A Unified ML Pipeline for Lunar Terrain Age Determination

Core idea

Lunar terrain age is traditionally estimated by manually counting impact craters — a slow, labor-intensive process. This work builds a scalable, automated alternative: a two-stage machine learning pipeline that first identifies genuine primary craters from a large, noisy catalog, then groups them into age-based clusters.

Methodology

1. Classification — a supervised model separates true primary impact craters from secondary craters and catalog noise across a dataset of 1.3M+ candidate craters.
2. Clustering — the confirmed primary craters are grouped into age-based clusters using an ensemble of unsupervised methods, combined by majority voting for robustness against any single algorithm's assumptions.
3. Validation — results are checked against known, named lunar terrains with independently established ages, including Mare Imbrium and Mare Frigoris.

Architecture

• Classifier: a ResNet-18-based model, achieving 95.6% validation accuracy at identifying primary craters, isolating roughly 1.06M primary craters from the original 1.3M+ candidates.
• Clustering ensemble: k-means, DBSCAN, HDBSCAN, and Gaussian Mixture Models, combined via majority voting into 4 age-based groups.

Results

The pipeline's age groupings align with the known relative ages of validated lunar terrains, supporting it as a scalable, reproducible alternative to manual crater-counting for large-scale lunar surface analysis.

Status

This paper was accepted at ICAITA 2026 (Beijing), but I wasn't able to attend in person to present it, so it won't be published from that track. I'm currently identifying a new venue to resubmit it to.