A deep-learning-based method with appearance embedding and multi-view photometric consistency aims to reconstruct highly detailed shape models of asteroids under various environmental illumination conditions.

Study conducted by Prof. Bo WU and his research team

Significant strides have propelled the lunar and planetary sciences forward in recent years, with notable achievements stemming from the burgeoning missions in Moon and Mars exploration conducted by the US, China, and other countries. Asteroid exploration and sample-return from asteroids have also attracted much attention, as asteroids are valuable sources for studying the evolution of the Solar System and may contain clues to help reveal the formation of terrestrial planets and the origins of water and life. The low gravity, non-uniform shape, and rapid rotation of asteroids present challenges for the sampling process from asteroids. To ensure a successful sample collection, it is crucial to conduct a detailed investigation of the asteroid’s surfaces to evaluate potential sampling sites that are both safe and scientifically valuable.

Existing methods, including stereo-photogrammetry (SPG)¹ and stereo-photoclinometry (SPC)² face limitations in the 3D surface reconstruction of asteroids. These methods tend to generate overly smooth surfaces and exhibit varying performance based on surface reflectance and albedo. With the aid of deep learning methods, a research team led by Bo WU, Associate Director of the Research Centre for Deep Space Explorations and Professor in the Department of Land Surveying and Geo-Informatics at the Hong Kong Polytechnic University, has leveraged remote sensing and deep learning for topographic analysis and proposed a promising approach to reconstruct the 3D surfaces of asteroids.

Writing in Astronomy & Astrophysics, the research team leverages the state-of-the-art neural radiance field (NeRF) model and presents Asteroid-NeRF, a novel deep-learning approach for reconstructing 3D surface models of asteroids³. Neural radiance fields (NeRF) is a prominent technique related to neural volume rendering (Figure 1). This method uses neural networks to represent a scene as a continuous function of 3D coordinates and viewing direction, employing colour fields and volume density to build high quality 3D scenes. Traditional volume rendering techniques involve projecting 3D volumetric data onto a 2D image plane. These techniques require sampling the volume at various depths and applying techniques like ray casting or ray marching to compute colour and opacity based on the data values. With the use of underlying volumetric data, neural network models are trained to predict colour and opacity values for points in 3D space using a continuous function. This approach improves efficiency by accelerating the rendering process without the need for extensive sampling data, while producing smooth interpolation in rendering complex volumetric scenes.

Figure 1. Volume rendering sampling of asteroid images

Riding on NeRF, the research team proposes the Asteroid-NeRF approach, which is capable of analysing images of asteroids with significant illumination differences and substantial shadows to build a 3D surface model. It utilises a continuous and global signed distance field (SDF) to encode the distance of a position from the surface in the 3D scene, providing a negative value for any point inside the object and a positive value for any point outside the object, thereby reconstructing the shape and radiance of the target. The approach incorporates novel appearance-embedding strategies to deal with asteroid images captured under varying illumination conditions and employs multi-view photometric consistency to optimise surface reconstruction.

The research team used hundreds of images of two asteroids produced by the Asteroid Multi-band Imaging Camera (AMICA) and NASA’s Origins, Spectral Interpretation, Resources Identification, and Security-Regolith Explorer (OSIRIS-REx) to train and test the models. Compared to benchmarking methods such as SPG and SPC, the Asteroid-NeRF models demonstrate refined detail and accuracy in capturing the overall structure and various topographic features of the asteroids, offering a new and effective solution for high-resolution 3D surface reconstruction of asteroids (Figure 2).
 

Note: Plus and minus symbols indicate different viewing directions

Figure 2. Overview of a 3D surface model of Itokawa (a near-Earth peanut-shaped/sea-otter-shaped asteroid) reconstructed by Asteroid-NeRF. Difference maps of (a) the Asteroid-NeRF model and the SPC model, and (b) the SPG model and the SPC model of Itokawa’s surface. (c) Corresponding images of Itokawa, for reference. One of the most significant differences between the Asteroid-NeRF and SPC models is the omission of a boulder by the latter model.

For years, WU’s research team has developed innovative deep-learning methods for enhanced topographic mapping and geomorphological analysis to support China’s lunar and Mars landing missions. Leveraging deep-learning and neural networks, the cutting-edge Asteroid-NeRF approach assists the research team in reconstructing refined 3D surfaces of asteroids, making valuable contributions to China’s future asteroid exploration endeavour.

The research is supported by grants from the Research Grants Council of the Hong Kong Special Administrative Region, China (RIF Project No.:R5043-19; GRF Project No.: PolyU 15210520; PolyU 15219821). The neural networks and surface models are available at https://zenodo.org/records/11174638.

1. Scholten, F., Preusker, F., Elgner, S., Matz, K.-D., Jaumann, R., Hamm, M., Schröder, S. E., Koncz, A., Schmitz, N., Trauthan, F., Grott, M., Biele, J., Ho, T.-M., Kameda, S., Sugita, S. (2019). The Hayabusa2 lander MASCOT on the surface of asteroid (162173) Ryugu – Stereophotogrammetric analysis of MASCam image data. Astronomy & Astrophysics, 632, L5. https://doi.org/10.1051/0004-6361/201936760

2. Barnouin, O. S., Daly, M. G., Palmer, E. E., Gaskell, R. W., Weirich, J. R., Johnson, C. L., … Turner, R. (2019). Shape of (101955) Bennu indicative of a rubble pile with internal stiffness. Nature Geoscience, 12(4), 247–252. https://doi.org/10.1038/s41561-019-0330-x

3. Chen, S., Wu, B., Li, H., Li, Z., & Liu Y. (2024). Asteroid-NeRF: A deeplearning method for 3D surface reconstruction of asteroids. Astronomy & Astrophysics, 687, A278. https://doi.org/10.1051/0004-6361/202450053

Prof. Bo WU Associate Director of the Research Centre for Deep Space Explorations and Professor in the Department of Land Surveying and Geo Informatics