Research update: Xenarch ML anomaly detection model demo
Machine learning has many applications in planetary science for identifying unusual or hard to detect patterns on planetary surfaces and in planetary datasets. A potential application for machine learning is in planetary SETI for detecting anomalous features that could be of non-terrestrial technological origin. Such machine learning algorithms have been developed and tested for use on the Moon. A general anomaly detection pipeline for any planetary surface has not yet been developed.
For this project, which I have called the Xeno-Archaeology or “Xenarch” Project, I have developed a demo ML model for detecting anomalous features on planetary surfaces that the model cannot reconstruct as natural geologic features. The goal is to create a model that can be generalized to other planetary bodies. The model is not complete, but I have created a demo which I applied to the Apollo 11 landing site.
The demo uses variational autoencoder (VAE) model trained only on natural geology. The training dataset includes images of the Moon (LROC), Mars (HiRISE) and desert Earth desert imagery (Google Earth Pro) that has been vetted to avoid technological or biological structures to avoid unintentionally training the model on features that it is supposed to flag as anomalies.
Once the model was trained, I applied the model to test images of the Apollo 11 landing site. The model successfully identified parts of the Apollo 11 site as anomalies but not top anomalies, indicating more training is needed to refine the model. Nonetheless, this demo model has been able to identify artificial structures on the moon as anomalies that cannot be reconstructed as natural geology and represents the beginnings of a potential model that could be used for planetary SETI as well as for general planetary exploration in identification of unusual planetary phenomena of importance to planetary science researchers and future space resource companies.
