Towards Intelligent Compression of Hyperspectral Imagery

For space applications, communications bandwidth is often contested and limited. Limited communications is particularly evident for space vehicles in low Earth orbit, where space-based sensor platforms that collect excess amounts of data will often have difficulty communicating this data in a timely manner. Data compression is a necessary step in transferring large files, such as high resolution images, as it allows a system to make more efficient use of communications bandwidth. Sensing systems continue to evolve with increased resolutions and data rates that effectively increase the overall amount of data. Hyperspectral cameras are one type of imaging sensors that produce vast amounts of data relative to conventional camera systems. Due to the data increase, hyperspectral sensors can potentially benefit significantly from data compression. CNN-JPEG is a state-of-the-art, neural network-based compression framework. This algorithm is a lossy, end-to-end image compression system that has been adapted from previous literature to compress hyperspectral imagery collected from the Airborne Visible Infrared Imaging Spectrometer (AVIRIS) sensor. By applying the CNN-JPEG algorithm to hyperspectral imagery, we achieve upwards of 17× compression ratios over the original image on average for each spectral band. This increased compression ratio comes at the cost of decreased reconstruction quality. Additionally, the amount of compression delivered by CNN-JPEG is highly dependent on the content of the image. CNN-JPEG provides a unique option for space platforms in which high compression is desired with acceptable losses in image quality.