Efficient In-Orbit CNN Updates

Building Artificial Intelligence (AI) models for Earth Observation (EO) satellites can be a difficult task since there is often no, or very limited, real training data available for a particular sensor. Even if data exists from a previous mission, a change in the sensor may make the model unsuitable. Therefore, being able to update the trained model, using the actual data provided by the sensor after the satellite is in orbit, is highly important. Thus, the main purpose of this work is to investigate and develop how to update remotely a model placed in a satellite in orbit, and how to control or specify the update size using the training parameters. The goal is to improve the accuracy performance of the model, using new data acquired from the in-orbit satellite, and to achieve this performance improvement under the criterion of reducing the size of the model update that is uplinked to the satellite. The method proposed selects which Convolutional Neural Network (CNN) layer weights must be modified and which must be fixed during training in order to maximize the accuracy increase and minimize the update file size. For a sample network, and without the proposed method, results show an update size of 44.5MB for the retrained network (with an original network size of 48.9MB), with retraining using the new data (such as that acquired from a satellite post-launch) resulting in an accuracy improvement from 78.4% to 79.9%. With the proposed Efficient Network Update (ENU) method, the generated post-training network update is only 18MB in size, but still achieves an accuracy of 78.9%. This demonstrates the reduced data bandwidth requirement to update the model, while gaining accuracy on the original trained network.