GNGTS 2021 - Atti del 39° Convegno Nazionale

395 GNGTS 2021 S essione 3.1 REGULARIZATION OF 2D SEISMIC SURVEYS VIA SYMMETRIC OUTPUT DEEP PRIOR V. Lipari 1 , F. Picetti 1 , P. Bestagini 1 and S. Tubaro 1 1 Politecnico di Milano Introduction Several seismic processing algorithms take advantage of seismic data with regular and dense spatial sampling. An important example in this regard is Surface Related Multiple Elimination (SRME), which uses the data itself for multiple prediction. For this reason, its performance is highly dependent on the data used. Ideally, both shot and receiver coordinates should be on the same regular grid satisfying the Nyquist criterion. Inpractice, geophysicists use a suite of regularizationmethods relying of different assumptions: for instance, moveout based techniques assume that the underlying velocity model and wavefield are simple; f-x (Spitz, 1991) or f-k interpolation (Gülünay, 2003) assume that spatially aliased events are locally comprised of plane waves; sparsity-based interpolation assumes that spatial sampling can be considered randomized. Often, a cascade of multiple approaches is applied demanding an extra effort for parameters tuning. More recently, also deep learning methods exploiting convolutional neural networks (CNNs) have been proposed for seismic interpolation tasks, for instance residual CNNs (Wang et al., 2019), conditional GANs (Oliveira et al., 2018), convolutional autoencoders (Mandelli et al., 2018) and CycleGANs (Pham and Fomel, 2019). Most of the proposed CNN interpolation strategies make use of the standard learning paradigm: (i) during a training stage, the CNN learns to perform the interpolation task on a dataset made of pairs of complete/decimated data, and (ii) a deployment stage where the trained network is used to interpolate different data. The achieved performance heavily depends on the size and the features of the training dataset and can be very difficult to generalize to different data. More recently, based on the concept of deep-image-prior (Ulyanov et al., 2018), CNN-based seismic interpolation methods that do not need a training stage have been presented (Liu et al., 2019; Kong et al., 2020). However, these methods were mainly devoted to the task of common shot gathers (CSG) interpolation. This work proposes a preliminary study of deep prior based 2D seismic data regularization. The goal of the proposed method is to obtain a complete acquisition geometry where both sources and receivers are located on the same dense regular grid. This is achieved by adopting 3D kernels in the MultiResolution Unet we exploit for CSG interpolation, and by forcing the output to be symmetric to introduce source-receiver reciprocity. Through a synthetic example based on the well-known Marmousi model, we show that the proposed technique is very promising for the task of seismic data regularization, and the modifications introduced into the network architecture greatly improve the performance of the method. Regularization by deep prior A dense 2D ( Regularization of 2D seismic surveys via symmetric output deep prior Vincenzo Lipari, Francesco Picetti 2 , Paolo Bestagini 2 and Stefano Tubaro 2 INTRODUCTION Several seismic processing algorithms take advantage of seismic data with regular and dens sampling. An important example in this regard is Surface Related Multiple Elimination (SRM uses the data itself for multiple prediction. For this reason, its performance is highly dependent on used. Ideally, both shot and receiver coordinates should be on the same regular grid satisfying the criterion. In practice, geophysicists use a suite of regularization methods relying of different assumpt instance, moveout based techniques assume that the underlying velocity model and wavefield are s x (Spitz, 1991) or f-k i erpolation (Gülünay, 2003) assume that spatially aliased events ar comprised of plane waves; sparsity-based interpolation assumes that spatial sampling can be c randomized. Often, a cascade of multiple approaches is applied demanding an extra effort for p tuning. More recently, also deep learning methods exploiting convolutional neural networks (CN been proposed for seismic interpolation tasks, for instance residual CNNs (Wang t a ., 2019), co GANs (Oliveira et al., 2018), convolutional autoencoders (Mandelli et al., 2018) and CycleGA and Fomel, 2019). Most of the proposed CNN interpolation strategies make use of the standard learning paradigm: (i) training stage, the CNN learns to perform the interpolation task on a dataset made of complete/decimated data, and (ii) a deployment stage where the trained network is used to in different data. The achieved performance heavily depends on the size and the features of the dataset and can be very difficult to generalize to different data. More recently, based on the c deep-image-prior (Ulyanov et al., 2018), CNN-based seismic interpolation methods that do no training stage have been presented (Liu et al., 2019; Kong et al., 2020). However, these meth mainly devoted to the task of common shot gathers (CSG) interpolation. This work proposes a preliminary study of deep prior based 2D seismic data regularization. The the proposed method is to obtain a complete acquisition geometry where both sources and recei located on the same dense regular grid. This is achieved by adopting 3D kernels in the MultiRe Unet we exploit for CSG interpolation, and by forcing the output to be symmetric to introduce receiver reciprocity. Through a synthetic example based on the well-known Marmousi model, that the proposed technique is very promising for the task of seismic data regularization, modifications introduced into the network architecture greatly improve the performance of the met REGULARIZATION BY DEEP PRIOR A dense 2 x ,t ) dataset where N sources and N receivers are located in the same positions on a spatial grid can be arranged as a 3D cube X = ( x S , x R , t ) where x S = x S 1 , x S 2 , ⋯ , x S N are source po x R = x R 1 , x R 2 , ⋯ , x R N are receiver positions and t = t 0 , ⋯ , T are time samples. In the ideal case the c x R slices are common shot gathers, constant x S slices are common receiver gathers and constant are the so-called time slices. Notice that source-receiver reciprocity implies that the time sli symmetric N ×N matrices. e sources and N rec iv rs are located in the same positions on a regular spatial grid can be arranged as a 3D cube Vincenzo Lipari, Francesco Picetti 2 , Paolo Bestagini 2 and Stefano Tubaro 2 INTRODUCTION Several seismic processing algorithms take advantage of seismic data with regular and sampling. An important example in this regard is Surface Related Multiple Elimination (S uses the data itself for multiple prediction. For this reason, its performance is highly depende used. Ideally, both shot and receiver coordinates should be on the same regular grid satisfyin criterion. In practice, geophysicists use a suite of regularization methods relying of different ass instance, moveout based techniques assume that the underlying velocity model and wavefield x (Spitz, 1991) or f-k interpolation (Gülünay, 2003) assume that spatially aliased event comprised of plane waves; sparsity-based interpolation assumes that spatial sampling can randomized. Often, cascade of multiple approaches is appl ed demanding a extra effort f tuning. More recently, also deep learning methods exploiting convolutional neural networks been proposed for seismic interpolation tasks, for instance residual CNNs (Wang et al., 2019 GANs (Oliveira et al., 2018), convolutional autoencoders (Mandelli et al., 2018) and Cycle and Fomel, 2019). Most of the proposed CNN interpolation strategies make use of the standard l arning paradig training stage, the CNN learns to perform the interpolation task on a dataset made complete/decimated data, and (ii) a deployment stage where the trained network is used t different data. The achieved performance heavily depends on the size and the features of dataset and can be very difficult to generalize to different data. More recently, based on t deep-image-prior (Ulyanov et al., 2018), CNN-based seismic interpolation methods that d training stage have been presented (Liu et al., 2019; Kong et al., 2020). However, these mainly devoted to the task f common shot gathers (CSG) int rpolation. This work proposes a preliminary study of d ep prior based 2D s ismic data regularization. the proposed method is to obtain a complete acquisition geometry where both sources and r located on the same dense regular grid. This is achieved by adopting 3D kernels in the Mul Unet we exploit f r CSG interpolation, and by fo cing the output to be symmetric to intro receiver reciprocity. Through a synthetic example based on the well-known Marmousi mod that the proposed technique is very promising for the task of seismic data regularizati modifications introduced into the network architecture greatly improve the performance of the REGULARIZATION BY DEEP PRIOR A dense 2D ( x ,t ) dataset where N sources and N receivers are located in the same positions spatial grid can be arr nged as a 3D c e X = ( x S , x R , t ) where x S = x S 1 , x S 2 , ⋯ , x S N are sourc x R = x R 1 , x R 2 , ⋯ , x R N are receiver positions and t = t 0 , ⋯ , T are time samples. In the ideal case x R slices are common shot gathers, constant x S slices are common receiver gathers and cons are the so-called time slices. Notice that source-receiver reciprocity implies that the tim symmetric N ×N matrices. here Vincenzo Lipari, Francesco Picetti 2 , Paolo Bestagini 2 and Stefano Tubaro 2 INTRODUCTION Several seismic proces ing algorithms take advantage of seismic data with regular and d sampling. An important example in this regard is Surface Related Multiple Elimination (SR uses the data itself for multiple prediction. For this reason, its performance is highly dependen used. Ideally, both shot and receiver co rdinates should be on the same regular grid satisfying criterion. In practice, geophysicists use a suite of regularization methods relying of different as u instance, moveout based techniques as ume that he underlying velocity model and wavefield a x (Spitz, 19 1) or f-k interpolation (Gülünay, 20 3) as ume that spatially aliased events comprised of plane waves; sparsity-based interpolation assumes that spatial sampling can b randomized. Oft n, a cascade of multiple approaches is ap lied demanding an extra ef ort fo tuning. More recently, also de p learning methods exploiting convolutional neural networks ( be n proposed for seismic interpolation tasks, for instance residual CNNs (Wang et al., 2019) GANs (Oliveira et al., 2018), convolutional autoencoders (Mandelli et al., 2018) and CycleG and F mel, 2019). ost of the proposed CN interpolation strategies make use of the standard learning paradigm training stage, the CNN learns to perform the interpolation task on a dataset made complete/decimated data, and (ii) a deployment stage where the trained network is used t different data. The achieved performance heavily d pends on the size and the features of dataset and can be very difficult to generalize to dif erent data. More recently, based on th de p-image-prior (Ulyanov et al., 2018), CNN-based seismic interpolation methods that do training stage have be n presented (Liu et al., 2019; Kong et al., 2020). However, these m mainly devoted to the task of ommon shot gathers (CSG) interpo ation. This work proposes a pr liminary study of deep prior based 2D seismic data regularization. the proposed method is to obtain a complete acquisition geometry where both sources and re located on the same dense regular grid. This is achieved by adopting 3D kernels in the Multi Unet w xploit for CSG i erpolation, and by forc ng the output to e symmetric to introd receiver reciprocity. Through a synthetic example based on the well-known Marmousi mode that the proposed technique is very promising for the task of seismic data regularizatio modifications introduced into the network architecture greatly improve the performance of the REGULARIZATION BY DE P PRIOR A dense 2D ( x ,t ) dataset where N sources and N receivers are located in the same positions o spatial grid can be arranged as a 3D cube X = ( x S , x R , t ) where x S x S 1 , x S 2 , ⋯ , x S N are source x R x R 1 , x R 2 , ⋯ , x R N are receiver positions and t = t 0 , ⋯ , T are time samples. In the ideal case t x R slices are com on shot gathers, constant x S slices are common receiver gathers and const are the so-called time slices. Notice that source-receiver reciprocity implies that the time symmetric N ×N matrices. are source positions, Vincenzo Lipari, Francesco Picetti 2 , Paolo B stagini 2 and Stefano Tubaro 2 INTRODUCTION Several seismic processing algorithms take advantage of seismic data wit sampling. An important example in this regard is Surface R lated Multiple uses the data itself for multiple prediction. For this reason, its performance is used. Ideally, both shot and receiver coordinates should be on the same regula criterion. In practice, geophysicists use a suite of regularization methods relying of instance, moveo t based techniques assume that the underlying velocity model x (Spitz, 1991) or f-k interpolation (Gülünay, 2003) assume that spatially compris d f plane waves; sparsity-based int rpolation assumes that spatial randomized. Often, a cascade of multiple approaches is applied demanding a tuning. More recently, also deep learning methods exploiting convolutional n been proposed fo seismic interpolation tasks, for instance residual CNNs (W GANs (Oliveira et al., 2018), convolutional autoencoders (Mandelli et al., 20 and Fomel, 2019). Most of the proposed CNN interpolation strategies make use of the standard le training stage, the CNN l arns to perform the interpolation task on a complete/decimat d data, and (ii) a deployment stage where the trained net different data. The achiev d performance heavily depends on the size and dataset and can be very difficult to generalize to different data. More recent deep-image-pr or (Uly nov et al., 2018), CNN-based seismic interpolation training stage have been presented (Liu et al., 2019; Ko g et al., 2020). H mainly devoted to the task of common shot gathers (CSG) interpolation. This work proposes a preliminary study of deep prior based 2D seismic data the proposed method is to obtain a complete acquisition geometry where bot locat d o the same dense regular grid. This is achieved by adopting 3D ker Unet we exploit for CSG interpolation, and by forcing the output to be sym receiver reciprocity. Through a synthetic example ased on the well-known that the proposed t chnique is very promisi g for the task of seismic da modifications introduced into the network ar hitecture greatly improve the perf REGULARIZATION BY DEEP PRIOR A dense 2D ( x ,t ) dataset where N sources and N receivers are located in the s spatial grid can be arranged as a 3D cube X = ( x S , x R , t ) where x S = x S 1 , x S 2 , x R = x R 1 , x R 2 , ⋯ , x R N are receiver positions and t = t 0 , ⋯ , T are time samples. In x R slices are common shot gathers, constant x S slices are common receiver ga are the so-called time slices. Notice that source-receiver reciprocity implie symmetric N ×N matrices. are receiv positions a d Vincenzo Lipari, Francesco Picetti 2 , Paolo B stagini 2 and Stefano Tubaro 2 I T TI Several seis ic proce sing algorithms take advantage of seismic data wit sa pling. An important example in this regard is Surface Related Multiple uses the data itself for multiple predi tion. For this reason, its performance is used. Idea ly, both shot and receiver c ordinates should be on the sa e regul criterion. In practice, geophysicists use a suite of regularization methods relying o instance, moveout based techniques a sume that the underlying velocity mode x (Spitz, 1 91) or f-k i erpolation (Gülünay, 2003) assum that spatia l compris d of pla e waves; sparsity-based interpolation assumes that spati l randomized. Often, a cascade of multiple appro ches is a plied dem ding a tuning. More recently, also deep learning methods exploiting convolutional n been proposed for seismic i terpolation tasks, for inst nce residual CNNs (W GANs (Oliveira et al., 2018), convolutional utoencoders (Mande li et al., 2 and Fo el, 2019). ost of the proposed C N interpolation strategies make use of the standard le training stage, the CNN learns to perform the interpolation task on a complete/decimated data, and ( i) a deployment stage where the trained ne di ferent data. The achieved performance eavily depends on the size and dataset and can be very di ficult to generalize to di ferent data. More recen d ep-image-prior (Ulyanov et al., 2018), CNN-based seismic interpolation trai ing stage have b n pr s nted (Liu et al., 2019; Kong et al., 2020). H ainly devoted to the task of co on shot gathers (CSG) interpolation. This work proposes a preliminary study of d ep prior based 2D seismic data the proposed method is to obtain a complete acquisition g ometry where bot located on the sa e dense regular grid. Th is achi v d by adopting 3D ker net we exploit for CS interpolation, and by forcing the output to be sym receiver reciprocity. Through a synthetic example based on the we l-known that the proposed technique is very promising for the task of seismic d difications introduced into the network architecture greatly improve the per E L IZ TI B D EP P I dense 2 ( x ,t ) dataset where N sources and N receivers are located in the spatial grid can be a ranged as a 3 cube X ( x S , x R , t ) where x S x S 1 , x S 2 , x R x R 1 , x R 2 , , x R N r r cei er o itions n t = t 0 , ⋯ , T are time samples. I x R slices are co on shot gathers, constant x S slices are com on receiver g are the so-ca led time slices. Notice that source-receiver reciprocity impli symmetric ×N matrices. a e ti e samples. In the ideal case the constant Regularization of 2D seismic surveys via symmetric output d Vincenzo Lipari, Francesco Picetti 2 , Paolo Bestagini 2 and Stefan INTRODUCTION Several seismic processing algorithms take advantage of sei sampling. An important example in this regard is Surface Rel uses the data itself for multi le prediction. For this reason, i s p used. Ideally, both shot and receiver coordinates should be on t criterion. In practice, geophysicists use a suite of regularization meth instance, oveout based techniques assume that the und rlying x (Spitz, 1991) or f-k interpolation (Gülünay, 2003) assum comprised of plane waves; sparsity-bas d interpolation assum randomized. Often, a cascade of ultiple approach s is applie tuning. More recently, also deep learning methods exploiting c been proposed for s ismic interpolation tasks, for instance resi GANs (Oliveira et al., 2018), convolution l autoencoders (Ma and Fomel, 2019). Most of the proposed CNN interpolation strategies make use of training stage, the CNN learns to perform the interpolati complete/decimated data, and (ii) a deployment stage where different data. The achieved performance heavily depends on dataset and can be very difficult to generalize to different dat deep-image-prior (Ulyanov et al., 2018), CNN-based seismic tr in ng st ge hav bee presented (Liu et al., 2019; Kong et mainly devoted to the task of common shot gathers (CSG) interp This work proposes a preliminary study of deep prior based 2 the proposed method is to obtain a complete acquisition geom l cated on the same dense regular grid. This i achieved by ad Unet we exploit for CSG interpolation, and by forcing the out receiver reciprocity. Through a synthetic example based on th that the proposed technique is very promising for the task modific tions introduc d into the n twork architecture greatly i REGULARIZATION BY DEEP PRIOR A dense 2D ( x ,t ) dataset where N sources and N receivers are l spatial grid can be arranged as a 3D cube X = ( x S , x R , t ) where x R = x R 1 , x R 2 , ⋯ , x R N are receiver positions and t = t 0 , ⋯ , T ar ti x R slices are common shot gathers, constant x S slices are com are the so-call d time slices. Notice that ource-receiver reci symmetric N ×N matrices. slices re common shot gathers, c nstant INTRODUCTION Several seismic processing algorithms take advantage of seismi sampling. An i portant example in th s regard is Su face Relate uses the da a itself for multiple prediction. For this reason, ts perfo us d. Ideally, both shot and receiver coordinates should be on the criterion. In practice, geophysicists use a suite of regularization methods instance, moveout based techniques assume that the underlyi g vel x (Spitz, 1991) or f-k interpolation (Gülünay, 2003) assum th comprised of pla e waves; sparsity-based interpo ation assum s rando ized. Often, a cascade of multiple approaches is applied de tuning. More recently, also deep learning methods exploiting conv be n proposed for seismic interpolati tasks, for instanc residual GANs (Oliveira et al., 2018), convol tional autoencoders (Mandel a d Fomel, 2019). Most of th propo ed CNN inter olation strategies make use of the training s age, the CNN learns to perform the interpolation complete/decimated data, and (ii) a deployment stage where th different data. The achieved performance heavily depends on th datas t and can b very difficult to generalize to different data. deep-image-prior (Ulyanov et al., 2018), CNN-based seismic int training stage have been presented (Liu et l., 2019; Kong et al., mainly evoted to the task of common shot gathers (CSG) interpola This work proposes a preliminary study of deep prior based 2D se the proposed method is to obtain a complete acquisition geometry l c ted on the sa e dense regular grid. T is is ac ieved by dopt Unet we exploit for CSG interpolation, and by fo cing the outpu receiver reciprocity. Throu h a synth ic example based on the w that the proposed techniqu is very promising for the task f mod fications introduced into the network architecture greatly impr REGULARIZATION BY DEEP PRIOR A dense 2D ( x ,t ) dataset where N sources and N receivers are loca spatial grid can be arranged s a 3D cube X = ( x S , x R , t ) where x S x R = x R 1 , x R 2 , ⋯ , x R N are receive positions and t = t 0 , ⋯ , T are time x R slices are com on shot gathers, cons a t x S slices are common are the so-called time slices. Notice that source-receiver recipro symmetr N ×N matrices. ice are common receiver gathers and constant egularization of 2D seismic su veys via symmetric utput deep prior incenzo Lipari, Francesco Picetti 2 , Paolo Bestagini 2 and Stefano Tubaro 2 TRODUCTION everal seismic processing algorithms tak advantage of seismic d ta with r gular and d nse spatial mpling. An important ex mp e in this regard is Surface Related Multipl Elimination (SRME), which es the data itself for multiple predictio . For this reason, its performance is highly dependent on the data ed. Ideally, both sh t and receiver coordinates should be on the same regular grid satisfying the Nyquist iterion. practice, geophysicists use a suite of regularization methods r lying f differe t assumption : or stance, moveout based techniques assume that h under yi g velocity model and wavef eld are simple; f- (Spitz, 1991) or f-k interpolation (Gülünay, 2003) assume that spatially aliased events are locally mprised of plane waves; sparsity-based i terpolation assumes that spatial sampling can be cons dered ndomized. Ofte , a cascade of multiple approaches is appli d d mand ng n ex ra effort for param ters ning. More rec ntly, also eep l arning m thods exploiting c nvolutional neur l networks (CNN ) have en proposed for seismic nterpolation tasks, f r instance residual CNN Wa g et al., 2019), conditiona ANs (Oliveira et al., 2018), convolutional autoencoders (Mandelli et al., 2018) and CycleGANs (Pham d Fomel, 2019). ost of the proposed CNN interpolation strategies make use of the st ard le rning paradigm: (i) during a aining stage, the CNN learns to perform the interpolation task on a d taset made of pairs of mplete/decima ed data, and (ii) a d ployment stage where the trained network is used to nt rpol e fferent data. The achieved performance heavily d ends on the size an the features of the train ng taset and can b very diff cult to ge eralize to diff rent data. Mo rec ntly, based on the concept of ep-image-prio (Ulyanov t al., 2018), CNN-based s ismic interpolation met ds that do not need a aining stage have be presented (Liu et al., 2019; Kong et al., 2020). However, these methods were ainly devoted to the task of common shot gathers (CSG) interpolation. his work proposes a preliminary study of deep pri r based 2D seismic data egularization. The goal of e proposed method is to obt in a complet acquisition ge metry where both sources and receivers are cated on the same d nse regular grid. This is achieved by ad pting 3D kernels in the MultiResolutio net we exploit for CSG interpolation, and by forcing the output to be symmetric to ntroduce source- ceiver reciprocity. Through a synthetic exampl based on the well-known Marmousi model, we show at the proposed technique is very promising for th task of seismic ata regularization, and the odifications introduced into the n twork architecture greatly improve the performance of the method. EGULARIZATION BY DEEP PRIOR dense 2D ( x ,t ) dataset where N sources and N receivers are located in the same positions on a regular atial grid can be arranged as a 3D cube X = ( x S , x R , t ) where x S = x S 1 , x S 2 , ⋯ , x S N are source positions, R = x R 1 , x R 2 , ⋯ , x R N are receiv positi d t = t 0 , ⋯ , T are time samples. In the ideal case the constant R slices are common sh t gathers, constant x S sl s are c mmon receive gathers and constan t slic s e the so-called ti slices. otice that source-receiver reciprocity implies that the time slices are mmetric N ×N matrices. slices are the so-called tim slices. Notice that source- receiver reciprocity implies t at the tim slices are symmetric INTRODUCTION Several seismic processing algorithms take advant sampling. An import nt ex mple in this reg rd is S uses th data itself for multiple predictio . For h s re sed. Id ally, both sho and rec iver coordin tes sho crit rion. In practice, g ophysicists u e a suite of reg lariz instance, moveout based techniques assume that the u x (Spitz, 1991) or f-k interpolat on (Gülünay, 20 com rised of plane waves; sp r ity-based interpolat r ndomized. Often, a casc de of m ltiple ppro che tuning. More recently, lso deep learning ethods e been proposed for s ismic interpolation tasks, for i s GANs (Oliveira et al., 2018), onvolutional autoenc nd F mel, 2019). Most of the proposed CNN i terpolation strategies m training stage, the CNN learns to perform the complete/decimated da , a d (ii) a deployment st different d ta. The achieved performance heavily dataset and ca be very difficult to generalize o di deep-image-prio (Ulyanov et al., 2018), CNN-base training stage have been prese ted (Liu et a ., 201 ma nly devoted to the task of common shot g hers ( This work proposes a pr liminary study f deep pri the proposed method is to obtain a complete a quisi l ca ed on the sam dense regula grid. This is achi Unet we exploit for CSG int rpol tion, and by f rci receiver reciprocity. Thr ugh a synthetic example b that the proposed technique is very pr mising fo modif cations introduced into t network architectur REGULARIZATION BY DEEP PRIOR A dense 2D ( x ,t ) dataset where N sources and N rec spatial grid can be arranged as a 3D cube X = ( x S , x x R = x R 1 , x R 2 , ⋯ , x R N are r ceiver positions and t = t 0 , x R slic s are common shot gathers, onstant x S slice are t so- alled ime slices. Notice that sourc -re sy m tric N ×N mat ices. mat i s. Any acquisition geo etry Any acquisition geometry G ( x S , x R ) is defined by a N ×N binary matrix M ( x S , x R ) (i.e., a m follows: 1 is defined by a Vincenzo Lipari, Francesco Picetti , Paolo B stagini and S INTRODUCTION Several seismic processing algorithms take advantage o sampl ng. An important example in this reg rd is Surfac uses the data itself for multiple prediction. For this reason, used. Ideally, both shot and eceiv r co dinates should be crit rion. In practice, geophysicists use a suite of regularization instance, moveout based techniques assume that the und rl x (Spitz, 1991) or f-k interpolation (Gülünay, 2003) comprised of plane waves; sparsity-based interpolation a randomized. Often, a cascade of multiple approaches is a tuning. Mor recently, also deep learning methods exploit been propos d for seismic interpolation asks, for in tance GAN (Oliveira et al., 2018), convolutional autoencoders and Fomel, 2019). Most of the proposed CNN interpolation strategies make u training stage, the CNN le rns to perform the interp complete/d cimated data, and (ii) a deployment stage w different d a. The achiev d performance heavily depen dataset and can be very difficult to generalize to differen deep-im ge-prior (Ulyanov et al., 2018), CNN-b sed sei training stage have been presented (Liu et al., 2019; Ko mainly devoted to the task of common shot gathers (CSG) i This work pr poses a preliminary study of deep prior bas the proposed method is to obtain a complete acquisition g located on the same dense regular grid. This i achieved Unet we exploit for CSG interpolation, and by forcing th receiver reciprocity. Thro g a synthetic example based that the proposed technique is very promising for the modifications introduced nto the network architec ure grea REGULARIZATION BY DEEP PRIOR A dense 2D ( x ,t ) dataset where N sources and N receivers spatial grid can be arrang as a 3D cube X = ( x S , x R , t ) x R = x R 1 , x R 2 , ⋯ , x R N are receiver positions and t = t 0 , ⋯ , T x R slices are common shot g thers, constant x S slices are are he so-called time slices. Notice t at source-receive symmetric N ×N matrices. binary matr x Any acquisition ge m try G ( x S , x R ) is def ned by a N ×N binary m trix M ( S , x R ) (i.e., a mask) a follows: 1 if x , x G (i. ., a mask) as follows: