[M-IS21] 惑星火山学

[MIS21-P08] 教師あり機械学習手法を用いた可視画像からの露頭領域抽出

*野口 里奈¹、庄司 大悟²、藤本 圭一郎²、下司 信夫³、白尾 元理、春山 純一² (1.新潟大学 自然科学系、2.宇宙航空研究開発機構、3.産業技術総合研究所)

キーワード：露頭、機械学習、地質調査、火星

The stratigraphic column is one of the description methods which summarize fundamental geological data such as layer thickness and constituent materials at a certain outcrop. Identification and discrimination of stratigraphic exposures are preliminary and fundamental work in the geological survey. Automation of such basic investigations is not as easy as it seemed at first sight because they are often obscured by talus and vegetation. In this study, we performed supervised machine learning to extract areas of stratigraphic exposures in visible images using u-net. We input augmented 14024 terrestrial outcrop images and those masked images to train the machine. As a result, we obtained 92% of accuracy for validation data. The trained model can extract stratigraphic exposures from inputted images, though there are some difficulties in color and untrained situations such as snow coverage. This autonomous detection of exposed stratigraphic structures will increase output from the huge storage of high-resolution images taken on terrestrial bodies. Autonomous detection of exposed stratigraphic structures from outcrop images will contribute to the remote-geological survey on the red planet.