Segmentation Convolutional Neural Networks for Automatic Crater Detection on Mars

Abstract

Machine learning segmentation techniques show great promise for automating historically tedious tasks for planetary scientists. One such task is crater counting, which is commonly used by the planetary science community to study the absolute and relative ages of planetary bodies. Developing effective segmentation neural networks for tasks such as crater counting involves multiple design choices in the network architecture and training set preparation. Here, the authors evaluate two target types, measure the impact of hyperparameters (kernel size, filters), and vary the amount of data used to train the models from using 3 to 15 of the 24 tiles. (Each tile is 30° by 30° and is within ±30° latitude.) The algorithm is trained using annotations of 2- to 32-km-radius Martian craters and THEMIS Daytime IR images. Pixel-based machine learning metrics like loss and accuracy are used during training and validation. In addition, crater count metrics such as the recall (the match ratio), the precision, and the F1 score are used to evaluate the performance and for model selection. The results enumerate how incorporating machine learning into the crater counting process is beneficial to planetary geologists, for example, by creating a list of craters in a region or suggesting potential degraded craters for further analysis. A segmentation network using convolutional neural networks is successfully implemented to find 65%–76% of craters in common with a human annotated dataset.