A Machine Learning Based Search for Lunar Anomalies

The Lunar Reconnaissance Orbiter (LRO) has been collecting high-resolution images (at ∼ 0.5-2 m/pixel linearly with its Narrow Angle Camera) of the Moon since 2009, amassing a large dataset of images and offering researchers the opportunity to study the surface of the Moon at unprecedented scale. Here, we aim to test the abilities of the β-Variational Autoencoder (VAE) created by Lesnikowski et al. (2024), an unsupervised learning model which identifies anomalous features across the Moon’s surface, locating not only scientifically useful geologic formations such as rockfall deposits, fresh impact craters, irregular mare patches, or volcanic pits/collapsed lava tubes, but also artificial objects such as landed spacecraft. This investigation further gauged the model’s ability to locate anomalous surface features, successfully recovering two places of interest (Plaskett Crater and Paracelsus C Crater) and numerous landed technological assets at a statistically significant rate.

The dataset used in this project contains images taken by the Lunar Reconnaissance Orbiter (LRO) over the last 17 years. The LRO was launched by NASA on June 18, 2009 with the goal of producing detailed mapping and remote sensing of the Moon’s surface. Looking up at the Moon, even with the naked eye, its diverse surface of large maria, highlands, and craters can be observed. The LRO offers a much closer perspective, allowing for the observation of finer details; this includes the likes of interesting surface anomalies such as irregular mare patches (IMPs), fresh impact craters, rockfalls, volcanic pits/collapsed lava tubes, other geological formations, and even landed technological assets.

In this work we examined the capabilities of the model introduced by Lesnikowski et al. (2024) which highlighted the viability of new science waiting to be discovered by re-examining existing datasets through the lens of cutting edge technology and advancements such as generative machine and deep learning models. In general, there are 2 major types of machine learning models: supervised and unsupervised. Supervised learning models are given both the raw data and the true corresponding classification or label for the data (e.g. Pinault et al. 2026). The model in this project avoids the use of labeled data, which often relies on a laborious, and biased, process to obtain, not by solely comparing pixel color values as done in standard computer vision algorithms (e.g. De la Torre 2022), but by employing the use of unsupervised learning where the algorithm learns a domain without explicitly being told what it is processing. It does this by uncovering underlying patterns and connections within the training set. Due to the Moon’s overarching homogeneity, any kind of surface anomaly stands out when compared to what the ‘average’ lunar surface looks like, allowing the algorithm to quantitatively rate surface features by comparing their appearance to the mathematically derived ‘average lunar surface.’

We employ a β-Variational Autoencoder (Higgins et al. 2017) trained on a dataset of LRO images (more detailed information can be found in Lesnikowski et al. 2024), learning the latent (underlying) representation of the statistically normal lunar surface. The LRO images are very large (∼50,000 x 5,000 pixels; each pixel is nominally 0.5m x 0.5m), so the data fed to the model is split into patches of a specified size for computational efficiency. This model’s patch size was chosen to be 64x64 pixels (∼32x32 m) to strike a balance between capturing large anomalies (caves, craters, etc.) and small anomalies (landed technological assets). The training dataset consisted of ∼1,000 images randomly selected across the Moon, totaling over 52 million patches to train on. The model ingested these patches, learning underlying patterns and connections between them, and then attempted to recreate them one at a time based on what it had learned - what it expects the lunar surface to look like. The model does not create a high-definition, perfect recreation of the input image. This behavior comes from the ‘β ’ regularization term, which alters the objective function to put more emphasis on capturing the underlying features of the data, rather than focusing on exact reconstructions.

By comparing the reconstructed image patch to the original image patch on a pixel-by-pixel level, the amount in which they differ is quantified as an “anomaly score.” Patches with high anomaly scores exhibit anomalous features, such as volcanic pits/collapsed lava tubes, while less anomalous patches exhibit fairly homogeneous regolith. While there is an inherent bias towards anomalies that take up the majority of a patch, the metric still manages to identify smaller anomalies.

To further test the performance of this model, we examined two community-discovered anomalies-of-interest to determine if the model would assign them significant anomaly scores. The chosen sites were Plaskett Crater and Paracelcus C Crater. The model was given LRO images that contained these anomalies and calculated the anomaly scores for each 64x64 pixel patch of the images. The model successfully identified patches containing these two targets of interest as anomalous, displaying its ability to highlight unique surface features, such as these two rock formations. We note that the Plaskett anomaly is a bedrock outcrop on a slope that has been acting as a source of rockfalls (see boulder track leading towards the left); the Paracelsus C anomaly includes two angular boulders that are most likely remnants of a past impact event.

Also, to expand upon the tests done in Lesnikowski et al. (2024), we fed an expanded suite of known landed technological assets into the model, including: Apollo 11 lander, Apollo 12 lander, Apollo 13 S-IVB crash site, Apollo 14 lander, Apollo 15 lander, Apollo 16 lander, Apollo 17 lander, Surveyor 3, Luna 24 lander, SMART1 crash site, Chang’e 3 lander, Change’e 5 lander, and Chang’e 6 lander. The results from the model examining images containing these technosignatures show that despite their small size within the image patches, the model still identified the majority of patches containing technosignatures as anomalous when compared to the average lunar surface. While the anomaly scores of the artificial anomalies are not as high as natural anomalies, this can be partially explained by the size of the training patches, as the anomaly score calculation has a bias towards objects that dominate the patches.

This investigation showed in a robust manner that the current version of the model can identify lunar surface anomalies well. Applying the model to a larger, if not global, dataset would unveil a plethora of anomalies. One of our team’s next goals is to create a global surface anomaly map of the Moon in hopes of not only discovering previously unknown structures that could prove useful or interesting to future manned missions (volcanic pits, rockfalls, recent impact craters, etc.), but also missing or unknown landed assets. Newly identified anomalous patches could then even be used in a self-supervised manner to improve the model’s performance.

This version of the model was trained based on certain hyperparameters that bias its performance, none more so than the patch size. By training more models on different sized patches, anomalies of different scales can be more readily identified. Combining these models would allow a more comprehensive examination of the lunar surface. The regularization term (β) could also be altered to make the recreated patch images less or more abstract. Testing different values of β across different patch sizes would be another way to test performance across different biases.

Beyond completing a global lunar anomaly search, this technique can be applied to any Solar System body that has high-resolution imaging of its surface, such as Mars or, in the near future, Mercury. The model can even be utilized as a launching point to train new models on new data of a different object using a technique called transfer learning, where a learned model in a certain domain is used as a starting point to train a new model in a similar domain. The code is available on GitHub for the use of the community.

Improvements to AI Systems:

1. Multi-Scale Anomaly Detection Ensemble

• Train multiple β-VAEs with varying patch sizes (e.g., 32×32, 64×64, 128×128, 256×256 pixels) and combine their anomaly scores via a weighted voting or attention mechanism.

• Capability: Detect anomalies across scales—from small artificial objects (landers, 5 m) to large geological features (craters, lava tubes, >100 m)—without size bias, producing a unified global anomaly map.

2. Adaptive β Regularization Scheduler

• Dynamically adjust β during training based on reconstruction loss distribution (e.g., increase β when the model overfits to common textures, decrease when it misses rare anomalies).

• Capability: Balance between capturing fine-grained surface details and abstracting global patterns, improving recall of subtle anomalies (e.g., fresh rockfalls, faint boulder tracks) without sacrificing precision.

3. Self-Supervised Anomaly Feedback Loop

• Use high-scoring anomaly patches (e.g., top 0.1%) as new training data in a self-supervised manner, re-weighting the loss to emphasize these regions in subsequent training iterations.

• Capability: The model continuously improves its ability to recognize novel anomalies over time, reducing false negatives and discovering new classes of features (e.g., buried impact ejecta, transient surface changes).

4. Cross-Domain Transfer Learning with Feature Alignment

• Pre-train the β-VAE on lunar imagery, then fine-tune on Mars, Mercury, or asteroid surfaces using a domain-adversarial network to align latent feature distributions.

• Capability: Rapid deployment of anomaly detection to other planetary bodies with limited labeled data, identifying caves, ice deposits, or landing hazards in new environments.

5. Anomaly Score Calibration via Contrastive Learning

• Add a contrastive loss term that pulls together latent representations of similar anomalies (e.g., all craters) and pushes apart distinct ones (craters vs. landers), using a small labeled set of known anomalies.

• Capability: Produce semantically meaningful anomaly scores that separate natural vs. artificial objects, enabling automated classification of technosignatures (e.g., distinguishing spacecraft from rockfalls).

6. Spatial Context-Aware Patch Aggregation

• Instead of scoring each 64×64 patch independently, use a sliding window with overlap and a graph neural network to propagate anomaly scores across neighboring patches.

• Capability: Reduce false positives at patch boundaries and detect anomalies that span multiple patches (e.g., a long boulder track or a collapsed lava tube), improving spatial coherence of the global anomaly map.

7. Real-Time Anomaly Prioritization for Mission Planning

• Integrate the model with a decision tree that ranks anomalies by scientific value (e.g., proximity to known IMPs, size, freshness) and accessibility (slope, illumination).

• Capability: Automatically generate a prioritized list of targets for future robotic or crewed missions, highlighting safe landing zones and high-interest geological sites.

8. Uncertainty-Aware Anomaly Scoring

• Add a Bayesian layer to the VAE’s decoder to output uncertainty estimates alongside reconstruction errors.

• Capability: Flag patches with high uncertainty as “needs human review,” reducing false alarms in noisy or poorly lit images and improving trust in autonomous analysis.

What the Improved AI System Can Do:

• Generate a global, multi-resolution anomaly map of the Moon (and later Mars/Mercury) with clear separation of natural vs. artificial features.

• Automatically discover new geological formations (e.g., hidden lava tubes, recent impact craters) and locate missing or unknown landed spacecraft with high confidence.

• Continuously self-improve by incorporating newly found anomalies into its training set, adapting to new data without manual labeling.

• Provide mission-ready outputs: prioritized landing sites, hazard maps, and scientific targets, all in real-time from orbital imagery.

The Lunar Reconnaissance Orbiter (LRO) has been collecting high-resolution images (at around 0.5-2 meters per pixel linearly with its Narrow Angle Camera) of the Moon since 2009, amassing a large dataset of images and offering researchers the opportunity to study the surface of the Moon at unprecedented scale. Here, we aim to test the abilities of the Beta-Variational Autoencoder (VAE) created by Lesnikowski et al. (2024), an unsupervised learning model which identifies anomalous features across the Moon's surface, locating not only scientifically useful geologic formations such as rockfall deposits, fresh impact craters, irregular mare patches, or volcanic pits/collapsed lava tubes, but also artificial objects such as landed spacecraft. This investigation further gauged the model's ability to locate anomalous surface features, successfully recovering two places of interest (Plaskett Crater and Paracelsus C Crater) and numerous landed technological assets at a statistically significant rate.