Adaptive SAR signal compression through Artificial Intelligence

Main goal of the ARTISTE study led by DLR has been to investigate the use of artificial intelligence (AI) for efficient on-board SAR data compression: In a first project phase, the raw and processed SAR Data were prepared for the Deep Learning-based data compression. Then, novel CNN-based architectures have been developed in order to implement an optimized adaptive data rate allocation for SAR raw data, depending on the local SAR image statistics, as well as on the target performance degradation due to quantization. The applicability of deep autoencoders for on-board data volume reduction has also been evaluated on different SAR image processing "stages", i.e. raw data, range-compressed data and fully processed images. The achieved performance has been assessed using standard quality metrics, together with the achievable data volume reduction with respect to state-of-the-art quantization methods employed on current SAR missions. The feasibility of the selected algorithms and functionalities have been evaluated, and a baseline hardware (HW) architecture and possible building blocks (such as, e.g., high-performance FPGAs) for the on-board implementation of the developed methods for multiple dynamic SAR compression selection has been proposed. The main drivers for this activity where the fact that next generation SAR systems will offer a giant leap in performance using large bandwidths and digital beam forming techniques in combination with multiple acquisition channels. These innovative spaceborne radar techniques have been introduced to overcome the limitations imposed by "conventional" SAR imaging for the acquisition of wide swaths and, at the same time, of fine resolutions, and they are currently being widely applied in studies, technology developments and even mission concepts conducted at various space agencies and industries. Such significant developments in terms of system capabilities are clearly associated with the generation of large volumes of data to be gathered in a shorter time interval, which, in turn, implies harder requirements for the onboard memory and downlink capacity of the system. In this scenario, the proper quantization of SAR raw data is of utmost importance, as it defines, on the one hand, the amount of onboard data and, on the other hand, it directly affects the quality of the generated SAR products. These two aspects must be traded off due to the constrained acquisition capacity and onboard resources of the system.