A new framework for hyperspectral image classification using Gabor embedded patch based convolution neural network

The contiguous acquisition of information in narrow wavelength in hyperspectral images poses complex problems in processing of the bands at various stages. The complexity arises due to the high dimensionality and the redundancy of information can easily be addressed with deep networks. In this research work, initially spatio-spectral features are fused by extracting the uncorrelated bands and exploit the texture patterns via exploratory factor analysis and Gabor filter respectively and embedded these features to the original cube underlying the assumption that the noise is heteroscedastic in each of the variable in factor analysis. Later, from the resultant Gabor embedded hyperspectral cube, extracted different number of patch cubes of sizes 25 x 25 x bands and trained an evolving newly designed deep network, three dimensional convolution neural networks, to classify the labels of hyperspectral cube. Experiments are conducted on the three bench mark datasets, namely, Indian Pines, University of Pavia and Salinas. The proposed method exhibits with high accuracy in performance over the state of the art methods as the convolution neural network is trained with Gabor embedded patches. The Overall Accuracy of the proposed method is 99.69%, 99.85% and 99.65% for Indian Pines, University of Pavia and Salinas dataset respectively.