Astronomy & Music: Lunar Anomalies #1

Are there objects or patterns on the Moon that don't look like they should be there, or have some other unusual characteristics? If there are, what are they?

I've started a project that is searching for anomalies on the Moon.

An ‘anomaly’ is something that isn’t ‘like the Moon’.

Here’s a description of the process I've developed:

I start with a CDR (Calibrated Data Records) image from LRO (Lunar Reconnaissance Orbiter) downloaded from https://pds.lroc.im-ldi.com/data/. I’m using the ones that are 5064x52224 pixels. I divide this images up into eight, 5064x6528 sections and call them a, b, c, d, e, f, g, and h.

To generate training data for an autoencoder, I take a section and I randomly select 7,575 64x64 subarrays. I train on these and generate a model of ‘the typical moon’ for that particular image section.

Here’s a typical image section showing the location of 200 (of the 7575) random sample locations:

It looks like it’s a pretty decent random distribution.

I then divide the image up into 7,575 64x64 images systematically (i.e., in a grid that samples the entire image). Each of these is fed into the autoencoder model, and each gets a ‘score’ based on the MSE (mean-squared error) of the output of the model.

I then sort these according to MSE value. The lower the MSE, the more similar to the trained model it is (and “more like” the Moon it is). The higher the MSE, the less similar to the model it is (and more unlike the Moon it is).

Judging from the sorted MSE curves, I’ve decided that the top 225 sections (i.e., the sections with the highest MSE values) are the ones I should pay most attention to when it comes to finding anomalies.

In the end of this process, I generate a sorted (highest to lowest) MSE list of subarrays. Note these plots are log scale on the y-axis:

Here are the top 225:

I take the top 225 of them and generate a visual display of each subarray (upper left has the highest MSE, lower ones proceed left to right to the bottom right) :

Here's where these 225 subarrays are located in the full image section:

As expected, many of these mark the locations of craters and/or rock outcroppings, which the autoencoder has deemed to be "not like the moon" (a high MSE):

Sofar in the very small amount of data I've examined (there are literally petabytes of LRO data, so I'll never get through it all), I've already found a few interesting things: