🔭 astrophysics

A Machine Learning Based Search for Lunar Anomalies

This paper evaluates the effectiveness of a Beta-Variational Autoencoder in analyzing Lunar Reconnaissance Orbiter imagery to successfully identify both natural geological anomalies and artificial spacecraft landings on the Moon.

Original authors: Cameron Kelahan, Daniel Angerhausen, Adam Lesnikowski, Valentin T. Bickel

Published 2026-08-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Cameron Kelahan, Daniel Angerhausen, Adam Lesnikowski, Valentin T. Bickel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Moon's Silent Diary and the Robot Detective

Imagine the Moon not just as a glowing rock in the night sky, but as a giant, silent diary written in stone. For billions of years, it has recorded every crash, every slide, and every change on its surface. In recent decades, humanity has sent a high-tech spy, the Lunar Reconnaissance Orbiter (LRO), to take millions of incredibly sharp photographs of this diary. These images are so detailed they can spot a single boulder from space. But looking at millions of photos by hand is like trying to find a specific grain of sand on a beach while blindfolded; it's impossible for human eyes alone.

This is where a branch of science called "machine learning" steps in. Think of machine learning as teaching a computer to be a super-smart detective. Instead of telling the computer exactly what to look for (like "find a crater"), we let it study thousands of pictures of "normal" moon dirt until it learns what the average surface looks like. Once it knows the "normal," it can instantly spot anything that looks weird or out of place. This is the core idea behind searching for "technosignatures"—which is just a fancy word for signs that technology, perhaps left behind by humans or even aliens, might be hiding on the surface. The big question is: Could there be something artificial, like a forgotten lander or a strange structure, hiding in plain sight among the rocks?

Teaching a Robot to Spot the Weird Stuff on the Moon

In this paper, a team of researchers decided to test a new kind of robot detective called a $\beta$ -Variational Autoencoder (VAE). You can think of this model as a digital artist that has spent years studying the Moon. Its job is simple: look at a picture of the Moon, try to redraw it from memory, and then compare its drawing to the original photo.

Here's how the magic happens:
The Moon is mostly boring and uniform—it's just gray dust and rocks. The robot learns what this "average" gray looks like. When it sees a patch of the Moon that is just normal dust, it draws a perfect copy. But when it sees something strange—like a fresh crater, a pile of rocks from a landslide, or even a piece of old space junk—it gets confused. It tries to draw the "normal" background, but the weird object doesn't fit. The difference between what the robot expected to see and what it actually sees is called an "anomaly score." The bigger the difference, the higher the score, and the more suspicious the spot becomes.

The researchers fed this robot a massive diet of data: about 1,000 images from the LRO, chopped up into tiny 64x64 pixel squares (roughly the size of a 32x32 meter patch on the ground). That's over 52 million little patches for the robot to study! The robot learned the patterns of the Moon's surface and then started grading every single patch it looked at.

What Did the Detective Find?

The team wanted to see if their robot detective was actually good at its job. They gave it two famous "suspicious" spots that other people had already found: Plaskett Crater and Paracelsus C Crater.

Plaskett Crater is a rocky slope where boulders have been tumbling down.
Paracelsus C has two sharp, angular boulders that look like they were left behind by a crash.

The robot didn't need to be told these were special. It looked at the images, calculated the scores, and immediately flagged these spots as highly anomalous. The average score for a normal patch was around 23, but these weird spots stood out clearly, proving the model could spot unique geological features without being programmed with a specific list of what to find.

But the real fun came when they tested the robot on technosignatures—the man-made stuff. They fed it images of famous landing sites, including:

The Apollo 11, 12, 14, 15, 16, and 17 landers.
The Surveyor 3 lander.
The Chang'e 3, 5, and 6 landers.
Even the crash sites of the Apollo 13 rocket and the Luna 24 lander.

The robot successfully identified most of these tiny, artificial objects as anomalies! It's important to note that the robot didn't give these man-made objects the highest possible scores. Why? Because the robot was trained on patches that were 32x32 meters wide. A tiny lander is just a speck in that big patch, so it doesn't dominate the picture like a giant crater does. However, the robot still spotted them significantly better than random chance. In fact, the model's ability to find these anomalies was 3.02 times better than just guessing.

How Sure Are They?

The researchers are very confident that this method works, but they are careful not to overhype it. They used a statistical test (the Kolmogorov-Smirnov test) which gave a p-value of 1.95x10⁻³⁵. In plain English, this number is so incredibly small that it proves the robot isn't just getting lucky; it is genuinely seeing a difference between normal moon dirt and spots with anomalies.

However, the paper doesn't claim to have found new alien bases or lost ships. It simply proves that this specific type of machine learning model is a powerful tool for scanning the Moon. The robot found the things we already knew were there, which suggests it's ready to go hunting for things we don't know about yet.

What's Next?

The team sees this as just the beginning. Their next goal is to run this robot over the entire Moon to create a global map of weird spots. They hope to find:

Hidden volcanic pits or collapsed lava tubes (great for future astronauts to hide in!).
Fresh craters or rockfalls.
Any missing or unknown pieces of technology left behind.

They also realize the robot has a "bias" because of the patch size it was trained on. If they train it on smaller patches, it might get better at spotting tiny things. If they train it on bigger patches, it might get better at spotting huge craters. They plan to mix different versions of the robot to get the best of both worlds.

Finally, this isn't just for the Moon. The same technique could be used on Mars, Mercury, or any other planet with high-resolution photos. The code is even available for anyone to use, meaning this digital detective is ready to help the whole world search for the strange and the unknown in our solar system.

Technical Summary: A Machine Learning Based Search for Lunar Anomalies

Problem Statement
The Lunar Reconnaissance Orbiter (LRO) has accumulated a vast dataset of high-resolution images (approximately 0.5–2 m/pixel) of the Moon since 2009. While this data offers unprecedented scale for surface study, manually searching for specific features—ranging from natural geological anomalies (e.g., rockfalls, fresh craters, volcanic pits) to artificial technosignatures (landed spacecraft)—is labor-intensive and prone to human bias. The central challenge addressed by this work is the development of an efficient, automated method to identify surface anomalies within this large dataset without relying on pre-labeled training data, which is often difficult and biased to obtain.

Methodology
The authors employ an unsupervised learning approach using a $\beta$ -Variational Autoencoder ( $\beta$ -VAE), a model originally introduced by Lesnikowski et al. (2024). Unlike supervised models that require ground-truth labels, this algorithm learns the latent representation of the "statistically normal" lunar surface by identifying underlying patterns within the training data.

Data Processing: The large LRO images (approx. 50,000 $\times$ 5,000 pixels) are segmented into 64 $\times$ 64 pixel patches (roughly 32 $\times$ 32 meters). This size was selected to balance the detection of large features (craters, caves) with smaller ones (landed assets).
Training: The model was trained on approximately 1,000 randomly selected LRO images, totaling over 52 million patches. The $\beta$ -VAE learns to reconstruct these input patches based on the "average" lunar surface it has learned. The $\beta$ regularization term is critical here; it shifts the objective function to prioritize capturing underlying data features rather than perfect pixel-for-pixel reconstruction, thereby forcing the model to learn the domain's general structure.
Anomaly Scoring: An "anomaly score" is calculated by comparing the reconstructed patch to the original input on a pixel-by-pixel basis. Patches with high scores indicate features that deviate significantly from the learned average lunar surface.

Key Contributions and Results
The study validates the model's ability to detect both natural and artificial anomalies through two primary tests:

Natural Anomalies: The model was tested on two community-discovered sites: Plaskett Crater (a bedrock outcrop acting as a rockfall source) and Paracelsus C Crater (containing angular boulders from a past impact). The model successfully assigned high anomaly scores to patches containing these features, distinguishing them from the homogeneous regolith.
Technosignatures: The model was evaluated against a suite of known landed technological assets, including all Apollo landers (11, 12, 14, 15, 16, 17), the Apollo 13 S-IVB crash site, Surveyor 3, Luna 24, SMART-1, and the Chinese Chang'e 3, 5, and 6 landers.
- Statistical Significance: The distribution of anomaly scores for patches containing known technosignatures was statistically distinct from those without, confirmed by a Kolmogorov-Smirnov test (p-value $1.95 \times 10^{-35}$ ).
- Performance: The model's average precision improved by a factor of 3.02 compared to a random baseline.
- Limitations Observed: While effective, the anomaly scores for artificial objects were generally lower than those for large natural anomalies. The authors attribute this to the patch size bias; the metric favors objects that dominate the patch, whereas small spacecraft occupy a smaller fraction of the 64 $\times$ 64 pixel area.

Significance and Claims
The paper claims that the current version of the $\beta$ -VAE model robustly identifies lunar surface anomalies, demonstrating the viability of re-examining existing LRO datasets with generative deep learning models. The authors assert that this approach offers a new perspective for discovering:

Previously unknown geological structures useful for future missions (e.g., volcanic pits, recent impacts).
Missing or unknown landed assets.

The work positions this technique as a scalable solution that can be extended to a global lunar anomaly map. Furthermore, the authors suggest the methodology is transferable to other Solar System bodies with high-resolution imaging (e.g., Mars, Mercury) via transfer learning. The code for the model is made available to the community on GitHub to facilitate further research. The paper concludes modestly, noting that future iterations could improve performance by training models on varying patch sizes and adjusting the $\beta$ regularization term to better detect anomalies of different scales.