Deep learning techniques for spacecraft detection and classification in orbital operations

As humanity’s in-orbit activities increased, so did the number of debris objects. The same can’t be said about policies that should have regulated such activities, especially in the first decades of space conquest. Today, after many years of unregulated reign, we face all the problems that this phenomenon brings. The rapid advancement of space exploration has driven the need for automated systems capable of accurately identifying and classifying spacecraft under various conditions. This thesis presents a machine learning-based approach to spacecraft detection and classification using image data. A two-stage process is implemented on the NASA PoseBowl dataset: first, YOLO’s object detection model is implemented to localize spacecraft within the image frames. Next, MobileNetV3, a classification model fine-tuned to detected spacecraft, leveraging cropped images to reduce background interference and improve classification accuracy. The classification model is trained on images from a chaser spacecraft’s perspective, achieving a high level of accuracy in both training and testing accuracy after extensive model tuning and refinement. This thesis work demonstrates that combining object detection with image classification significantly enhances the accuracy of spacecraft identification, offering a robust solution for future space operations such as docking, rendezvous, and other on-orbiting servicing operations. This research contributes to the growing field of autonomous spacecraft systems, with potential applications in satellite management, more in particular in space debris removal – a rising concern that needs to be addressed for a sustainable future.