effect. In order to reduce the redundancy and increasing the performance of the classification methods, feature selection algorithms have been carried out to remove irrelevant features and highlight the efficient features of dataset to achieve a significant accuracy with minimum costs. The feature selection methods are typically presented in three categories based on how they combine the selection algorithm and the model building: filter-based methods which select suitable features using a search method; wrapper methods that validate the selected features with a classifier; and embedded methods which utilize the profits of two prior methods. There have been many studies related to feature selection not only in developing novel methods but also in application of the methods to hyperspectral image classification. To our knowledge, there is no any general analysis over hyperspectral remotely sensed datasets, involving a wide range of feature selection methods to compare them in the same experimental environments. In this work, a comprehensive experimental analysis with seventeen mostly used state of art feature selection algorithms is conducted extensively analyzed with two well-known classifiers, that are K-nearest neighbours and support vector machines, on seven common hyperspectral remotely sensed datasets. The contribution of this thesis is to present an extensive benchmark study on using feature selection algorithms with hyperspectral datasets to help researchers to comprehend the behaviour of feature selection methods on different cases. The analysis of feature selection algorithms are carried out by considering different number of training samples and different number of ranked features count. Besides, the methods are assessed based on four evaluation criteria which are classification accuracy, stability of feature selection methods, ability of ranked features to separate the classes of a dataset, and computational cost. According to the results obtained from the experiments, the filter-based methods, which are improved by mutual information measures, are more profitable than the other filter-based methods, even wrapper and embedded techniques. Although, filter methods are known as unstable method, they achieve accurate classification results as well as low computational time. Some wrapper and embedded methods also perform significant classification accuracy while filter-based methods also enhance a higher level of generalization.