Research Projects

Highlighted research projects and technical deep-dives.





Work in Progress — Page under construction
Planetary Science Deep Learning Image Synthesis

Multimodal Image Transformation Transformer for the Lunar Surface

A unified any-to-any generation model that translates between heterogeneous lunar image modalities using a single transformer architecture.

2025

Overview

The lunar surface is observed through a wide variety of remote sensing instruments — reflectance maps, albedo images, thermal emission, topography (DTMs), slope maps, and more. Each modality captures different physical properties, yet not all are simultaneously available for any given region. This project presents a single, unified Transformer capable of translating any available input modality (or combination thereof) into any target modality, enabling virtual data synthesis for data-scarce scenarios.

Key Contributions

  • Any-to-any generation: A single model handles all pairwise and multi-source-to-target translation tasks across heterogeneous lunar modalities.
  • Multimodal conditioning: The transformer architecture accepts variable numbers of input channels, fusing multi-source information to improve synthesis quality.
  • Lunar data coverage: Trained on aligned multi-modal datasets derived from LRO instruments (LROC WAC, LOLA, Diviner, etc.).
  • Geophysical plausibility: Generated outputs are evaluated for physical consistency with real observations, not just perceptual quality.

Supported Modalities

Albedo (WAC) Topography (LOLA) Thermal (Diviner) Slope Maps Reflectance
arXiv preprint · 2025
arXiv Paper All Papers
Planetary Science Deep Learning Anomaly Detection

Lunar Technosignatures: A Deep Learning Approach to Detecting Apollo Landing Sites

Enhancing the search for anomalies on the Moon by evaluating state-of-the-art machine learning architectures against actual Apollo landing sites.

2025 Read more

Overview

Uncovering anomalies on the lunar surface is a critical step in exploring the Moon's geological history. Finding these unique data points often requires slow and biased manual inspections by domain experts. This research automates the search for technosignatures by using confirmed Apollo landing sites to ground truth and benchmark three state-of-the-art deep learning algorithms.

Key Contributions

  • Algorithm evaluation: Tested and compared three modern anomaly detection algorithms (EfficientAD, Cut&Paste, and AnoViT) on LRO Narrow-Angle Camera data.
  • Ground truth validation: Used the highly localized Apollo 15 and Apollo 17 landing sites to empirically grade the methods, replacing subjective interpretations.
  • Vision Transformer advantages: Demonstrated that the transformer-based AnoViT method outperforms convolutional network approaches in precision and providing clean global anomaly maps.
  • Open source deployment: All code, evaluation metrics, and pre-trained models are publicly accessible to push future lunar research.

Technologies

Vision Transformers (AnoViT) ResNet18 (CNNs) LRO NAC Data PyTorch
VISIGRAPP 2025
View Project Details

More projects coming soon.