"AIR·RingSee" Integrated Machine Unveiled to Boost Remote Sensing Image Interpretation
Oct 27, 2023
A significant advancement in the field of remote sensing of China has emerged, as researchers from the Aerospace Information Research Institute (AIR), Chinese Academy of Sciences (CAS), unveiled their latest product – the "AIR·RingSee" integrated machine for remote sensing intelligence training and inference on October 26, 2023.
Collaborating with leading research organizations across China, AIR scientists developed this system which overcomes challenges related to lightweight foundational models and facilitates efficient deployment, offering a streamlined solution that includes data, models, platforms, and hardware. It significantly improves processing speed, particularly in scenarios requiring rapid data analysis.
Historically, remote sensing tasks, particularly image segmentation, posed a formidable challenge. The remote sensing foundation model generally takes more than an hour to process a single image for 50-kilometer by 50-kilometer area on a single GPU, far from meeting the demands of real-time data processing. "AIR·RingSee" boasts four core attributes: multi-task with high precision, efficient model training and inference, cost-effective flexible deployment, and comprehensive independent innovation.
In terms of precision, this model supports tasks covering nine major categories and 36 subcategories of downstream tasks, including land feature extraction, regional change detection, etc. Compared with traditional network models, "AIR·RingSee" has exhibited an impressive 6% to 12% increase in average multi-task accuracy.
Efficiency is another hallmark of "AIR·RingSee". Equipped with model fine-tuning technology, "AIR·RingSee" only needs to update less than 5% training parameters to achieve the comparable results on new tasks. In the Ascend (Shengteng) environment, the inference speed experiences a remarkable 3.5-fold increase. Moreover, the machine can be deployed in a variety of edge scenarios, such as satellite, airborne, or vehicle-based setups, at a significantly lower cost than the traditionally high-power, high-consumption large servers.
This "AIR·RingSee" integrated machine can seamlessly integrates with Ascend AI environment and the MindSpore framework, uniting software and hardware into a harmonious whole.
This innovation has profound implications for diverse industries, including natural resource management, infrastructure development, agriculture, and emergency response. For instance, in the complex task of fine-grained feature classification in complex environments, the model has exhibited a remarkable accuracy increase of over 20% compared to traditional methods, while drastically reducing the time required for image processing to just minutes per frame.
As for the future, the "AIR·RingSee" integrated machine will evolve continually, pushing the boundaries of remote sensing technology and opening up new frontiers in intelligent interpretation.