Multilinear compressive learning and reconstruction of hyperspectral data with U-Net architecture for enhanced spectral imaging

Hyperspectral (HSI) data provide vast amounts of data which contain rich spatial and spectral information about each pixel of an image, such information is useful in detecting objects and identifying materials. However, the information provided by the HSI image is challenging to process. Hence, the compression of HSI makes it easier to utilize the data without the need for any heavy computation. Traditional compressive sensing techniques are adequate for lower dimensional signals, but they fail to preserve the essential spectral information provided by HSIs while performing compression. The proposed method aims to implement a Multilinear Compressive Learning (MCL) model for ideal compression of HSIs and a CNN based U-NET architecture is used for effective reconstruction of the compressed signal. The core of the algorithm involves two key blocks: (a) A multidimensional compressive sensing block for performing compression along every dimension, preserving spectral and spatial correlations. The employed sensing matrix is learned with respect to the given data for effective compression, (b) Reconstruction of the signal using a CNN based U-NET algorithm, which consists of encoders and decoders for effective reconstruction of the signal. Band selection is also performed in this technique to optimally select the most informative bands. The proposed model was implemented on the Indian Pines, Pavia University, and Salinas scenes datasets. The compression and reconstruction results were evaluated using performance metrics.