地震地质 ›› 2024, Vol. 46 ›› Issue (3): 739-755.DOI: 10.3969/j.issn.0253-4967.2024.03.013

• 新技术应用 • 上一篇    下一篇

基于RS-Conv的多尺度神经网络LiDAR点云断裂带提取方法

宋冬梅1)(), 王浩1),*(), 冯家兴2), 单新建3), 王斌1)   

  1. 1) 中国石油大学(华东), 海洋与空间信息学院, 青岛 266580
    2) 东方通用航空摄影有限公司, 太原 030032
    3) 中国地震局地质研究所, 北京 100029
  • 收稿日期:2023-02-14 修回日期:2023-03-29 出版日期:2024-06-20 发布日期:2024-07-19
  • 通讯作者: *王浩, 男, 1997年生, 硕士, 主要从事三维激光点云地物提取方面的研究, E-mail: z20160115@s.upc.edu.cn。
  • 作者简介:

    宋冬梅, 女, 1973年生, 2003年于中国科学院沈阳应用生态研究所获景观生态学专业博士学位, 教授, 现主要研究方向为海洋灾害遥感与遥感影像智能算法研究, E-mail:

  • 基金资助:
    国家重点研发计划项目(2019YFC16D9202-4); 国家自然科学基金(U22A20586); 国家自然科学基金(41701513); 国家自然科学基金(41772350); 山东省自然科学基金(ZR2022MD015); 中国地震科学实验场地震构造探查系统项目共同资助。

A FRACTURE ZONE EXTRACTION METHOD FOR LIDAR POINT CLOUD BASED ON MULTI-SCALE NEURAL NETWORK WITH RS-CONV

SONG Dong-mei1)(), WANG Hao1),*(), FENG Jia-xing2), SHAN Xin-jian3), WANG Bin1)   

  1. 1) College of Oceanography and Spatial Information, China University of Petroleum, Qingdao 266580, China
    2) Oriental General Aerial Photography Co., Taiyuan 030032, China
    3) Institute of Geology, China Earthquake Administration, Beijing 100029, China
  • Received:2023-02-14 Revised:2023-03-29 Online:2024-06-20 Published:2024-07-19

摘要:

断裂带与地震、 滑坡等自然灾害的发生有着密切关系, 精准提取断裂带不仅可为地震断层的定量化研究提供指导, 还可为地震灾害风险评估及防震减灾决策的制定提供科学依据。针对现有方法中LiDAR点云断裂带提取不完整、 连续性差及错误率高等问题, 文中提出了一种基于 RS-Conv 的多尺度神经网络LiDAR点云断裂带提取方法, 以便更好地解决复杂地形区域的断裂带自动提取问题。该方法首先构建不同空间尺度的邻域点集, 从而更全面地考察点云的局部几何结构特征。考虑到RS-Conv算子能够很好地表征中心点与邻域点的空间关系, 文中以RS-Conv算子作为卷积模块构建了多尺度神经网络模型, 以提取出LiDAR点云不同尺度的深层次特征, 对其进行堆叠并输入到全连接层, 以完成对断裂带点的提取。最后, 在ISPRS点云数据集、 川滇点云数据集和鲜水河数据集上对文中所述方法与张量分解方法和Deep Neural Networks(DNN)方法进行了对比实验, 结果表明, 文中方法的分类精度最高, 分类总误差最低仅为0.3%, 较其他方法降低了0.91%~2.79%, 证实了该方法在点云断裂带提取方面的优越性。

关键词: LiDAR点云, 多尺度邻域点集, 深度学习, 断裂带提取

Abstract:

Fracture zones are geological formations resulting from the strong movement of the Earth’s crust, often manifesting as fragile and sensitive areas. These zones are closely linked to natural disasters such as earthquakes and landslides. Accurate extraction of fracture zones is crucial for quantitative studies of earthquake faults, providing a scientific basis for risk assessment and decision-making in earthquake prevention and mitigation. Thus, an in-depth study to determine their distribution patterns and surface geometry is essential for understanding earthquake dynamics and mechanisms.

This paper addresses the shortcomings of existing methods in extracting fracture zones from LiDAR point clouds, which often suffer from incomplete extraction, poor continuity, and high error rates. We propose a method based on a multi-scale neural network with RS-Conv to improve the automatic extraction of fault zones in complex terrain regions. Fracture zones exhibit complex morphologies and scale features; therefore, single-scale neighborhood point sets fail to reveal their intrinsic structural information fully. Our approach begins by constructing neighborhood point sets at different spatial scales to comprehensively examine geometric features at various levels within the point cloud. The RS-Conv operator effectively portrays the spatial relationship between the center point and neighboring points. We then build a multi-scale neural network model using the RS-Conv operator as the convolution module. This model captures the spatial relationships in the point cloud, efficiently extracting deep features at different scales. The extracted multi-scale features are concatenated to form a richer and more comprehensive feature representation, which is inputted into a fully connected layer to classify the centroid and solve the fracture zone extraction problem. We compared our method with the Tensor Decomposition and Deep Neural Networks(DNN)methods using the ISPRS point cloud dataset, the Sichuan-Yunnan point cloud dataset, and the Xianshuihe dataset. Results show that our method achieves the highest classification accuracy across all three datasets. Specifically, our method’s total classification error is only 0.3%, a reduction of 0.91% -2.79%compared to other methods. This significant error reduction demonstrates the accuracy, stability, and reliability of our proposed method in handling complex point cloud data. The main conclusions of this study are as follows:

(1)The construction of neighborhood point sets at different scales reveals that the combination of these scales significantly impacts the model’s classification performance. Selecting appropriate scale combinations is crucial for optimizing the model’s classification accuracy, facilitating better distinction between fracture zone points and non-fracture zone points.

(2)Compared to traditional and machine learning methods, the deep learning network model developed in this study shows significant advantages in extracting fracture zones from point clouds. The model can automatically learn deep features from point cloud data and process large-scale, high-dimensional point cloud datasets, thereby achieving more accurate fracture zone extraction in complex terrain conditions.

(3)Comparative experiments on different datasets further demonstrate the proposed method’s generalization ability. It is effective not only in extracting fracture zones under single terrain conditions but also in maintaining stable performance across multiple terrain conditions. This adaptability enhances the extraction of fracture zones in various terrain scenarios.

In conclusion, the method proposed in this paper offers a novel approach to fracture zone extraction. It achieves higher classification accuracy compared to existing traditional and machine learning methods, effectively addressing the challenge of fracture zone extraction in complex terrain areas.

Key words: LiDAR point cloud, multi-scale neighborhood point sets, deep learning, fracture zone extraction