Automated crater detection using machine learning

Craters are among the most studied geomorphic features in the Solar System because they yield important information about the past and present geological processes and provide information about the relative ages of observed geologic formations. This work focuses on improving the features extracted from satellite imagery in order to more accurately detect craters. The first focus is on improving the accuracy of methods based on semi-automatic Haar image features by only considering subsets of the features. Using feature selection methods for black-box optimization such as genetic algorithms and randomized variable elimination we are able to achieve better performance. The second focus was to learn the optimal filters and features based on training examples and replace the semi-automatic Haar features with full-automatic convolutional filters. For this a Convolutional Neural Network (CNN) called CraterCNN is designed which outperforms all existing methods and achieves up to 90% on the standard crater benchmark dataset. Then the GoogLeNet inception architecture is used to further improve the benchmark and achieve up to 93% F1-Score. In order to decrease the computational cost of CNN models to make global Martian analysis possible a convolutional feature selection method called RandomOut is proposed. This method identifies convolutional filters which have been abandoned by the network by using the convolutional gradient norm and reinitializes them during training. RandomOut method enables CNNs to increase their accuracy to that of a network containing more filters but without the computational cost of actually adding more filters. This dissertation showcases significant progress in the field of automated crater detection and provides methods that can be applied to many other areas of automated planetary science.