COMPRESSION WITH SIDE INFORMATION, REVISITED: TAKING ADVANTAGE OF GROUND SEGMENT IMAGE ARCHIVES

Nowadays vast archives of imagery of the Earth surface are available from past observations and keep being updated with new data. At the same time, compressions systems onboard satellites struggle to meet stringent bandwidth constraints. Previously acquired images could be fruitfully used as side information to reduce the amount of information to be transmitted, but they are unavailable to the satellites as they are stored in datacenters at the ground segment. In this work, we revisit the concept of distributed source coding where an encoder that has no explicit access to the side information can still compress to its conditional entropy, provided a suitable decoder exploiting it. In the past, this scheme was limited by the ability to model the correlation between the newly acquired image on the satellite and the archival data. However, we show that decoders based on deep learning can effortlessly address this shortcoming of older approaches. We present designs of compression systems where the onboard encoder operates independently on the images and ground-segment AI-based decoders decompress the encoded data together with the side information of previously acquired images to lead to a higher-quality reconstruction at a given rate. We experimentally show that significant rate-distortion gains can be achieved even when the side information is already compressed, distant in time or affected by clouds.