日本地球惑星科学連合2018年大会

講演情報

[JJ] ポスター発表

セッション記号 M (領域外・複数領域) » M-GI 地球科学一般・情報地球科学

[M-GI27] データ駆動地球惑星科学

2018年5月23日(水) 13:45 〜 15:15 ポスター会場 (幕張メッセ国際展示場 7ホール)

コンビーナ:桑谷 立(国立研究開発法人 海洋研究開発機構)、長尾 大道(東京大学地震研究所)、堀 高峰(独立行政法人海洋研究開発機構・地震津波海域観測研究開発センター)

[MGI27-P02] GEOFCM:位置情報を活用した地球化学データのクラスタリング手法の提案

*吉田 健太1桑谷 立1安本 篤史2原口 悟1上木 賢太1岩森 光1 (1.国立研究開発法人海洋研究開発機構、2.京都大学地質学鉱物学教室)

キーワード:多変量解析、ジオコーディング、領家帯

Geochemical data from geological samples show compositional trends that reflect the material differentiation and assimilation occurred during certain geological processes. These trends often comprise groups in a multidimensional compositional space and are distributed in real space as geological units ranging from millimeters to kilometers in scale (e.g., Ueki and Iwamori, 2017). Therefore, spatial contextual information combined with chemical affinities could provide fundamental information about the sources and generation processes associated with the samples.

However, conventional clustering algorithms such as k-means and fuzzy c-means (FCM) cluster analysis do not fully utilize the spatial distribution information of geologic samples. In this study, we propose a new clustering method for geochemical datasets with location coordinates. A spatial FCM algorithm originally constructed for image segmentation was modified to deal with a sparse and unequal-spaced dataset. The proposed algorithm evaluates the membership function modified using a weighting function calculated from neighboring samples within a certain radius.

We applied new algorithm to a geochemical dataset of granitoids in the Ina-Mikawa district of the Ryoke belt that was compiled by Haraguchi et al. (2017), showing that samples collected from the same geological unit are likely to be classified as the same cluster. Moreover, overlapping geochemical trends are classified consistently with spatial distribution, and the result is robust against noise addition compared with standard FCM analysis.

The proposed method can be calculated in the “GEOFCM” Excel® sheet provided as supplementary material and on our website (http://dsap.jamstec.go.jp). Geological datasets with precise location coordinates are becoming increasingly available, and the proposed method can help find overviews of complicated multidimensional data structure.