2023–24Undergraduate Research · Sonoma State
Eclipse Phase Classifier
Undergraduate research for NASA's Eclipse Megamovie 2024, which collected 80,000+ citizen-science photos of the April 2024 total solar eclipse. The images arrived unlabeled, so the goal was to train a model that could accurately sort them into eight categories: the stages of the eclipse from partial to total, and being able to determine which photos did not feature an eclipse. I built the keyboard-driven labeling tool that cut labeling from one to two minutes per image to a few seconds, which let the team build a human-verified training set much faster, and I helped train and evaluate the CNN that learned from it.
- —~88% accuracy across 8 classes vs ~70% for a SIFT bag-of-words baseline
- —Labeling tool took each image from 1–2 minutes to a few seconds
- —Oversampled the training set to correct a heavy "total eclipse" skew
PythonCNNOpenCVSIFT + KNN