Hyper Spectral Image Compression Using a Non-Linear Iterative Partial Least Squares Algorithm

Abstract

Hyper spectral image has drawn great attention in applications like crop-disease detection, geological mapping, remote-sensing, environmental monitor, and defense surveillance, etc. Nonetheless, the high dimensionality and big data volume due to many spectral bands present an important obstacle for efficient storage and transmission. To address this issue, we propose a hybrid compression framework that decreases both spatial and spectral redundancies and is capable of accurately modeling non-linearities. The discrete wavelet transform provides a frequency domain representation of the image features, minimizing the spatial redundancy. Tucker Decomposition exploiting spectral redundancy through higher-order singular value decomposition by exploiting inter-band correlations. These techniques assume linear relationships between parameters, and to get around this limitation, the proposed method includes a non-linear iterative partial least squares model, which effectively models the non-linear relationships. We validate the proposed method on four public hyperspectral datasets such as Cuprite, Pavia University, Indian Pines Corrected, and Salinas. Experimental results show a very high compression factor of 280, which is better than Tucker and CANDECOMP/PARAFAC (CP) decomposition methods by margins of 199 and 61, respectively. Peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM)—High PSNR and SSIM values are attained despite the high compression. In addition, the proposed method outperforms traditional tensor decomposition approaches such as CP and Tucker, and recent methods for hyperspectral image compression, including hybrid transform with wavelet-based transform coding, long short-term memory autoencoder based compression as well as convolutional neural network-based lossy compressive methods under consistent experimental conditions.