Excellence
The number of satellites being launched into orbit is increasing rapidly every year, and with it the complexity and capabilities of each satellite continues to grow dramatically. Today, near full earth coverage by optical sensors is achieved daily by civilian spacecraft, and soon civilian SAR (Synthetic Aperture Radar) will achieve a similar daily coverage. The ever growing amount of spaceborne data will need new solutions to get that data to the ground, because the available downlink is always a significant bottleneck in space system design. More sophisticated on- board data processing and storage will allow future iterations of spacecraft to achieve higher performance in smaller packages. Current solutions present limitations in computational performance, memory capacity and performance, and data reliability in very small form factors. SOPHOS will design and implement enabling technology for high-end data products produced on-board spacecraft via the implementation of more power efficient high performance space processing chains for various Low-Earth Orbit (LEO) missions, with a focus on SAR, which is one of the most data intensive space applications currently used and thus, e.g. also high resolution optical applications are covered by the SOPHOS approach as well. This implementation will be achieved through the optimisation of the payload processing and data storage system accompanied by the use of COTS (Commercial Off-The-Shelf) components and the miniaturisation of high-performance hardware in combination with robust firmware and software with heritage in high-end space applications. SOPHOS will combine state-of-the-art industrial computing technologies (COTS) including high-end FPGAs and GPU equipped SoCs, along with advanced and scalable processing capabilities. The modules developed within SOPHOS will allow for higher data processing performance in small and nanosatellite platforms, with the ability to deliver more data from data-intensive applications including SAR earth observation.
Objectives
SOPHOS will consider advanced algorithms for on-board SAR data processing on-board raw data compression as well as full SAR image formation. These state-of-the-art algorithms are not currently available in an optimized and space-qualified form, although some of the partners have carried out preliminary investigations in previous H2020 projects (e.g. S3NET and S4Pro). We intend to develop optimized yet portable implementations of these algorithms and benchmark as well as run them on space-proven hardware within this project to comply with the target TRL 6. The corresponding specific and measurable objectives of SOPHOS map directly onto the requirements of future flexible, multi-user SAR missions and are further detailed in the following paragraphs.
Objective 1: Enhance data access to on–board mass storage
SOPHOS targets a storage technology independent architecture with high speed data rates of at least 4 Gbit/s, with the possibility to integrate modules with latest high capacity non-volatile and volatile storage devices, also in hybrid fashion, to support high instrument and downlink data rates as well as high speed access. Due to the SOPHOS ́s targeted modular architecture scaling of size, capacity, access rates and operational storage profiles is achieved beyond current space mass memory implementations. Synchronous flash devices are planned to be used to further increase the storage data throughput. The improvement in access rate is estimated to be about 10x higher than competitive products.
Objective 2: Miniaturisation of HW modules within the processing chain
By using advanced highly integrated chip platforms in combination with the New Space approach, using mainly COTS components (e.g.: automotive grade), an outstanding functional integration is made possible, as well as a reduction in power consumption for all functional modules of the anticipated architecture (i.e. processing and storage). By this, SOPHOS targets to miniaturise payload processing and mass memory HW boards/system by a factor of 2-4 in surface area, by reducing the number of interfaces and using standard physical interfaces. Since cube- and nano-satellites typically cannot accommodate regular size HW boards that are currently being used for high-performance processing, different standards for PCB sizes (i.e. CPCI Serial Space or nanosatellite formats) will be adopted for HW sizing in order to provide an easy scalability of the system and enhance the modularity.
Objective 3: Improvement in on-board SAR raw data compression
Efficient SAR raw data compression is of primary importance, as it directly affects the volume of stored data, and in turn the acquisition capability of the system, as well as the SAR imaging quality. SOPHOS proposes a novel approach for performance-optimized SAR raw data quantization which allows for a joint optimization of the resource allocation and, at the same time, of the resulting SAR products quality. The algorithm will be first implemented as a python prototype and tested to predefined performance requirements, and a benchmark with state-of-the-art SAR data compression standards will be provided.
Objective 4: Improvement in/Feasibility of on-board SAR image formation
SAR image formation onboard the satellite is indispensable to achieve very low latency information products. Onboard SAR image formation is also an indispensable prerequisite for many other applications based on higher- level information products derived from onboard image analysis, e.g. using machine learning techniques. SOPHOS proposes a high-quality focusing kernel to enable the generation of high resolution (spatial resolutions of at least 0.5m) SAR image products onboard the satellite.
Objective 5: Optimization of portable processing SW routines to enable enhanced on-board SAR processing
Development of portable and high-performance SAR image formation and adaptive compression SW modules is a major goal of SOPHOS. Aim of work is to implement portable SAR related SW with high impact to accuracy and near real-time provision of data products. It will combine and improve several approaches in order to increase portability, robustness and reliability for further space and terrestrial applications.
Previous developments performed by iTUBS/UZL within the S3Net and S4Pro (H2020 projects) with linear algebra as well as optical image and SAR processing routines proof that handcrafted solutions (e.g. individual assembler level optimizations as employed in OpenBLAS or OpenCV) show a lack of support for most recent hardware architectures and are providing significantly less than the achievable performance.
Depending on the routines significant acceleration is achievable by using e.g. a self-optimizing SW framework. The performance difference will be measured through application benchmarks conducted during the project.
Objective 6: Agility
Performing data processing close to any sensor is a trend that has been adopted widely for any terrestrial sensor used ranging for lidar on automobiles, cameras for surveillance, or cell phones. This trend will naturally spread upwards and find use in space, where high fidelity sensors are being used to generate even higher data volumes that demands more bandwidth to bring the data to ground. The data can with advantage be prepared and filtered on orbit to relay critical data with a priority to users, and to present a downlink cost saving to operators while presenting filtered metadata prepared on-orbit to any user. The full SpaceCloud system offers a novel solution that enables a fundamental change for processing data on-orbit with the option to filter data with different objectives based on state-of-the-art AI/ML techniques. Different AI/ML models can be applied to the same dataset to extract e.g. both flying objects, land based objects, and maritime objects in parallel. Furthermore, SpaceCloud offers an integrated solution that enables both newly developed code and legacy with years of development time to hosted and used on the same system, thus provide an unseen flexibility to users with minimum development effort. This is enables by the fact that SpaceCloud carries libraries like ROCm, TensorFlow, PyTorch, to orbit that can be re-used for updating and designing new functionality long after satellite launch.
Objective 7: Flexibility/Modularity
The SOPHOS hardware will target an objective of fitting in a very small physical envelope, capable of being integrating in current nanosatellite buses. This involves the design of the boards in a 10cm x 10cm board shape, with specific cut-outs and mounting hardware to be compatible with current nanosatellite hardware. Because of the small size of each individual board, this will involve a stack of more than one board being used to accommodate all the EEE components required by the mass memory unit. The SOPHOS mass memory unit will also be designed with modularity in mind, to allow for the potential expansion of storage capacity or the use of alternative external interfaces. It is well known that most space programs have a mission-specific set of payload interfaces, and designing the mass memory unit to accommodate different configurations of external interfaces is essential for its success as a product.
The Payload Processing Module is a heterogeneous onboard computing platform comprised of an AMD V1000 Series 14 nm embedded family of SOCs and the Microchip PolarFire FPGA, both selected for their tolerance to space radiation. The combination of the x86-64 bit SoC (incl. CPU and GPU) with an ARM based FPGA, which also serves as a trusted control node for the larger SoC, forms together with VPU accelerators like Intel Movidius MyriadX, a modular high performance compute solution suitable for a variety of computational tasks. The OS based on Linux provides a complete software stack, from the Linux kernel driver to compiler support and common libraries and software for machine learning. On top of this, SpaceCloud Framework constitutes the orchestration layer for containerised applications and services, which enabling onboard cloud computing, similarly to Earth-based cloud services.
The development of highly optimized SAR processing algorithms will take a modular approach in which individual signal processing operations remain independent. Common SAR signal processing operations, such as transform and interpolation routines, thereby become reusable in future. The software development plan also aims at flexibility by addressing both CPU and GPU architectures, to cater to a diverse set of space-borne hardware platforms.