U-Net Machine Learning Model to Analyze Ice Formations on Pluto’s Sputnik Planitia

This thesis details the process of building a machine learning model that classifies and labels the ice sublimation pits across Pluto’s Sputnik Planitia. Using primarily generated data, I worked to build and train a U-Net model to run image analysis on these ice pits to be able to identify the pit shape and the center of the pit automatically. The U-Net, a variation on the Convolutional Neural Network, is a powerful architecture for working with low-dimensional data and with smaller data sets. U-Nets are both time-efficient and power-efficient models, even when trained on largely computer-generated or augmented datasets, making it the ideal architecture for building a model to analyze Pluto. By building on the data augmentation code from my advisor, Dr. Alan Howard, I generated sufficient ice pit images to train and tune the U-Net to be able to identify and label ice pits in new images.