A comparative study of different ICA algorithms for hyperspectral image analysis

Independent Component Analysis (ICA) is an unsupervised source separation method that has been widely used in signal processing. In recent years ICA has been exploited also in remote sensing image processing, and particularly in hyperspectral image analysis to extract useful information for classification and unmixing problems. In this paper a comparative study between the three most widely used ICA implementations (FastICA, Infomax, JADE) is presented. ICA is exploited in terms of feature extraction technique in order to retrieve class discriminant information for the supervised classification. Two different experiments are proposed and analyzed. Firstly, the spectral dimensionality is reduced by employing a supervised feature selection (FS), followed by ICA in order to obtain a new feature space. In the second experiment ICA is applied on the whole dataset and the most informative features are selected from the FS algorithm. The final results are compared and presented in terms of classification accuracies, obtained by Support Vector Machine (SVM).