A multi-service edge-AI architecture based on self-supervised learning

This paper presents the development of an efficient artificial intelligence (AI) methodologies for onboard satellite image processing. The proposed system features a fast and efficient neural network designed to process radiometrically corrected multispectral optical images directly onboard satellites. The architecture is composed of a backbone feature extractor that generates semantically meaningful feature representations of input data, which are shared with multiple task-specific heads for various applications, including image classification, segmentation, and object detection. Training employs a self-supervised learning approach, significantly reducing the need for labelled data, with only small application-specific datasets required. The flexible design allows new tasks to be added without retraining the entire model or making major code changes. To ensure suitability for onboard use, the model is optimized for efficiency and low energy consumption through the use of quantization techniques and efficient deep learning modules. Key applications include cloud segmentation, fire detection, and flood detection, which demand low-latency responses for early warning and damage assessment.