Raw SAR Data Compression with Deep Learning

Most current on-board raw SAR data reduction techniques are based on adaptive quantization. Next-generation SAR systems will be drastically more data intensive, with increased swath sizes, improved spatial resolution and multi-frequency acquisitions. Thus, new data compression approaches that significantly decrease the bits to be downlinked as well as stored on-board are paramount. Neural data compression is an emerging scheme where the encoder, decoder and entropy model are learned simultaneously and tailored for a given dataset. In this work, we apply this to raw SAR data from the Sentinel-1 stripmap mode. We utilize convolutional neural networks to represent the analysis and synthesis transforms and learn an entropy model such that the data are encoded with low bit-rate. Preliminary results on this scenario enable a significant reduction in the average bits-per-sample utilized, while still maintaining high quality reconstructions with respect to amplitude and phase metrics. This illustrates the great potential of neural data compression for raw SAR data towards more efficient data systems.