互補(bǔ)盲點(diǎn)策略和U型Transformer的地震數(shù)據(jù)去噪

打開(kāi)文本圖片集

中圖分類號(hào)：TP399 文獻(xiàn)標(biāo)志碼：A 文章編號(hào)：1001-3695（2025）07-018-2056-08

doi： 10.19734/j. issn.1001-3695.2024.12.0516

Abstract：Randomnoisedenoisingcanefectivelyimprovesthesignal-to-noiseratio（SNR）ofseismicdata.Blindspot-driven unsuperviseddenoising methodsdonotrequirelabeleddataandcanautomaticallextractfeatures，buttheyignore noisecorrelations，leading tosuboptimalperformance.Toaddress thisisse，thispaperproposed thecomplementaryblindspotstrategy andU-shapedTransformerseismicdenoising framework（CBUTS）.Firstly，thecomplementaryblind-spotstrategyusedtrace maskingandrandommaskingforcomplementarysamplingtoefectivelyweakenthespatialconnectionsof noise.Secondly，visibleblindspotlossfunction integrateddenoisedresultsfromboth non-blindandblindspots，reducing informationlo.Finall， the Transformer-based U-shaped blindspotnetwork（STU-Net）enhancedthecaptureof globalandlocal features，further weakened thenoisecorrelations，andmore accuratelypredictedvalidsignals.Experimentalresultsshow that，compared to classicalandadvancedsupervisedand unsupervised methods，CBUTSachievesbeterperformance indenoising noiseand preserving thecontinuityofseismicevents.Analysisandcomparisonconfirmtheapplicabilityof the method toseismicdata denoising.

Keywords：seismic data denoising；unsupervised；blind spot strategy；Transformer

0引言

由于勘探環(huán)境和測(cè)量設(shè)備等因素的影響，地震數(shù)據(jù)在采集過(guò)程中不可避免地會(huì)引入隨機(jī)噪聲。（剩余15864字）