Systems and methods for seismic image generation using deep learning networks
Methods and systems are disclosed. The method include generating a simulated seismic training dataset, where the simulated seismic training dataset includes upgoing signals and downgoing signals, each recorded on hydrophones and geophones and training a deep learning network, using the simulated seismic training dataset, to separate the upgoing signals from the downgoing signals. The method also includes obtaining field hydrophone data and field geophone data pertaining to a subsurface region of interest, recorded by a seismic acquisition system including hydrophones and geophones, and separating the field hydrophone data and field geophone data into predicted upgoing signals. The method further includes generating a seismic image of the subsurface region of interest based, at least in part, on the predicted upgoing signals.
1 . A method, comprising:
generating a simulated seismic training dataset, wherein the simulated seismic training dataset comprises upgoing signals and downgoing signals, each recorded on hydrophones and geophones;
training a deep learning network, using the simulated seismic training dataset, to separate the upgoing signals from the downgoing signals;
obtaining, from a seismic acquisition system comprising hydrophones and geophones, field hydrophone data and field geophone data from a subsurface region of interest, the field hydrophone data and the field geophone data comprising observed upgoing signals and observed downgoing signals;
wherein the observed downgoing signals are formed in part by reflected downgoing ghost signals;
separating, using the trained deep learning network, the observed upgoing signals from the observed downgoing signals to yield predicted upgoing signals with the reflected downgoing ghost signals removed;
generating a seismic image of the subsurface region of interest based, at least in part, on the predicted upgoing signals;
identifying, using a seismic interpretation workstation, a drilling target based, at least in part, on the seismic image, and
planning a planned borehole trajectory, using a borehole planning system, based, at least in part, on the drilling target.
2 . The method of claim 1 , wherein generating the simulated seismic training dataset comprises:
creating, using a random number generator, a plurality of reflectivity surfaces; and
convolving a wavelet with the plurality of reflectivity surfaces.
3 . The method of claim 1 , wherein the deep learning network is a U-net neural network.
4 . The method of claim 3 , further comprising at least one operation selected from the group consisting of: a convolution, a max pooling, a transpose convolution, and a concatenation.
5 . The method of claim 1 , wherein generating the simulated seismic training dataset further comprises at least one selected from the group consisting of: creating coherent noise, random noise, and amplitude distortion.
6 . The method of claim 1 , wherein the simulated seismic training dataset further comprises:
an input dataset, comprising a linear superposition of upgoing signals and downgoing signals;
a first output dataset comprising upgoing signals; and
a second output dataset comprising downgoing signals.
7 . The method of claim 1 , wherein the hydrophones and geophones are arranged in collocated pairs, each pair comprising one hydrophone and one geophone.
8 . The method of claim 1 , wherein the seismic acquisition system comprises ocean bottom cables.
9 . The method of claim 2 , wherein the plurality of reflectivity surfaces comprises hyperbolas.
10 . The method of claim 1 , comprising:
drilling, using a drilling system, a borehole guided by the planned borehole trajectory.
11 . A non-transitory computer-readable memory comprising computer-executable instructions stored thereon that, when executed on a processor, cause the processor to perform steps comprising:
generating a simulated seismic training dataset, wherein the simulated seismic training dataset comprises upgoing signals and downgoing signals, each recorded on hydrophones and geophones;
training a deep learning network, using the simulated seismic training dataset, to separate the upgoing signals from the downgoing signals;
obtaining, from a seismic acquisition system comprising hydrophones and geophones, field hydrophone data and field geophone data from a subsurface region of interest, the field hydrophone data and the field geophone data comprising observed upgoing signals and observed downgoing signals;
wherein the observed downgoing signals are formed in part by reflected downgoing ghost signals;
separating, using the trained deep learning network, the observed upgoing signals from the observed downgoing signals to yield predicted upgoing signals with the reflected downgoing ghost signals removed;
generating a seismic image of the subsurface region of interest based, at least in part, on the predicted upgoing signals;
identifying, using a seismic interpretation workstation, a drilling target based, at least in part, on the seismic image, and
planning a planned borehole trajectory, using a borehole planning system, based, at least in part, on the drilling target.
12 . The non-transitory computer-readable memory of claim 11 , wherein generating the simulated seismic training dataset comprises:
creating, using a random number generator, a plurality of reflectivity surfaces; and
convolving a wavelet with the plurality of reflectivity surfaces.
13 . The non-transitory computer-readable memory of claim 11 , wherein the deep learning network is a U-net neural network.
14 . The non-transitory computer-readable memory of claim 11 , wherein generating the simulated seismic training dataset further comprises at least one selected from the group consisting of: creating coherent noise, random noise, and amplitude distortion.
15 . The non-transitory computer-readable memory of claim 11 , wherein the simulated seismic training dataset further comprises:
an input dataset, comprising a linear superposition of upgoing signals and downgoing signals; and
an output dataset comprising separate upgoing signals and downgoing signals.
16 . The non-transitory computer-readable memory of claim 11 , wherein the hydrophones and geophones are arranged in collocated pairs, each pair comprising one hydrophone and one geophone.
17 . A system, comprising:
a field hydrophone and a field geophone; and
a computer system, configured to:
generate a simulated seismic training dataset, wherein the simulated seismic training dataset comprises upgoing signals and downgoing signals, each recorded on hydrophones and geophones;
train a deep learning network, using the simulated seismic training dataset, to separate the upgoing signals from the downgoing signals;
obtain, from a seismic acquisition system comprising the hydrophone and geophone, field hydrophone data and field geophone data from a subsurface region of interest, the field hydrophone data and the field geophone data comprising observed upgoing signals and observed downgoing signals;
wherein the observed downgoing signals are formed in part by reflected downgoing ghost signals;
separate, using the trained deep learning network, the observed upgoing signals from the observed downgoing signals to yield predicted upgoing signals with the reflected downgoing ghost signals removed;
generate a seismic image of the subsurface region of interest based, at least in part, on the predicted upgoing signals;
identify, using a seismic interpretation workstation, a drilling target based, at least in part, on the seismic image, and
plan a planned borehole trajectory, using a borehole planning system, based, at least in part, on the drilling target.
18 . The system of claim 17 , wherein the deep learning network is a U-net neural network.