Hyperspectral image compression with deep learning: A review

The integration of spectroscopy and digital imaging produces a three-dimensional data cube known as a Hyperspectral Image (HSI), where each pixel captures a spectrum spanning wavelengths from 400 nm to 2500 nm. HSIs have become increasingly indispensable across a wide range of applications, including remote sensing, military operations, medical diagnostics, food inspection and environmental monitoring. However, the rapid advancement of hyperspectral imaging technology and the growing reliance on HSIs have introduced significant challenges in storage and transmission due to their high dimensionality and substantial data volume. To address these challenges, various compression techniques have been developed, ranging from traditional methods to deep learning-based approaches. Traditional methods, such as wavelet transforms and discrete cosine transforms, have been widely used for decades but may now be deemed less effective compared to more advanced deep learning solutions. Deep learning-based techniques excel at learning complex patterns through extracting adaptive features, modeling non-linear relationships, and facilitating end-to-end learning, thereby offering superior performance in HSI compression. In this article, we provide a comprehensive review of deep learning-based HSI compression techniques, discussing their methodologies, advantages, limitations, and performance. A detailed comparison of these algorithms is systematically presented in Table Table 5, offering valuable insights for researchers and practitioners in the field.