As the industry faces challenges in integrating Artificial Intelligence (AI) into Earth Observation (EO) satellites, Ubotica’s CTO, Aubrey Dunne, presented a paper on this topic at the recent European Data Handling & Data Processing Conference. His paper, titled ‘Efficient In-Orbit Convolutional Neural Network (CNN) Updates,’ tackles an important, but often ignored, issue related to deploying effective AI applications on satellites.
The Challenge: Building Robust AI Models for EO Satellites
Building AI models for Earth Observation (EO) satellites is challenging due to the scarcity of real-world training data for specific sensors and satellites. Remotely updating these models with actual in-orbit data is critical, a challenge that the paper seeks to address.
The issue of limited real-world data is exacerbated by the diverse array of sensors found on Earth Observation satellites. Changes in sensor configurations or in satellite pointing stability can render existing models less effective, even when data from previous missions are available. Thus, it’s crucial to update trained models with real-time data once the satellite is operational. The paper delves into methods for remotely updating these in-orbit models and outlines how to manage the update size through training parameters.
The objectives are twofold:
- Maximise Accuracy: Enhance model performance using fresh in-orbit data.
- Minimise Update Size: Reduce the data volume needed for the update, saving crucial bandwidth.
The Solution: Efficient In-Orbit Convolutional Neural Network Updates
The paper focuses on determining which CNN layer weights should be adjusted, and which should remain fixed during training, to both improve accuracy and minimise the update size for uplinking to in-orbit satellites. The method supports user control of the balance between update size and accuracy improvement. A Convolutional Neural Network (CNN) is a specialised AI algorithm that is ideal for application to image data. It segments images into smaller sections to identify key features, making it highly efficient for tasks such as object detection and classification. This capability is particularly advantageous for Earth Observation satellites.
The paper presents compelling results for a sample network. Without using the Efficient Network Updates (ENU) method, the update size was 44.5MB with an original network size of 48.9MB, achieving an accuracy improvement from 78.4% to 79.9%. However, with the ENU method, the update size was reduced to just 18MB while still achieving an accuracy of 78.9%. This shows the method’s efficacy in balancing update size with accuracy improvements.
Conclusion: A Step Forward in EO Satellite Technology
Aubrey’s presentation and paper offers a viable solution to a key challenge to deploying AI in space, and sets the stage for further innovations in the field.