Hyper-Spectral Image Compression by Joint Spatial Spectral Dimension Reduction Using Thresholded Principal Component Analysis

Hyperspectral image analysis techniques have advanced considerably in the past decade, resulting in specialized hyperspectral applications. Because the spectral dimension of hyperspectral cubes is very broad, preserving a large number of these is a challenging task. To address this challenge, various compression methods have been employed over the past decades. Dimensional reduction methods are the most straightforward for compressing hyperspectral images, however, a significant amount of processing power is required for reasonable performance, which is reflected in long run-duration and high costs. In this work, a new method of dimensionality reduction compression based on a principal component analysis (PCA) algorithm is presented. The spectral dimension was reduced by thresholding principal components (PCs), while the spatial dimensions were reduced by average pooling/max pooling close groups of pixels. The proposed method was compared to three common dimensional reduction methods and tested on a database of 60 hyperspectral cubes. In terms of compression time to PSNR, the proposed method outperformed the other methods. The evaluation results indicate that the proposed method optimizes the tradeoff between compression performance and low complexity to obtain maximum performance with minimal runtime.