Performance Optimization of Hyperspectral Image Compression Algorithms

Hyperspectral imaging is widely used in various fields. This thesis refers to a specific remote sensing application. It aims to find and implement an algorithm able to compress hyperspectral images which are used to detect the presence of ice on metallic surfaces. The basic concept is that different materials bring to different spectral signatures, which allows us to recognize them through specific software. A hyperspectral image is obtained with a special camera which collects information of the same spatial scene at different spectral bands, i.e. narrow portions of the electromagnetic spectrum. As a consequence, the amount of data to be processed is huge, and has to be reduced without losing essential information for spectral recognition. Specifically, according to the state of the art, lossless compression algorithms bring a poor compression ratio even if they preserve totally the original information, while lossy ones achieve better compression, getting rid mostly of spectral and spatial redundancy. In the thesis, the first analysed technique uses PNG and JPEG standards applied to all images corresponding to each spectral band of the HSI. Both techniques are applied in combination with a two-dimensional Wiener filter, in order to get rid of the noise and obtain a better compression ratio. Another implemented and studied lossless technique is based on the Recursive Least Squares (RLS) filter, used for pixel prediction on the previous spectral bands. Eventually, the last algorithm investigated in this thesis, provides hyperspectral data representation in tensor notation and it is based on the combination of the two-dimensional Discrete Wavelet Transform (2D-DWT) and the Tucker Decomposition in its Alternating Least Square (ALS) implementation.