Researchers Successfully Train Machine Learning Model in Outer Space for the First Time

2023-07-31

Researchers have successfully trained a machine learning model on a satellite in outer space for the first time. This achievement enables real-time monitoring and decision-making for a range of applications, including disaster management and forest logging. The project has been summarized in the publication "Rapid Model Inference and Training on Satellites," which is available on the preprint server arXiv. The work was also presented at the International Geoscience and Remote Sensing Symposium (IGARSS) conference on July 21, 2023. Data collected by remote sensing satellites forms the basis for many critical activities, including aerial mapping, weather forecasting, and monitoring forest logging. Currently, most satellites can only passively collect data as they lack the ability to make decisions or detect changes. Instead, the data must be transmitted to Earth for processing, which often takes several hours or even days. This limits the ability to identify and respond to rapidly emerging events, such as natural disasters. To overcome these limitations, a team of researchers led by Dr. Vít Růžička from the University of Oxford's Department of Computer Science took on the challenge of training the first machine learning program in outer space. During 2022, the team successfully proposed their idea to the "Journey Through the Stars" mission, which publicly solicited project proposals and is set to be executed on the ION SCV004 satellite launched in January 2022. In the fall of 2022, the team uploaded the code for the project to the already orbiting satellite. The researchers trained a simple model that can directly detect changes in cloud cover from aerial images on the satellite, which contrasts with training on the ground. The model is based on a method called "subsample learning," which allows the model to learn the most important features with only a small number of samples. A key advantage is that the data can be compressed into a smaller representation, making the model faster and more efficient. Vít Růžička explains, "The model we developed, called RaVAEn, first compresses large image files into a vector of 128 numbers. During the training phase, the model learns to retain only the information values in that vector that are relevant to the changes it is trying to detect—in this case, the presence of clouds. As only a very small classification model needs to be trained, the training speed is very fast." The first part of the model is used to compress newly seen images and is trained on the ground, while the second part (which determines whether an image contains clouds) is trained directly on the satellite. Typically, developing machine learning models requires the computational power of a set of connected computers for multiple rounds of training. In contrast, the team's micro-model completed the training phase in approximately 1.5 seconds using over 1300 images. When testing the model's performance on new data, the team automatically detects the presence of clouds within a tenth of a second. This involves encoding and analyzing scenes equivalent to an area of approximately 4.8 x 4.8 km² (equivalent to nearly 450 football fields). According to the researchers, the model can be easily adjusted to perform different tasks and use other forms of data. Vít Růžička adds, "After completing this demonstration, we now plan to develop more advanced models that can automatically distinguish interesting changes, such as floods, fires, and forest logging, from natural variations, such as seasonal changes in leaf color." "Another goal is to develop models for more complex data, including images from hyperspectral satellites. For example, this could detect methane leaks, which have a critical impact on climate change." Performing machine learning in outer space can also help overcome the problem of satellite sensors being affected by harsh environmental conditions, requiring regular calibration. Vít Růžička states, "The system we propose can be used for non-uniform satellite constellations, where reliable information from one satellite can be applied to the rest of the constellation. For example, this can be used to recalibrate sensors that degrade over time or experience rapid environmental changes." Source: https://techxplore.com/news/2023-07-successfully-machine-outer-space.html

RECENT AI NEWS

RECENT AI TOOLS