Concept

Deployment of edge-AI systems empowered with accurate DNNs for feature extraction is the most promising approach to reduce the burden of high-volume data load for satellites, ground-stations and ground-processing servers.

Edge-AI system can reduce satellites’ data downlink latency along and eventually opening the path towards development of novel and competitive EO space-based value-added services (space-VAS).

However, downsizing techniques for complex DNNs show strong limitations in retaining original processing accuracy, which hindered edge-AI applicability. As such, inability to produce small size yet accurate DNNs in the past has allowed deployment of only very basic image segmentation NNs on-board satellites, which has become a bottleneck towards development of more intelligent space-VAS.

Edge SpAIce’s approach aims at hatching full potential of edge-AI technology by drastically improving NN architectures reduction efficacy to prevent accuracy loss and bolster AI technology deployment in space by demonstrating its application in a challenging use-case scenario for marine plastic litter detection.

To do so, Edge SpAIce will develop a DNN (50-100M parameters) starting from publicly available database on remote marine plastic litters detection (e.g., MARIDA22), adopted and amended to actual target satellite, in parallel improving capabilities of both Agenium Space’s distillation tool and CERN’s open-source HLS4ML tool for accurate DNN optimisation and deployment onto SoC-FPGAs used in space.

With the additional purpose of reinforcing EU leadership role and competitiveness in the global space market, Edge SpAIce will further showcase deployment of the accurate-DNN developed during the first stage and optimized for edge computing onto Europe’s manufactured space dedicated FPGA from NanoXplore.