Sea Ice Classification Using Edge AI for Synthetic Aperture Radar Satellite On-board Image Processing

Kurzfassung

In this contribution we present automated sea ice classification on Sentinel-1 imagery for satellite on-board processing, which has the potential to significantly speed up product delivery by enabling direct downlink to stakeholders. Synthetic aperture radar (SAR) satellite imagery serves as an important data source for the creation of sea ice charts. Existing ice services today largely rely on manual expert analysis, an approach that is challenged by the trend towards large constellations of satellites with frequent acquisitions. Methods involving AI have been developed to automate ice charting, but considerable time already passes from image acquisition to downlinking raw data to a ground station for processing. In the scope of the dAIEDGE Network of Excellence, a neural network for semantic segmentation is trained on the AI4Arctic challenge dataset and implemented on an FPGA-based AI accelerator. By applying optimization techniques such as quantization and pruning, our on-ground prototype is able to classify a full image (25 Mpx) in 38 seconds. An F1 score of 80.6% across the test dataset is achieved, competitive with publications using the same dataset. Our results show the feasibility of deploying re-configurable edge AI directly on-board of satellites by utilizing dedicated hardware acceleration.