Automated Crater Morphology Characterization on the Moon Using an Unsupervised Neural Network

Craters provide windows into a planet’s geology, and their morphologies provide important information into the properties of the body that was impacted and the processes that have occurred since the crater formed. While previous lunar crater classification has been done by hand, it can be time consuming given the sheer number of craters that are on the Moon. We propose the use of unsupervised machine learning techniques for a systematic, time efficient approach to crater characterization. Unsupervised machine learning techniques deal with finding patterns in unlabeled data. One of their main purposes is to achieve a form of dimensionality reduction, i.e., condensing data into less features while still retaining the most important aspects. Using the python deep learning packages Keras and Tensorflow, we construct an autoencoder to categorize the morphology of simple craters and explore the natural clustering that results when crater images are put through the autoencoder.

We first create a trial data set composed of 46 images of craters, obtained by the Lunar Reconnaissance Orbiter Camera Wide Angle Camera (LROC WAC). The images are then put through an autoencoder neural network where they are first encoded, or compressed, into a set of features called the ‘bottleneck’ layer or vector. This layer is subsequently used to decode or reconstruct the original input. Following this, we plot the locations of each crater within the outputted latent space of the bottleneck layer of the autoencoder, therefore theoretically plotting images with similar features close to one another. We find our techniques can distinguish pristine craters from degraded craters and those with rims breached by the lunar mare basalts.

For our second trial, we compile remote sensing data (including LROC Narrow Angle Camera [NAC] mosaics, Mini-RF radar, and normalized nighttime temperatures) for 100 craters. 50 of these craters are cold spot craters, which are thermophysically distinct, and the other 50 are non-cold spot craters of similar size. Similarly, these will be put through the autoencoder to evaluate the viability of using an autoencoder for automated detection of morphologies that distinguish cold spot craters.