An entropy-based speed up for hyperspectral data classification via CNN

The paper presents an empirical study concerning the use of Convolutional Neural Networks (CNN) for hyperspectral data classification. The type and the size of input data have been analyzed to both optimize accuracy and training time. An entropy-based method for selecting the number of CNN input features has also been adopted; for fixed operational setup, it allows to preserve accuracy rate, while greatly decreasing the training time. Experimental results refer to source spectra and their reduced version through principal components analysis (PCA) and show that a proper selection of principal components allows to greatly reduce the computing time, while still guaranteeing high classification accuracies.