I've extended the capabilities of lunar anomaly detection by adding a form of 'Method of Experts'.
As previously described, my lunar anomaly detector uses an Autoencoder to detect small patches on the moon that are 'not the moon' according to the training data. The training data for the autoencoder is over seventy-five hundred 64x64 pixel image 'segments', randomly selected from a ~5km by ~6.5km region on the moon.
The anomaly detector then systematically goes through the same image and gives each 64x64 pixel 'segment' a score based on how much unlike the moon it is (that is, how well the autoencoder is able to re-create the input image). The higher the score, the more unlike the moon.
Because the autoencoder is trained on a random set of data, it seems reasonable to assume that different training sessions will generate different autoencoder models. This is indeed the case.
So now I can see how much different models agree with each other, and generate two new data products:
1. An 'agreement map' that shows the agreement between eight different models. Sofar in my observations, about 70% of the top 256 scores agree across all models.
2. An 'agreement/score map' that shows a superposition of where the models agree and what 'segments' are most anomalous
Here's a typical region (from LRO dataset M1424544254RC):
Here is the 'agreement map' between the 'top 256' most anomolous regions among the eight different models (red indicates the 178 segments where all eight models agreed, grayscale indicates lower levels of agreement):
Here's the 'agreement map' superimposed onto the input LRO image. The grey areas are hard to see, but the red areas show up ok:
And now here is the 'agreement / score map' -- a combination of the agreement and the total score between the eight different models:
The bright areas are where the models agree and the scores are high. This might be useful for detecting regional or large-scale anomalies in addition to specific anomalies.
