A comparative assessment of several processing chains for hyperspectral image classification: What features to use?

Hyperspectral image classification is a very active research area. Over the last years, several advanced feature extraction techniques have been integrated in processing chains intended for this purpose. In the context of supervised classification, the good generalization capability of machine learning techniques such as the support vector machine (SVM) can still be enhanced by an adequate selection of the number of features to be used for classification purposes. This number depends on the size of the available training set, which opens the way for the incorporation of supervised techniques for feature extraction in addition to more classic unsupervised ones. In this paper, we particularly investigate the issue of how many (and what type of) features can be used effectively for SVM-based classification. For this purpose, we consider different types of feature extraction strategies - unsupervised and supervised - in the context of different types of processing chains (all based on the SVM as the baseline classifier). We also explore the role of different dimensionality estimation techniques. Our study, conducted using a variety of hyperspectral scenes collected by different instruments, provides practical observations regarding the utility and number of features needed for different analysis scenarios.