HyVIC: A Metric-Driven Spatio-Spectral Hyperspectral Image Compression Architecture Based on Variational Autoencoders

To address the issue of spatial-spectral redundancy inherent in hyperspectral images and the insufficient consideration of this problem by existing methods due to the growing amount of remote sensing data, researchers proposed the HyVIC architecture based on variational autoencoders (VAE). This architecture utilizes a configurable spatial and spectral feature learning module, enabling independent control over both types of features to improve compression efficiency. The study indicates that the balance between spatial and spectral feature learning is crucial for reconstruction fidelity, and accordingly, an indicator-based hyperparameter selection strategy was proposed. Tests on two benchmark datasets showed that HyVIC can maintain high reconstruction quality across a wide range of compression ratios, with its BD-PSNR increasing by up to 4.66 dB compared to state-of-the-art methods. This research provides guidelines for learning variational hyperspectral image compression, and the relevant code and pre-trained weights are available publicly.