Learning-Based Hyperspectral Image Compression Using A Spatio-Spectral Approach

Advances in hyperspectral imaging have led to a significant increase in the vol-ume of hyperspectral image archives. Therefore, the development of efficient andeffective hyperspectral image compression methods is an important research topic inremote sensing. Recent studies show that learning-based compression methods areable to preserve the reconstruction quality of images at lower bitrates compared totraditional methods 1. Existing learning-based image compression methods usu-ally employ spatial compression per image band or for all bands jointly. However,hyperspectral images contain a high amount of spectral correlations which neces-sitates more complex compression architectures that can reduce both spatial andspectral correlations for a more efficient compression. To address this problem, wepropose a novel Spatio-Spectral Compression Network (S2C-Net).S2C-Net is a flexible architecture to perform hyperspectral image compression,exploiting both spatial and spectral dependencies of hyperspectral images. It com-bines different spectral and spatial autoencoders into a joint model. To this end, alearning-based pixel-wise spectral autoencoder is initially pre-trained. Then, a spa-tial autoencoder network is added into the bottleneck of the spectral autoencoder forfurther compression of the spatial correlations. This is done by applying the spatialautoencoder to the output of the spectral encoder and then applying the spectraldecoder to the output of the spatial autoencoder. The model is then trained usinga novel mixed loss function that combines the loss of the spectral and the spatialmodel. Since the spatial model is applied on the output of the spectral encoder,the spatial compression methods that are optimised for 2D image compression canbe used in S2C-Net in the context of hyperspectral image compression.In the experiments, we have evaluated our S2C-Net on HySpecNet-11k that isa large-scale hyperspectral image dataset 2. Experimental results show that S2C-Net outperforms both spectral and spatial state of the art compression methods forbitrates lower than 1 bit per pixel per channel (bpppc). Specifically, it can achievelower distortion for similar compression rates and offers the possibility to reachmuch higher compression rates with only slightly reduced reconstruction quality. References1 F. Zhang, C. Chen, and Y. Wan, A survey on hyperspectral remote sensingimage compression, in IEEE IGARSS, 2023, pp. 74007403..2 M. H. P. Fuchs and B. Demir, Hyspecnet-11k: A large-scale hyperspectraldataset for benchmarking learning-based hyperspectral image compression meth-ods, in IEEE IGARSS, 2023, pp. 17791782.