This project aims to advance the state-of-the-art of onboard machine learning for small satellites and produce a new class of intelligent satellite systems capable of automatically refining onboard object detection models on satellites in orbit.
We plan to develop a solution based on semi-supervised learning algorithms since they support improving a pretrained model using the unlabeled images captured by the satellite. Such a solution allows efficient use of the resources (e.g., hardware and data) available in orbit since the satellite system incorporates the following tasks: autonomously processing the captured input images to filter those of interest, transmitting them to ground stations for further analysis, and automatically improving the recognition performance of the onboard object detection model.
The focus of this project is to automatically increase the model’s accuracy in detecting relevant events or objects, which results in enhancing the quality of the transmitted data and, therefore, promoting faster human intervention for critical events. Furthermore, it to aims to create a generic solution with considerable potential to help various satellite applications that employ object detection algorithms regardless of the domain, e.g., environmental monitoring and agriculture.
The primary outcomes of this project consist of robust computational methods based on SSOD to enable the automatic improvement of object detection models on small satellites in orbit. Regarding academic results, when allowed by the SmartSat CRC, we expect to publish our findings in the form of scientific papers targeting top-tier international conferences and journals in computer vision, machine learning, and remote sensing topics.
Finally, we expect our work will also deliver practical outcomes such as open-source implementations of our methods, annotated datasets, and further cooperation with academic and industry partners of the SmartSat CRC.