Onboard Machine Learning-based Compression of Synthetic Aperture Radar (SAR) Images Using FPGA/MPSoC Hardware

Kurzfassung

Synthetic Aperture Radar (SAR) data is a critical modality in remote sensing, offering unparalleled capabilities for earth observation under all weather conditions and during both day and night. Despite its numerous advantages, SAR missions face significant challenges in their ground segment, particularly in data handling and transmission. The number of ground stations available for data downlink cannot scale proportionally when the number of satellites and their sensing resolution increase, leading to bottlenecks in data transmission. Consequently, data compression plays a central role in SAR missions, as it can alleviate the burden on ground stations and improve operational efficiency. Recent advances in SAR data processing, such as onboard focusing, have further highlighted the need for effective data compression strategies. These innovations enable real-time SAR data processing and transmission, paving the way for efficient onboard data handling [1]. However, developing practical, high-performance compression solutions that meet the constraints of onboard hardware remains challenging. In this work, we present a comprehensive framework for the onboard compression of SAR data, addressing challenges from both algorithmic and hardware implementation perspectives. As a baseline, we employ the methodology described in [2], a state-of-the-art approach for simultaneous despeckling and compression of SAR images. The lightweight model, based on an autoencoder architecture, incorporates quantization and entropy encoding of the latent variables [3] - fundamental steps in traditional data compression. Additionally, the method leverages a self-supervised training strategy to effectively eliminate speckle, achieving competitive performance in both tasks. The model operates within a split-execution paradigm: compression occurs onboard the satellite (transmitter side), while reconstruction is performed on the ground (receiver side). This split enables efficient use of bandwidth and computational resources. However, translating this algorithm into a practical onboard implementation necessitates careful compromises between compression rates, reconstruction quality, and hardware constraints. A highly suitable hardware infrastructure for onboard SAR data processing is the Field Programmable Gate Array (FPGA), known for its flexibility and suitability for parallel computations. FPGAs are widely used in earth observation missions for their low-power consumption and potential for radiation tolerance. However, deploying Machine Learning models on FPGAs is non-trivial and requires significant expertise and design effort. To address these challenges, we employ Vitis AI [4]. The toolchain facilitates the quantization and compilation of deep learning models, easing their deployment on FPGA-based systems. We evaluate our implementation on a Zynq UltraScale+ evaluation board. Zynq devices are heterogeneous Systems-on-Chip (SoC), including FPGA-based programmable logic and CPU-based processing systems. In our development workflow, we refactor the dataflow of the compression algorithm to optimize its performance on FPGA hardware. Specifically, we focus on minimizing data movement between the programmable logic (FPGA) and the processing system (MPSoC), as these operations are often bottlenecks in such systems [5]. Additionally, we address key practical considerations such as model size, latency, and power consumption to ensure suitability for deployment onboard CubeSats, which face strict constraints in size, weight, and power (SWaP). Our results demonstrate that the proposed implementation achieves a favorable trade-off between task-specific metrics, e.g., peak signal-to-noise ratio and compression rate, and hardware resource utilization. Despite the challenges of migrating CPU- or GPU-based algorithms to FPGA-based systems, we achieve a light and fast solution viable for real-world onboard deployment. This work represents a step forward in developing efficient, hardware-compatible SAR data compression techniques, contributing to the broader goal of enabling scalable and efficient SAR missions in the era of NewSpace. References [1] L. P. Garc´ıa, G. Furano, M. Ghiglione, V. Zancan, E. Imbembo, C. Ilioudis, C. Clemente, and P. Trucco, “Advancements in Onboard Processing of Synthetic Aperture Radar (SAR) Data: Enhancing Efficiency and Real-Time Capabilities,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 17, pp. 16625–16645, 2024. [2] F. Sica, N. Foix-Colonier, and J. Amao-Oliva, “Self-Supervised Joint SAR Image Compression and Despeckling,” in IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium, pp. 2188–2190. [3] D. Minnen, J. Ball´e, and G. D. Toderici, “Joint Autoregressive and Hierarchical Priors for Learned Image Compression,” in Advances in Neural Information Processing Systems, vol. 31, Curran Associates, Inc., 2018. [4] AMD, “Vitis AI Overview v3.5 User Guide.” [5] S. Mandapati, U. Balss, and H. Breit, “Real Time Floating Point SAR Focusing on FPGA,” in EUSAR 2024; 15th European Conference on Synthetic Aperture Radar, pp. 60–65.