Regularized Singular Value Decomposition Based Multidimensional Convolutional Neural Network for Hyperspectral Image Classification

Abstract

This paper develops a novel regularized singular value decomposition-based multidimensional convolutional neural network (RSVD-MCNN) to extract spatial and spectral-spatial information from the hyperspectral image (HSI). A hyperspectral image consists of narrow spatial and spectral band information that makes the nonlinear, invariant and discriminant wavelength. The variable nature of this wavelength introduces a challenge to perform the precise classification of HSI. In this paper, the idea of low rank matrix decomposition in multidimensional deep learning algorithm is proposed for the first time in HSI to precise classification and target recognition. First, we develops a framework for the decomposition of each pixel information of HSI to get the more useful spectral information from the visible energy radiation in the entire detection bands. Second, we design a multidimensional convolutional neural network (CNN) carrying the combination of 3-D and 2-D CNN to classify the spectral-spatial semantic feature information from the HSI. We then propose a novel deep learning network combined with above two features, in which the proposed network provides the precise classification with near to 100% accuracy. An Experimental validation with mostly used HSI dataset is studied to measure the effectiveness of the proposed RSVD-MCNN architecture and compared with other relevant existing techniques.