A deep learning approach to crater detection

Kurzfassung

Detecting craters can be useful for a broad variety of tasks, such as monitoring the development of celestial bodies over time or for navigational purposes. Deep learning as a tool for solving complex tasks in image processing became increasingly popular in the recent years. While there are a few works targeting crater detections with deep learning, most focus on monitoring celestial bodies, for which the requirements are different than for the purpose of navigation, where less computing power is available and additional crater parameters must be recovered from the image. In this work, a deep learning approach to crater detection is developed, implemented and evaluated. Due to the absence of proper training data a way to generate the training data with the Planet and Asteroid Natural Scene Generation Utility (PANGU) simulator is implemented. With this simulator a data set consisting of 20000 images and crater labels was generated. A segmentation approach is then used to create probability distributions for crater rims and centers which is then post-processed with conventional methods by using domain knowledge. The researched approaches are evaluated in terms of precision, recall and general accuracy. The different crater parameters are evaluated by categorizing the detections into true positives, false positives and false negatives which are then analyzed separately. Furthermore, the execution time of both approaches is compared. Lastly, the approach is evaluated by applying it to images taken on recent missions to the moon and other celestial bodies.