Transformer-based Approach for Accurate Asteroid Spectra Taxonomy and Albedo Estimation

Abstract China plans to launch a probe (Tianwen-2) around 2025, mainly to explore the near-Earth asteroid 2016 HO3 (469219, Kamo’oalewa). The mission involves close-range exploration, landing, and mining operations that require three-dimensional modeling of the asteroid, which requires prior knowledge of its material composition and uniformity. This information is crucial in progressive or ground exploration processes. Our research focuses on high-precision intelligent inversion of complex physical properties of asteroids based on spectral data, providing support for further analysis of asteroid materials, density, and structure. We have developed a platform for asteroid spectral classification, albedo estimation, and composition analysis, which includes three types of neural networks based on the Transformer attention mechanism: one for spectral classification, achieving a four-class classification accuracy of 94.58% and an 11-class classification accuracy of 95.69%, a second one for albedo estimation, with an average absolute error of 0.0308 in S-type asteroid albedo estimation, and the third one for composition analysis, with a predicted spectral angular distance of only 0.0340 and an rms error of 0.1759 for the abundance of endmembers. These results indicate that our network can provide high-precision asteroid spectral classification, albedo estimation, and composition analysis results. In addition, we utilized the platform to analyze and provide results for six asteroids.