地学前缘 ›› 2019, Vol. 26 ›› Issue (4): 22-32.DOI: 10.13745/j.esf.sf.2019.7.6

• 岩石大数据研究 • 上一篇    下一篇

基于卷积神经网络和火山岩大数据的构造源区判别

葛粲，汪方跃，顾海欧，管怀峰，李修钰，袁峰   

  1. 1. 合肥工业大学 资源与环境工程学院, 安徽 合肥 230009
    2. 合肥工业大学 矿集区立体探测实验室, 安徽 合肥 230009
    3. 合肥工业大学 安徽省矿产资源与矿山环境工程技术研究中心, 安徽 合肥 230009
    4. 安徽省矿产资源储量评审中心, 安徽 合肥 230001
    5. 安徽省地质调查院, 安徽 合肥 230001
  • 收稿日期:2018-01-29 修回日期:2018-07-16 出版日期:2019-07-25 发布日期:2019-07-25
  • 作者简介:葛粲(1986—)，男，博士，副教授，固体地球物理专业，主要从事地壳结构研究和大数据分析。
  • 基金资助:
    国家重点研发计划项目(2016YFC0600209);国家自然科学基金项目(41504042,41702353,41672069,41820104007);中央科研基本业务费资助项目(JZ2019HGTB0071,PA2018GDQT0020)

Tectonic discrimination based on convolution neural network and big data of volcanic rocks

GE Can,WANG Fangyue,GU Hai'ou,GUAN Huaifeng,LI Xiuyu,YUAN Feng   

  1. 1. School of Resources and Environmental Engineering, Hefei University of Technology, Hefei 230009, China
    2. Laboratory of Three-Dimension Exploration for Mineral District, Hefei University of Technology, Hefei 230009, China
    3. Anhui Province Engineering Research Center for Mineral Resources and Mine Environments, Hefei University of Technology, Hefei 230009, China
    4. Anhui Province Evaluation Center for Mineral Resources Reserves, Hefei 230001, China
    5. Geological Survey of Anhui Province, Hefei 230001, China
  • Received:2018-01-29 Revised:2018-07-16 Online:2019-07-25 Published:2019-07-25
  • Supported by:
     

PDF (PC)

443

摘要： 早先的构造源区判别图由于受时代、研究区域、研究思路以及研究手段、分析技术和样本数量的限制, 存在某些不足，导致部分学者在研究中遇到各种困惑和矛盾。在大数据的冲击下，部分传统图解的可靠性正在接受考验。本文提出了一种将地球化学数据二维图像化的方法，将GEOROC数据库中来自11个构造环境的火山岩数据生成了34 468张灰度二维码图像。根据深度学习理论和方法构造了卷积神经网络(CNN)模型，利用其中75%的二维码数据进行自动学习和训练。该模型可以对不同来源的火山岩数据进行有效分类，总体分类准确度可达95%以上。该模型具备较好的泛化能力，可以作为日常工具辅助人工进行火山岩样本的构造源区的判别。

 

关键词: 大数据, 二维码, 卷积神经网络, 构造源区判别

Abstract: The traditional tectonic discrimination graphs have some shortcomings due to the limitation of available analytical methods and techniques of times. This has led to some confusions and contradictions for scholars using the graphs in their researches. Under the impact of big data, the reliability of some traditional tectonic discrimination graphs is being tested. In this paper, we proposed a method for the two-dimensional visualization of geochemical data. Using this method, we converted the geochemical compositions of volcanic rocks from 11 tectonic environments registered in the GEOROC database into 34468 two-dimensional coded images. Relying on deep learning method, 75% of the images were used to learn and train automatically to construct the convolution neural network (CNN) model, which can be employed to classify volcanic rocks into tectonic groups at an overall accuracy of 95%. This model has good generalization capability and can be routinely used to distinguish the tectonic source regions of volcano rock samples.

Key words: big data, two-dimensional code, convolution neural network, tectonic discrimination

中图分类号: