Learning Deep Spectral Features for Hyperspectral Data Using Convolution Over Spectral Signature Shape

Deep convolutional neural networks learn the spatial image features automatically, for classifying a hyperspectral image. Learning the spectral features automatically is equally important in analyzing the hyperspectral image. However, most of the earlier work treat a hyperspectral pixel as a n dimensional vector (n = no. of bands) and a separate convolution is performed over the depth. The features so learned are stacked together with the spatial features and are used for further processing. The semantics of the learned spectral features are completely ignored and are not interpretable in these approaches. We propose a simple transformation of the hyperspectral pixel to two-dimensional spectral graph (shape) and then the convolution over the same. This results in learning the spectral features that can be interpreted using spectroscopic knowledge of the material. We compared our approach with some of the common deep learning approaches for the hyperspectral data. The improvements are evident from the experiments.