CNN with ICA-PCA-DCT Joint Preprocessing for Hyperspectral Image Classification

In this paper a simpler convolutional neural network with a joint pre-processing is proposed for hyperspectral image classification. Primarily the spectral dimension of raw hyperspectral data cube is reduced in a unique fashion by using PCA and DCT such that the data is reduced effectively but having much information intact for classification task. The raw data cube is divided into two small spectrally reduced cubes, the first cube (PCA cube) is a simple PCA based dimension reduction considering few top principal components and the second cube (PDCT cube) performing DCT as preliminary step which confined maximum energy into low frequencies and then subsequently applying PCA by selecting same number of principal components as in the first PCA cube. After that both PCA and PDCT cubes are fused together, furthermore ICA is carried out on fused data cube to make the output classes much independent for next steps. In the final pre-processing step, the ICA performed data cube is divided into small square patches having labeled center pixel and a fixed size neighboring pixels by considering that in hyperspectral image neighboring pixels are highly correlated and having same class label. These square patches are fed into our simpler convolutional neural network which effectively and automatically extract the suitable features for our classification prediction job. The results validated that our acclaimed model which mainly exploit a novel pre-processing tactic and simpler but effective CNN performs enormously well in comparison to the other compared models and can be used as an effectual classification model for hyperspectral images in particular.