中圖分類號:P631 文獻標識碼:A DOI:10.13810/j.cnki.issn.1000-7210.20240517
Abstract:The gravity anomaly inversion,which infers the density distribution subsurface anomalies from sur face gravity data,is an essential toolin geophysical explorationand is widelyappliedin fields suchasoilfields, mineral deposits,geological structures,and underground works detection. Traditional gravity inversion methods face challenges complex computation,low resolution,and dependence on prior information for inversion results. However, deep learning-based gravity anomaly inversion techniques show significant advantages,par ticularly in terms improving inversion accuracy and reducing computation time,without the reliance on initial models or prior information. This paper reviews the development and limitations traditional gravity anomaly forward and inversion methods and summarizes the current research on deep learning-based gravity inversion methods.Meanwhile,it introduces the improvements and innovations diffrent gravity inversion problems in four respects,including data preparation, network models,network optimization,and network validation.Additionally,,the application effect various gravity inversion methods on the measured data from Vinton Dome in Louisiana,the USA,and the San Nicolas ore deposit in Mexico. The multi-task framework CDUNet yields the most accurate inversion depth values on data Vinton Dome,while the 3D U-Net+ + network obtains clearer and more accurate inversion results on the data the San Nicolas ore deposit than the U-Net network. Keywords: gravity anomaly inversion,deep learning,data-driven,network model,network optimization
黃興業(yè),胡青青,鄺文俊,等,深度學(xué)習(xí)重力異常反演方法發(fā)展綜述[J].石油地球物理勘探,2025,60(4):1046-1058. HUANG Xingye,HU Qingqing, KUANG Wenjun,et al. Review deep learning-based gravity anomaly inversion methods[J]. Oil Geophysical Prospecting,2025,60(4) :1046-1058.
0 引言
下密度異常體的空間位置、幾何結(jié)構(gòu)、密度等參數(shù),廣泛應(yīng)用于石油勘探、礦產(chǎn)勘查、地質(zhì)結(jié)構(gòu)分析和地下工程探測等領(lǐng)域。傳統(tǒng)的重力反演方法受重力場的多解性以及核函數(shù)隨深度衰減等問題的影響,對重力異常反演通過地表重力場異常數(shù)據(jù)推算地深部區(qū)域的研究難以獲得高精度的反演結(jié)果。
近年來,深度學(xué)習(xí)技術(shù)取得了快速的發(fā)展,在各類任務(wù)中應(yīng)用效果顯著。Krizhevsky等1提出了AlexNet模型,研究表明深度學(xué)習(xí)對處理復(fù)雜重力場具有很好的效果;Simonyan等2提出的VGGNet深層網(wǎng)絡(luò)在圖像識別方面具有較好的效果;Szegedy等提出的GoogLeNet網(wǎng)絡(luò)實現(xiàn)了多尺度特征提取;He等4提出的ResNet網(wǎng)絡(luò)解決了網(wǎng)絡(luò)加深后梯度消失的問題;Ronneberger等[5提出的端到端網(wǎng)絡(luò)結(jié)構(gòu)U-Net在醫(yī)學(xué)圖像處理和語義分割等任務(wù)中得到了廣泛應(yīng)用;Dosovitskiy等將VisionTransformer引入自注意力機制進行圖像處理,挑戰(zhàn)了卷積神經(jīng)網(wǎng)絡(luò)在圖像處理領(lǐng)域的主流地位。
深度學(xué)習(xí)因其具有強大的非線性映射能力而為重力反演提供了新的解決方案。利用大量數(shù)據(jù)訓(xùn)練網(wǎng)絡(luò)模型,使其能夠從數(shù)據(jù)中自動學(xué)習(xí)特征,突破了傳統(tǒng)反演方法對先驗信息的高度依賴性,并在復(fù)雜的地下密度結(jié)構(gòu)預(yù)測中表現(xiàn)出更好的泛化能力。針對不同反演問題的特點,人們提出了不同改進技術(shù):在數(shù)據(jù)準備階段,引入隨機游走模型[提高合成數(shù)據(jù)的模型復(fù)雜度,增強了網(wǎng)絡(luò)的泛化能力;……