基于SegNet卷积神经网络的三维重力反演方法  

A 3D Gravity Inversion Method Based on SegNet Convolutional Neural Network

作　　者：国为政 张毅 Guo Weizheng;Zhang Yi(School of Information Engineering,Institute of Disaster Prevention,Sanhe Hebei 065201,China;Key Laboratory of Geological Survey and Evaluation of Ministry of Education,China University of Geosciences,Wuhan Hubei 430074,China)

机构地区：[1]防灾科技学院信息工程学院,河北三河065201 [2]中国地质大学(武汉)地质探测与评估教育部重点实验室,湖北武汉430074

出　　处：《工程地球物理学报》2025年第2期304-315,共12页Chinese Journal of Engineering Geophysics

基　　金：国家自然科学基金面上项目(编号:42274113)。

摘　　要：重力反演是恢复地下目标的密度分布,用于解释地球的内部结构和分布的一种手段。然而,传统重力反演方法在处理复杂地质结构时存在一定的局限。近年来,随着深度学习(Deep Learning,DL)算法的快速发展,基于卷积神经网络(Convolutional Neural Network,CNN)的重力反演也取得了一定的好效果。本文基于SegNet(Semantic Segmentation Network,语义分割网络)卷积神经网络,提出了一种改进的重力反演方法,构建了4种密度异常体模型对应的训练数据集。通过对比理论模型反演结果和重建密度的均方根误差(Root Mean Square Error,RMSE),得出本方法对密度异常体的位置、密度以及埋深具有较好的反演效果;抗噪能力良好,同时具有较强的泛化能力,并且本方法被应用于墨西哥圣尼古拉斯硫化物铜锌矿区实际数据的反演,反演结果与真实边界吻合,高密度区域能清晰地显示出来,密度分布结果与真实区域吻合,进一步证明了本文所提方法的正确性和有效性。Gravity inversion serves as a critical technique for recovering the density distribution of subsurface targets,aiding in the elucidation of the Earth’s internal composition and distribution.However,traditional gravity inversion methods exhibit limitations when applied to complex geological structures.In recent years,the rapid advancement of deep learning(DL)algorithms has facilitated significant improvements in gravity inversion techniques based on convolutional neural network(CNN).This paper introduces an enhanced gravity inversion method based on the SegNet(semantic segmentation network)CNN.We have developed training datasets corresponding to four density anomaly volume models.By comparing the root mean square error(RMSE)between theoretical model inversion results and reconstructed densities,our proposed method demonstrates superior performance in accurately determining the location,density,and burial depth of density anomalies.Additionally,this approach exhibits robust noise resistance and generalization capabilities.The method was successfully applied to the inversion of real data from the San Nicolas massive sulfide copper-zinc deposit in Mexico,yielding inversion results that align well with actual boundaries.High-density regions are clearly displayed,and the density distribution closely matches the realregions,further validating the correctness and effectiveness of the proposed method.

关 键 词：深度学习 语义分割网络 重力异常 重力反演 

分 类 号：P631.1[天文地球—地质矿产勘探]