Abstract

The lunar south pole is a high-priority region for robotic exploration in support of goals to establish a sustained presence. However, the thermal and solar environment poses challenges for rover design, mission planning, and trafficability. Accurate thermal modeling in these regions has the potential to improve predictions of regolith properties, such as compaction, that affect mobility.

The Diviner Radiometer Experiment aboard the Lunar Reconnaissance Orbiter has generated terabytes of surface temperature brightness data since 2009. Deep learning offers a path to extract insight from datasets of this scale. Building on prior reference work in the literature that trained a Variational Autoencoder to learn latent representations of the thermophysical model from Diviner data, this thesis identifies gaps in the original procedure, particularly the lack of detail around data selection and characterization needed to achieve ≤ 10 K reconstruction loss and robust learning of all latent variables.

Two Gaussian Process Regression interpolation methods are compared: one uninformed (Profiles-v1) and one enforcing expected lunar temperature shapes (Profiles-v2). Profiles are generated from Diviner data collected between July 2009 and September 2023 across 47 areas of interest. Two sampling strategies, sample by metric and Principal Component Analysis sampling, are used to improve data diversity. Density plots are used to visualize dataset distributions. Across both interpolation and sampling strategies, reconstruction loss remains around 50 K. Results are consistent when evaluated against a third-party Variational Autoencoder implementation, suggesting the original reference work omits critical methodological details. This thesis contributes insights into how data selection and interpolation influence deep learning performance on lunar thermal data.