Research update: Xenarch Model Version 2
The growing interest in the search for extraterrestrial intelligence in scientific circles makes worthy the consideration of searching for technological structures of non-terrestrial origin (i.e., non-terrestrial technosignatures) on planetary surfaces. Xenarch is an ML-based technosignature detection model for planetary surfaces. It is based on a Variational Autoencoder (VAE) that is trained on only natural geology using the metrics reconstruction error, contextual analysis, gradient anomaly, latent density and edge regularity. If it encounters something it cannot reconstruct as natural geology, it will flag it as an anomaly for human review as a potential technosignature.
I applied the Xenarch model to the Apollo 11 landing site based on Lunar Reconnaissance Orbiter Camera (LROC) images. Using the LROC image data, the earliest versions of the model (collectively Xenarch Version 0) were able to identify the Apollo 11 lander as one of the 10 top anomalies but not as the number one anomaly. Instead, the earliest versions of the model identified a field of boulders in a nearby crater as the number one anomaly.
Xenarch Version 1, which reduced the weight of the edge regularity metric and made it more balanced with the other metrics, was more accurate. Version 1 identified the Apollo 11 lander as one of the top 3 anomalies, but still identified a boulder field as the top anomaly. Xenarch Version 2, using higher resolution imagery from the Chandrayaan 2 Orbital High-Resolution Camera (OHRC), was able to identify the lunar lander as the number 1 anomaly with a 99.58% confidence rate.
The model is still a work progress, but the identification of a known artificial structure as the number one anomaly in the test dataset is a big step for this model. Once fully developed, Xenarch could also be applied to search for other types of technosignatures, including atmospheric technosignatures on exoplanets or objects passing through the solar system as interstellar interlopers.
