Sierra was born and raised in Haleiwa on the island of O’ahu. She is currently a senior at the University of Hawaii at Manoa pursuing a bachelors of Computer Science. She has an interest in Cosmology and Artificial Intelligence. After graduation, she plans to pursue a career with a focus on those subjects. In her free time, Sierra enjoys playing the guitar, spinning fire, and gaming.
Home Island: O’ahu
High School:
Institution when accepted: University of Hawaii at Manoa
Project Site: Space Systems Command (SSC), Kihei, Maui
Mentor: David Morris
Project Title: Deep Learning in Support of Space Domain Awareness
Project Abstract: Space domain awareness (SDA) is a term describing the study and monitoring of satellites and debris in Earth orbit. Machine learning models are currently being used by Space Systems Command (SSC) for this purpose. This project aims to improve current machine learning models used to track objects in space, as well as to optimize the current image processing pipeline. A critical aspect of SDA is the detection of stars to locate objects in space. Accurately identifying the positions of stars is essential for determining the precise locations of objects in Earth orbit, which is important for avoiding collisions and detecting changes. Existing models struggle with learning the shape of star streaks, which are elongated representations of stars in sidereal tracking images. These streaks can vary in length and angle, making it challenging for current models to predict them with a high enough efficacy and efficiency. In this project, we modify the current star detection neural network by appending convolutional layers to the feature extractor of the pretrained model, as well as to the end of the model, thus allowing the network more capacity to learn star streak angles and lengths. This project also aims to reduce image processing time and increase performance by analyzing and determining the optimal number of star annotations, and by re-engineering sections of the codebase in a more efficient manner. We demonstrate that these modifications drive improved model performance, increased annotation efficiency, and improved astrometric accuracy.