Automated detection of lunar rockfalls using a Faster Region-based Convolutional Neural Network

Manual detection of small surface features such as lunar rockfalls in high resolution imagery is a difficult, ineffective, and time-consuming process. These rockfalls however are valuable tools to investigate high priority science questions such as the Moon's past and current endo- and exogenic activity as well as its evolution. Other applications include remote evaluation of geotechnical properties of lunar terranes for future human and robotic exploration as well as in situ resource utilization. However, these applications are currently limited by the small total number of identified rockfalls on the lunar surface.

This study implemented a novel method to automatically detect and classify rockfalls in Lunar Reconnaissance Orbiter Narrow Angle Camera (NAC) images using a Region-based Convolutional Neural Network (R-CNN). The network has been trained with a data set of 2932 original rockfall images. In order to avoid overfitting, the initial training data have been augmented during training using random image rotation, scaling, and flipping. Testing images were labelled by human operators during a manual rockfall mapping campaign and have been used for R-CNN performance evaluation. Testing showed that the R-CNN is capable to correctly identify and classify more than 81% (122/151) of the human-based detections. In addition, the network correctly identified more than 227 rockfalls in the used NAC images that have not been mapped by the human mapping campaign. Less than 5% of all classifications are false positives (18/378). Overall network performance is dependent on the used confidence threshold. Mean processing time of a single NAC image in the R-CNN is around 10 seconds using computers optimized for GPU-processing, which is orders of magnitudes faster than a human operator. The processing speed allows to efficiently exploit the currently available NAC data stack with more than 1 million images in a reasonable timeframe. The combination of speed and detection performance can be utilized to produce large scale or global lunar rockfall distribution maps for utilization by the scientific and engineering community.

The workflow is entirely based on open source software. The developed tool will potentially be accessible online on the Moon Trek platform that is being developed by NASA JPL at Caltech, USA.