CNN-based Automatic Identification of Rock Particles on Small Bodies

Abstract:

The surfaces of small bodies, such as Itokawa, Ryugu, Bennu, and Churyumov-Gerasimenko, are generally covered with countless rock particles, including boulders, cobbles and pebbles. The total number of rock particles, their positions and shapes, aerial distributions, and size distributions, are critical for understanding the origins and surface processes of small bodies. However, the identifications of such particles are challenging due to the irregularity in the particle shapes, which often overlaps each other with the blurred profiles within the limited image resolutions. Here, we present a computational approach for the automatic identification of particles based on the image feature extraction algorithm utilizing the convolutional neural networks (CNNs). We prepare images of asteroid Itokawa and the laboratory-based simulated surface by using the simulated material (simulant) of the regolith, and carefully identify thousands of particles manually by eyes. With the data of profiles, the model is trained, enabling the particles on asteroid Itokawa to be identified without the aid of manual analysis. The particles identified automatically are mapped on the shape model of Itokawa and their sizes are measured, resulting in the Cumulative Size-Frequency Distribution (CSFD) of particles having the power-law index of -3.48±0.08, which is consistent with the CSFD obtained by the previous research. The approach of this study can rapidly identify numerous particles, which can be a promising tool for analyzing countless images and deciphering the geologic records on small bodies.