Good knowledge of Lunar surface morphology is essential for lunar exploration and scientific research. In particular, the production of lunar surface Digital Terrain Models (DTMs) of high quality has been a long-standing challenge in planetary mapping. Near-global coverage by highest-resolution images (up to 0.5 m / pixel) have been obtained by the Narrow Angle Camera (NAC) on board Lunar Reconnaissance Orbiter (LRO). However, the number of stereo observations, which are traditionally used for DTM production with photogrammetry technique, is limited. Hence, we must resort to single-view images for terrain reconstruction. Just recently, a new image inversion technique, based on Convolutional Neural Networks (CNN), has been successfully introduced for retrieval of planetary DTMs from single-view orbiter images.

In this project, we have designed a novel CNN architecture and have developed a software pipeline to generate large-area high-resolution DTMs from multiple adjacent single-view NAC images. We use coarse-resolution Selenological and Engineering Explorer and LRO Elevation Models (SLDEM) for training of the deep learning algorithm.

Our technique has been applied for example to the Chang’E-4 landing area.  We will focus on the following issues in the future:

1. The high-precision co-registration of images taken in different resolution and under different illumination conditions
2. Robust software pipeline for consistent high-quality DTMs in spite of terrain diversity.
3. Automated data management and large-area mapping.