日本地球惑星科学連合2019年大会

講演情報

[J] 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-GI33] データ駆動地球惑星科学

2019年5月27日(月) 13:45 〜 15:15 ポスター会場 (幕張メッセ国際展示場 8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)、上木 賢太(国立研究開発法人海洋研究開発機構)、加納 将行(東北大学理学研究科)

[MGI33-P03] 人工データを用いた機械学習による地球化学的データ処理および鉱山地域での重金属バックグラウンドを考慮した河川環境評価手法の提案

*佐藤 憲太1澤山 兼吾1土屋 範芳1 (1.東北大学大学院環境科学研究科)

キーワード:環境評価、河川底堆積物、重金属類

Acid mining drainage is one of the most important sources of heavy metals in river. Environmental standards are evaluated by the effects on the Ecosystem and human body. However, the current environmental standards do not consider natural background of heavy metals, therefore the method which can evaluate the effects on the environment concerning such heavy metals is needed. Then we propose to apply machine learning to evaluation effects by heavy metals. Using natural soil data and mineral data as training data in machine learning, it will become possible to evaluate the effects concerning heavy metals naturally included in soil. However, it is very important to choose the proper training data set in machine learning. In this study, we created training data artificially. If it can be done that Verifying correctness of this artificial training data, the range of application of machine learning expands. The objective of this study is, firstly, to verify the validity of artificial training data through case study, secondly, to establish the method to evaluate effects on the environment by heavy metals considering natural background of heavy metals. Focusing on the mine area where there are effects by heavy metals has already been confirmed, we digitized the effects by mine as the evaluation score by heavy metals using Support Vector Regression (SVR) and using the element content ratio as attribute. The element content ratio of mine was used as the ore training data and the element content ratio of the area surrounding the mine was used as the training data representing heavy metals naturally included in soil. We created data of the ore training data based on literature data about the mine. Furthermore, we verified validity of the artificial data through comparing three digitized assessment of contamination by the statistical method in the area, by the previous research, and by machine learning. As a result, the evaluation score by each method increased at the same point, and it is verified that the evaluation by machine learning using artificial training data is correct. If we use machine learning using artificial training data, it may improve further development of machine learning in the geochemical data set.