09:45 〜 10:00
[PPS03-04] Automatic detection and measurement of boulders size and shape on Ryugu using deep learning
キーワード：リュウグウ、深層学習
Size and shape distributions of boulders on asteroids give constraints on conditions of their origins and growths, which are essential for evaluating the collisional history of asteroids. In laboratory experiments, the axial ratios of fragments from catastrophic disruptions were found to be 2: √2: 1[1]. Therefore, measuring boulders can support to decide whether asteroids experienced catastrophic disruptions or not.
In research of asteroid Ryugu, because Ryugu is thought to be formed by re-accumulation after the catastrophic disruption of its parent body, boulder measurement may reveal the formation process of Ryugu. Michikami et al. [2] investigated the boulder size distributions and three-axial ratios on Ryugu by using images taken by Hayabusa2’s telescopic camera (Optical Navigation Camera telescopic, ONC-T [3]). They counted about 16,000 boulders (5 cm to 7 m in diameter), and concluded that the boulders were formed mainly by the catastrophic disruption of its parent body. They also found that the boulder size distributions showed a slightly different slope (power-index) in each image, suggesting different boulders inflow and formation processes in different locations.
The total number of close-up images by a visible band of ONC-T is ~300 that covers various locations of Ryugu. While analyzing surface images of the entire Ryugu and examining regional differences should be useful to reveal the formation process, Michikami et al. [2] counted boulders in just 14 images, because of a time-consuming task of counting boulders manually. Therefore, in this study, we adopt a deep learning method, the Instance Segmentation, that can not only detect boulders in an image, but also measure their shapes and sizes at the same time automatically. We adopt Mask R-CNN [4] for the boulder detection and size measurement, which is one of the practical deep learning methods for the Instance Segmentation task. We modified the model suitable for boulder detection. In addition, to enable the training, we have prepared annotation dataset in which numerous boulders are extracted manually. The total number of images was 275, and the total number of boulders was about 13,000 in this study. The goal of this study is to obtain the size distributions and three-axial ratios of the boulders and examine how regional differences emerged during the formation process of Ryugu. Also, the contribution of this study is to provide our annotation dataset to the community for enabling further deep learning studies widely.
We experimentally trained the Mask R-CNN model with the annotated boulders in 100 images. The results of inference are shown in the figure below. Then, we confirmed that our model can detect individual boulders automatically and enables to measure their two-axial lengths by conducting elliptical fitting. The size distributions from the inference results with 25 images showed power-law distributions as in [2] and their indices were basically similar to [2], while the locations of the inference images in this study are different from images used in [2].
Although not all boulders, especially for apparently small boulders with less than few ten pixels, have been detected yet, the resulted size distributions were consistent with those in previous studies.
We will analyze more images taken by ONC-T to investigate regional differences. We will also improve the model structure of the Mask R-CNN to detect smaller boulders, and for all images taken by the ONC-T, the size distribution and three-axial lengths of boulders will be automated.
References: [1] Fujiwara et al. Nature 272.5654 (1978): 602-603. [2] Michikami et al. Icarus 331 (2019): 179-191. [3] Kameda et al. Space Sci. Rev., 208 (2017): 17-31. [4] He, Kaiming, et al. Proceedings of the IEEE international conference on computer vision. 2017.
In research of asteroid Ryugu, because Ryugu is thought to be formed by re-accumulation after the catastrophic disruption of its parent body, boulder measurement may reveal the formation process of Ryugu. Michikami et al. [2] investigated the boulder size distributions and three-axial ratios on Ryugu by using images taken by Hayabusa2’s telescopic camera (Optical Navigation Camera telescopic, ONC-T [3]). They counted about 16,000 boulders (5 cm to 7 m in diameter), and concluded that the boulders were formed mainly by the catastrophic disruption of its parent body. They also found that the boulder size distributions showed a slightly different slope (power-index) in each image, suggesting different boulders inflow and formation processes in different locations.
The total number of close-up images by a visible band of ONC-T is ~300 that covers various locations of Ryugu. While analyzing surface images of the entire Ryugu and examining regional differences should be useful to reveal the formation process, Michikami et al. [2] counted boulders in just 14 images, because of a time-consuming task of counting boulders manually. Therefore, in this study, we adopt a deep learning method, the Instance Segmentation, that can not only detect boulders in an image, but also measure their shapes and sizes at the same time automatically. We adopt Mask R-CNN [4] for the boulder detection and size measurement, which is one of the practical deep learning methods for the Instance Segmentation task. We modified the model suitable for boulder detection. In addition, to enable the training, we have prepared annotation dataset in which numerous boulders are extracted manually. The total number of images was 275, and the total number of boulders was about 13,000 in this study. The goal of this study is to obtain the size distributions and three-axial ratios of the boulders and examine how regional differences emerged during the formation process of Ryugu. Also, the contribution of this study is to provide our annotation dataset to the community for enabling further deep learning studies widely.
We experimentally trained the Mask R-CNN model with the annotated boulders in 100 images. The results of inference are shown in the figure below. Then, we confirmed that our model can detect individual boulders automatically and enables to measure their two-axial lengths by conducting elliptical fitting. The size distributions from the inference results with 25 images showed power-law distributions as in [2] and their indices were basically similar to [2], while the locations of the inference images in this study are different from images used in [2].
Although not all boulders, especially for apparently small boulders with less than few ten pixels, have been detected yet, the resulted size distributions were consistent with those in previous studies.
We will analyze more images taken by ONC-T to investigate regional differences. We will also improve the model structure of the Mask R-CNN to detect smaller boulders, and for all images taken by the ONC-T, the size distribution and three-axial lengths of boulders will be automated.
References: [1] Fujiwara et al. Nature 272.5654 (1978): 602-603. [2] Michikami et al. Icarus 331 (2019): 179-191. [3] Kameda et al. Space Sci. Rev., 208 (2017): 17-31. [4] He, Kaiming, et al. Proceedings of the IEEE international conference on computer vision. 2017.