Context free band reduction using a convolutional neural network

In this paper, we present a method for content-free band selection and reduction for hyperspectral imaging. Here, we reconstruct the spectral image irradiance in the wild making use of a reduced set of wavelength-indexed bands at input. To this end, we use of a deep neural net which employs a learnt sparse input connection map to select relevant bands at input. Thus, the network can be viewed as learning a non-linear, locally supported generic transformation between a subset of input bands at a pixel neighbourhood and the scene irradiance of the central pixel at output. To obtain the sparse connection map we employ a variant of the Levenberg-Marquardt algorithm (LMA) on manifolds which is devoid of the damping factor often used in LMA approaches. We show results on band selection and illustrate the utility of the connection map recovered by our approach for spectral reconstruction using a number of alternatives on widely available datasets.