[Paper Review] Unsupervised Distribution Learning for Lunar Surface Anomaly Detection

Adam Lesnikowski, Valentin Bickel+1 more|arXiv (Cornell University)|Jan 14, 2020

Space Science and Extraterrestrial Life11 references4 citations

TL;DR

This paper proposes an unsupervised variational autoencoder (VAE) for detecting lunar surface anomalies, such as the Apollo 15 landing module, without prior labeling or hyperparameter tuning. The model learns a data density representation from high-resolution LRO NAC images, achieving an AP of 0.49 and PaTR of 0.055 on an unseen test set, demonstrating effective zero-shot technosignature detection.

ABSTRACT

In this work we show that modern data-driven machine learning techniques can be successfully applied on lunar surface remote sensing data to learn, in an unsupervised way, sufficiently good representations of the data distribution to enable lunar technosignature and anomaly detection. In particular we train an unsupervised distribution learning neural network model to find the Apollo 15 landing module in a testing dataset, with no dataset specific model or hyperparameter tuning. Sufficiently good unsupervised data density estimation has the promise of enabling myriad useful downstream tasks, including locating lunar resources for future space flight and colonization, finding new impact craters or lunar surface reshaping, and algorithmically deciding the importance of unlabeled samples to send back from power- and bandwidth-constrained missions. We show in this work that such unsupervised learning can be successfully done in the lunar remote sensing and space science contexts.

Motivation & Objective

• To develop an unsupervised machine learning method for detecting unknown lunar surface anomalies, including human-made artifacts like landers.
• To evaluate whether explicit data distribution learning can enable detection of technosignatures without prior knowledge of their physical form or shape.
• To validate the method on an unseen test set (Apollo 15) using only a model trained on Apollo 17 data, without any hyperparameter adjustment.
• To demonstrate the feasibility of using unsupervised density estimation for scientific discovery in planetary science, such as locating new impact craters or resources.
• To enable efficient prioritization of unlabeled data for transmission in bandwidth-constrained space missions.

Proposed method

• Trained a deep variational autoencoder (VAE) with eight convolutional and transposed convolutional layers, batch normalization, and a learned latent code of size 2^6 (64) to 2^11 (2048), to learn a compressed, disentangled representation of lunar surface patches.
• Used the reparameterization trick to enable backpropagation through stochastic sampling, and applied beta-VAE regularization to encourage disentangled and meaningful latent features.
• Minimized a loss function combining reconstruction error and KL divergence between the learned latent distribution and a standard normal prior, with beta-weighted KL to improve representation quality.
• Extracted anomaly scores from the VAE by measuring reconstruction error—higher error indicates lower likelihood under the learned data distribution.
• Evaluated performance using precision-recall metrics, including Average Precision (AP), Precision at Top R (PaTR), and Early Out-of-Mean Value (EoMV).
• Applied the model to 64×64 pixel patches from LRO Narrow Angle Camera (NAC) images, using a 32-pixel stride to ensure spatial overlap and coverage of the Apollo 15 and 17 landing sites.

Experimental results

Research questions

• RQ1 Can an unsupervised distribution learning model detect a known lunar lander (Apollo 15) in a test set without any fine-tuning or hyperparameter adjustment?
• RQ2 How well does a VAE-based anomaly detection system generalize to a new, unseen landing site when trained only on a different one (Apollo 17)?
• RQ3 To what extent can unsupervised data density estimation identify scientifically relevant anomalies such as landers or impact craters without prior knowledge of their appearance?
• RQ4 How do reconstruction quality and latent space interpolation reflect the model’s ability to learn meaningful representations of lunar surface data?
• RQ5 Can this approach be used to prioritize data transmission in future space missions with limited bandwidth?

Key findings

• On the Apollo 17 validation set of 2000 patches, the model achieved an Average Precision (AP) of 0.25 and a Precision at Top R (PaTR) of 0.333, with an Early Out-of-Mean Value (EoMV) of 111, indicating strong early detection of positive samples.
• On the unseen Apollo 15 test set of 8000 patches, the model achieved an AP of 0.49 and a PaTR of 0.055, with an EoMV of 45.8, demonstrating effective zero-shot detection of the landing module without any tuning.
• The precision-recall curve showed a high concentration of positive samples in the top-ranked results, with the top three predictions being positive, indicating strong anomaly discrimination.
• Visual inspection of reconstructions and latent space interpolations confirmed that the VAE learned a meaningful, smooth representation of lunar surface textures and features, with stable and plausible generation.
• The model’s performance on the unseen Apollo 15 site, despite no hyperparameter tuning, confirms the robustness of unsupervised distribution learning for planetary-scale anomaly detection.
• The results suggest that unsupervised data density estimation can enable discovery of unknown or non-publicly disclosed lunar landings, impact craters, and resources, even with limited labeled data.