基于钻孔伪标签半监督深度学习的沈阳市浅表地层三维建模研究  

3D Geological Modeling of Shallow Strata in Shenyang City Using Semi-supervisedDeep Learning Method with Pseudo-labels Based on Boreholes

作　　者：徐学闯 张恒兵[2] 符韶华[2] 刘升传[2] 王徐磊 郭甲腾[1] 敦力民[2] XU Xuechuang;ZHANG Hengbing;FU Shaohua;LIU Shengchuan;WANG Xulei;GUO Jiateng;DUN Limin(School of Resources and Civil Engineering,Northeastern University,Shenyang 110004;Shenyang Geotechnical Investigation&Surveying Research Institute Co.,Ltd.,Shenyang 110004,China)

机构地区：[1]东北大学资源与土木工程学院,辽宁沈阳110004 [2]沈阳市勘察测绘研究院有限公司,辽宁沈阳110004

出　　处：《地理与地理信息科学》2023年第6期9-17,共9页Geography and Geo-Information Science

基　　金：国家自然科学基金项目(42172327);中央高校基本科研业务费专项资金项目(N2201022)。

摘　　要：三维地质模型对于城市地下空间资源开发及工程建设具有重要价值,但基于大规模钻孔数据的城市三维地质建模及地层预测仍缺乏有效手段。该文针对大范围第四系冲积平原地貌的浅层三维地质建模问题,提出一种基于钻孔数据伪标签的半监督深度学习自动建模方法。首先将升采样的钻孔数据输入人工神经网络,然后对待预测网格点添加高置信度伪标签数据,最后结合地貌单元约束预测网格的地层属性。依据沈阳市三环范围内1.2万条岩土工程勘察钻孔数据,应用该方法在阶地、漫滩等地貌单元约束下建立了城市整体三维地层模型,与隐式曲面建模结果对比表明,该方法建立的三维地质模型有效揭示了56个标准地层的覆盖关系,建模结果整体与沈阳市地质情况相符,局部与钻孔数据保持一致,预测能力较强,可为基于大规模、多地层、分布不均匀钻孔数据的高精度深度学习三维地质模型构建提供借鉴。3D geological model is of great value for the development of urban underground space resources and engineering construction.However,there is still a lack of effective means for urban 3D geological modeling and stratum prediction based on a large amount of borehole data.Aiming at the problem of shallow 3D geological modeling of large-scale Quaternary alluvial plain landform,this paper proposes a automatic modeling method using semi-supervised deep learning with pseudo-labels based on borehole data.The method first inputs the upsampled borehole data into the artificial neural network,then adds high-confidence pseudo-label data to the grid points to be predicted,and finally predicts the stratum attributes of the grid with geomorphic unit constraints.Based on the data of 12000 geotechnical engineering survey borehole data within the scope of the third Ring road of Shenyang City,the 3D stratum model of the city is established.Compared with the implicit modeling results,the 3D geological model established by the proposed method effectively reveals the coverage relationship of 56 standard strata under the constraints of terrace,floodplain and other geomorphic units.The overall modeling result is consistent with the geological conditions of Shenyang City,while the local results are consistent with the borehole data,demonstrating good predictive ability.It provides a reference idea for high-precision deep learning 3D geological modeling from large-scale,multi-stratum and uneven distribution borehole data.

关 键 词：数字城市 三维地质建模 深度学习 钻孔数据 半监督学习 伪标签 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] P628[自动化与计算机技术—控制科学与工程]