A project led by the University of Oxford has trained a machine learning model in space for the first time, aboard a satellite. This achievement could revolutionize the capabilities of remote sensing satellites by enabling real-time monitoring and decision-making for a range of applications.
Data collected by remote sensing satellites is essential for many key activities, including aerial mapping, weather forecasting and deforestation monitoring. Currently, most satellites can only collect data passively because they are not equipped to make decisions or detect changes. Instead, the data must be transmitted to Earth for processing, which usually takes several hours or even days. This limits the ability to identify and respond to rapidly occurring events such as a natural disaster.
To overcome these limitations, a group of researchers led by DPhil student Vít Růžička (Department of Computer Science, University of Oxford) took on the task of training the first machine learning program in space. During 2022, the team successfully submitted their idea to the Dashing through the Stars mission, which announced an open call for project proposals to be carried out aboard the ION SCV004 satellite, launched in January 2022. During the fall of 2022, the team uplinked the code for the program to the satellite, which is already in orbit.
The researchers trained a simple model to detect changes in cloud cover from aerial images directly on board the satellite, as opposed to training on the ground. The model was based on an approach called multiple shots, which allows the model to learn the most important features to look for when it only has a few samples to train on. A key advantage is that the data can be compressed into smaller representations, making the model faster and more efficient.
Vít Růžička explained: “The model we developed, called RaVAEn, first compresses large image files into vectors of 128 numbers. During the training phase, the model learns to maintain only informative values in this vector; ones that relate to the change it’s trying to detect (in this case, whether a cloud is present or not). This leads to extremely fast training by using only a very small classification model.
While the first part of the model for compressing the newly seen images was trained on the ground, the second part (which decided whether the image contained clouds or not) was trained directly on the satellite.
Normally, developing a machine learning model would require several rounds of training using the power of a cluster of interconnected computers. In contrast, the team’s small model completed the training phase (using more than 1,300 frames) in about one and a half seconds.
When the team tested the model’s performance on new data, it automatically detected whether a cloud was present or not within a tenth of a second. This involved encoding and analyzing the scene of an equivalent area of approximately 4.8 x 4.8 km2 (equivalent to almost 450 football pitches).
According to the researchers, the model could be easily adapted to perform different tasks and use other forms of data. Vít Růžička added: “Having achieved this demonstration, we now intend to develop more advanced models that can automatically distinguish between changes of interest (such as floods, fires and deforestation) and natural changes (such as natural changes in leaf color over the seasons Another goal is to develop models for more complex data, including images from hyperspectral satellites. This could enable the detection of methane leaks, for example, and have key implications for the fight against climate change.”
Performing machine learning in space could also help overcome the problem that onboard satellite sensors are affected by harsh environmental conditions, requiring regular calibration. Vít Růžička said: “Our proposed system could be used in inhomogeneous satellite constellations, where reliable information from one satellite can be used to train the rest of the constellation. This could be used, for example, to recalibrate sensors that have degraded over time or undergone rapid changes in the environment.”
Professor Andrew Markham, who supervised Víto’s DPhil research, said: “Machine learning has huge potential to improve remote sensing – the ability to squeeze as much intelligence into satellites as possible will make space sensing increasingly autonomous. This would help overcome problems with inherent delays between acquisition and action by allowing the satellite to learn from onboard data. Vít’s work serves as an interesting proof of principle.”