Exponential Feature Extraction and Learning for Pixel-Wise Hyperspectral Image Compression

Hyperspectral images are captured over a wide range of the electromagnetic spectrum providing detailed information about the Earth’s surface. Hyperspectral imaging is widely used in agriculture, astronomy, molecular biology, physics, etc. Due to the very large size of information that the remotely-sensed hyperspectral data cube contains, transmission is a challenge. Various hyperspectral image compression techniques have been proposed in the last decades. We propose a new lossy compression technique that is based on the Fast Fourier Transform, negative exponential feature extraction, and machine learning. For the Pavia University data set, we obtained a compression rate of approximately 11, while preserving an important amount of the information in the original scene. Furthermore, we visualized the decompressed data starting from only two retained parameters and we evaluated the results by employing various metrics.