AI for Optimized Raw Data Quantization in SAR Systems

Next-generation SAR systems will feature high-resolution wide-swath acquisitions, resulting in a significant increase of the onboard data volume to be acquired by the system. This causes severe constraints in terms of onboard memory requirements and downlink capacity. In this scenario, an efficient onboard quantization of the raw data is of utmost importance, representing a trade-off between achievable product quality and consequent on-board data volume. In this paper, we investigate the use of artificial intelligence (AI), and in particular of deep learning (DL), for flexible and on-board SAR raw data quantization. The aim is to derive an optimized and adaptive data rate allocation given a set of desired performance metrics and requirements in the resulting focused SAR image without relying on a priori information on the acquired scene. The obtained bitrate maps (BRMs) can then be dynamically used as input to a state-of-the-art BAQ quantizer to perform the on-board raw data digitization. The proposed method aims at directly linking the characteristics of the SAR raw data to performance parameters computed in the focused SAR domain, without the necessity for performing on-board focusing. For optimizing the proposed DL model architecture, we consider multiple target performance parameters such as the Signal-to-Quantization Noise Ratio (SQNR), the InSAR coherence loss or the interferometric phase error, extending the capabilities of the architecture and, ideally, providing multiple bitrate estimations for a single input scene at a time, depending on the specific application requirement. The proposed method allows for an efficient joint optimization and reduction of the data rate and of the resulting performance setting a new paradigm for data reduction in future SAR systems.