Autonomous crater detection on asteroids using a fully-convolutional neural network

This paper shows the application of autonomous Crater Detection using a Fully-Convolutional Neural Network (FCN) on a main belt asteroid, Ceres. The FCN, namely a U-Net, is trained on optical images of the Moon Global Morphology Mosaic based on data collected by the Lunar Recoinnassance Orbiter (LRO), using manual crater catalogues as annotations for the creation of ground truth data. The Moon-trained network will then be tested on optical images of Ceres taken during the Dawn mission; as this network is adapted to a different domain from the one upon which it was already trained, this task is accomplished by means of a Transfer Learning (TL) approach. In particular, this work will take into consideration a homogeneous TL, because the image features and labels share the same qualities in both domains. The Moon-trained model has been fine-tuned using 100, 500 and 1000 additional annotated images of Ceres, reaching a testing accuracy on 350 never before seen images of 96.24%, 96.95% and 97.19%, respectively. Annotations for Ceres are taken from the Zeilnhofer crater catalogue, containing 44.594 craters with diameter greater than 1 km in a latitude range of 84.66°S-89.62°N, full longitude: therefore, a near-global coverage of the asteroid surface will be taken into consideration. The output of the U-Net is a grey-scale mask containing predicted craters: it will be post-processed applying global thresholding for image binarization and a template matching algorithm to extract craters positions and radii in the pixel space: this allows the creation of a definitive binary mask. These post-processed craters will be counted and compared to the ground truth data in order to compute image segmentation metrics: precision, recall and their harmonic mean known as F1 score. Precision, which is the ratio between matched craters and all the craters found by the network, reached values of 70.51%, 84.43% and 83.45%. Recall, defined as the ratio between matched and catalogued craters, achieved performances of 48.27%, 53.32% and 58.97%; the lower values are due to the fact that when precision is high, recall is automatically penalized and vice versa. The F1 score is introduced to sum up the effects of both the above indices: its reached percentages are 57.30%, 65.36% and 69.11%. These results are considered to be encouraging because they are comparable with other works that use the U-Net in a Crater Detection context