Satellite-ground collaborative upgrade framework for onboard object detection based on semisupervised learning

Remote sensing image object detection algorithms based on deep learning encounter domain shift issues between training and on orbit data, degrading model performance. This paper proposes a satellite-ground collaborative upgrade framework using semi-supervised learning. It pre-trains the model, filters on-board data, and conducts semi-supervised training on the ground. Experiments with the SSDD dataset and real-shot SAR images show that, with YOLOv5s as the base algorithm, after 50 minute semi-supervised training, the pre-trained model’s mAP50 in the target domain improves by 12%. This validates the framework’s effectiveness in enhancing target domain performance and addressing domain shift.