SEISMOLOGY AND GEOLOGY ›› 2024, Vol. 46 ›› Issue (2): 277-296.DOI: 10.3969/j.issn.0253-4967.2024.02.003

• Review • Previous Articles     Next Articles

APPLICATION OF DEEP LEARNING IN ACTIVE TECTONICS AND GEOMORPHOLOGY

LIU Xin1,2)(), WANG Shi-rou1)(), SHI Xu-hua1,2),*(), SU Cheng1),*(), LU Chen-yan1), QIAN Xiao-yuan1), SUN Qiao-yang1,2), DENG Hong-dan1,2), YANG Rong1,2), CHENG Xiao-gan1,2)   

  1. 1) Key Laboratory of Geoscience Big Data and Deep Resource of Zhejiang Province, School of Earth Sciences, Zhejiang University, Hangzhou 310058, China
    2) Structural Research Centre of Oil and Gas Bearing Basin of Ministry of Education, Hangzhou 310058, China
  • Received:2023-09-01 Revised:2023-11-08 Online:2024-04-20 Published:2024-05-29

深度学习在活动构造与地貌研究中的应用

刘鑫1,2)(), 王诗柔1)(), 石许华1,2),*(), 苏程1),*(), 鲁晨妍1), 钱晓园1), 孙侨阳1,2), 邓洪旦1,2), 杨蓉1,2), 程晓敢1,2)   

  1. 1) 浙江省地学大数据与地球深部资源重点实验室, 浙江大学, 地球科学学院, 杭州 310058
    2) 教育部含油气盆地构造研究中心, 杭州 310058
  • 通讯作者: *石许华, 男, 1982年生, 研究员, 博士生导师, 主要从事构造地貌、 活动构造和地震地质究, E-mail: shixuhua@zju.edu.cn。苏程, 男, 1985年生, 副教授, 硕士生导师, 主要从事遥感方向, 数字图像处理研究, E-mail: sc20184@zju.edu.cn
  • 作者简介:

    刘鑫, 男, 1998年生, 现为浙江大学构造地质学专业在读博士研究生, 主要从事活动构造与地震地质研究, E-mail:

    共同第一作者: 王诗柔, 女, 1999年生, 现为浙江大遥感与地理信息系统专业在读学博士研究生, 主要从事遥感方向深度学习研究, E-mail:

  • 基金资助:
    国家自然科学基金(41972227); 国家自然科学基金(41941016); 国家自然科学基金(51988101); 浙江省钱江人才计划项目(QJD190202); 浙江大学百人计划项目

Abstract:

The research on active tectonics and geomorphology involves extensive sub-topics, including the kinematics of crustal movements, the processes underlying the evolution of landforms, and the associated dynamic mechanisms. These sub-topics are intricately connected with the interactions between the Earth’s endogenic and exogenic processes. In the contemporary realm of the Earth system science, research in active tectonics and geomorphology has become a hot topic for interdisciplinary study. The advancement in big data research coupled with the progressive developments in deep learning technologies has furnished this field of study with a voluminous array of data sources and the requisite analytical tools for technical analysis. In recent years, the application of big data and deep learning technologies in this research field has yielded a series of outstanding results, fostering new research directions, and ushering the discipline into a new phase. In this paper we synthesize existing research to outline the data sources pertinent to the study of active tectonics and geomorphology, including field geological survey, unmanned aerial vehicle (UAV)-based photography, aerial photography, and remote sensing observations. Then, we discuss in-depth examination of the recent innovations progresses in deep learning algorithms, including but not limited to convolutional neural networks(CNNs), deep Gaussian processes, and autoencoders. This article further elaborates on innovative applications of deep learning in the study of active tectonics and geomorphology. These include the identification of changes in glacier extent, monitoring volcanic activity and deformation, recognizing river systems, precise surveillance of landslide events, as well as observations of lithospheric deformation co-seismic surface ruptures.

Based on the summary of prior studies, this paper showcases a distinct application instance. By employing convolutional neural networks(CNNs)within the realm of deep learning image analysis and utilizing UAV-obtained high-resolution images, we undertake the automated detection of structural fractures in granite rocks in Meizhou island, in the southeast of Fujian province, China. In fault damage zones, structural rock fractures are widely developed, and the study of their orientation, system, and secondary characteristics is of great importance for determining their mechanisms of development and the multi-phase tectonic activity events in the region. Under conventional methodologies, the study of structural fractures in rocks is time-consuming and requires considerable manual effort in conducting exhaustive field surveys and detailed interpretation of cartographic representations. However, the application of deep learning can greatly enhance the efficiency of cartographic work. This application case has improved the classic deep learning framework by developing a CNN model specifically designed for the extraction of complex features and multi-scale rock fractures. This model achieved rapid identification of over 9 000 fractures with varied shapes and complex distributions within 55 minutes, attaining an accuracy of 85% and a recall rate of 89%. These findings demonstrate that deep learning significantly enhances operational efficiency in comparison to manual statistical methods for the automated identification of rock structural fractures, while also maintaining exceptional accuracy in fracture detection. Based on the results identified by deep learning, it can be clearly observed that two sets of fractures, oriented NE and NW, develop on the granite outcrops in the study area. According to previous research and the cross-cutting relationships of the fractures, it is known that NE-oriented fractures formed earlier than NW-oriented fractures, corresponding respectively to the Indosinian Movement and the expansion movement of the South China Sea in the tectonic history of South China. Through the automated extraction of deep learning models, the workload of manual mapping can be greatly reduced, yielding results consistent with actual geomorphological phenomena.

Key words: machine learning, deep learning, active tectonics, geomorphology, automatic identification

摘要:

活动构造与地貌学主要涉及活动构造的运动学、 地貌的演化过程及其相关动力机制, 该研究方向是近几十年来地球系统科学交叉研究的热点之一。随着大数据与机器学习研究的发展, 活动构造与地貌学的研究和深度学习的结合已成为该领域中受到广泛关注的新兴研究方向, 并产出了大量优秀成果。文中总结并综述了现今深度学习在活动构造与地貌研究中的数据来源, 以及利用深度学习的方法定量化解决活动构造与地貌中的重要科学问题(包括冰川识别、 火山活动与地貌、 水系分析、 滑坡监测和地表形变等)。基于对上述实例的探索, 文中运用深度学习中的卷积神经网络, 对华南东南部福建地区的花岗岩岩石构造裂缝开展了基于高精度无人机航拍影像的深度学习自动识别。所搭建的卷积网络模型在55min的运行时间内自动识别出人工需消耗近一周才可识别的9 000余条裂缝, 并获得了85%的查准率与89%的查全率, 表明该模型在准确识别构造裂缝的同时显著提升了工作效率。文中最后讨论并展望了未来深度学习方法在活动构造与地貌学领域的发展前景。

关键词: 机器学习, 深度学习, 活动构造, 地貌, 自动识别