Recently, deep learning techniques based on Convolutional Neural Networks (CNN) have started to be used for the classification of hyperspectral images. These techniques present high computational cost when preprocessing stages are applied. In this paper, a GPU (Graphical Processor Unit) implementation of a spatial-spectral supervised classification scheme based on CNNs and applied to remote sensing datasets is presented. The scheme comprises convolution filters for processing the spectral information and a patch around each pixel to take the spatial information into account. To reduce the size of the filters, the dimensionality of the dataset is previously reduced using Principal Component Analysis (PCA). In order to achieve an efficient GPU projection, different techniques and optimizations have been applied such as the use of the deep learning framework Caffe. Speedups of up to 38.66× over the Pavia University dataset are obtained together with competitive classification accuracies.
Keywords: Hyperspectral, Classification, Convolutional neural network, Deep learn- ing, Caffe, GPU