Transforming Remote Sensing: Launch of 10-Billion-Parameter 'RingMo' Model V3.0

The Aerospace Information Research Institute (AIR) of the Chinese Academy of Sciences has recently launched the "RingMo" Version 3.0, a remote sensing interpretation foundation model with over 10 billion parameters. This cutting-edge model marks a significant advancement in automated, high-precision analysis of remote sensing data and is expected to make waves in the sectors such as emergency response, land management, marine monitoring, and urban development.

Developed in collaboration with Prof. FU Kun’s team at AIR and Peng Cheng Laboratory, this model is designed to transform how data is processed and utilized in the era of big data. As advancements in remote sensing data acquisition continue to flourish, innovations in artificial intelligence are redefining data service models and enhancing the capabilities of remote sensing technology.

SUN Xian, a researcher at AIR, uses an analogy to explain the significance of model parameters: “If we think of the foundation model as a robot, the ‘parameters’ are like its ‘memory cells’ and ‘knowledge cells.’” He further explains that smaller parameter scales limit the model's ability to retain and process information, making it suitable only for simpler tasks. Larger parameter scales, however, enable the model to handle more complex information, allowing it to tackle challenging tasks with greater precision and details.

"RingMo" Version 3.0 introduces a breakthrough heat conduction-based architecture, which addresses the performance and speed limitations that often hinder remote sensing model inference. This innovation significantly enhances both learning and inference speeds. For instance, training 10 million data samples now requires only a third of the previous number of GPUs, while processing speeds have increased by about 2.4 times. 

The model has shown outstanding generalization and interpretation capabilities, effectively analyzing data from various sensors, including visible light, synthetic aperture radar, thermal infrared, and multispectral satellite data. It can also automate high-precision interpretations of data from different aerial platforms, such as aircraft and drones.

Moving forward, the research team will focus on expanding the deployment of "RingMo" Version 3.0 across numerous industries within the national economy.