3D narrow-azimuth streamer survey

The survey is a narrow-azimuth marine streamer acquisition with 8 receiver cables of length 5.5 km. There are 5186 shots in total with 20% utilised per iterations.

The raw shot gathers that are input into the process have all free-surface effects retained. No pre-processing is applied to the traces beyond bandpass filtering for progressively widening the frequency range through blocks of iterations. 

With the limited cable length, the model updates need to penetrate below the reach of the diving waves for broadband velocity recovery at the target zone.

Single receiver cable from survey

Filtered 3-20Hz

Frequency Sweep

The panels show 6 stages of the inversion - predicted traces (left) and field recordings they are being matched against (right). The field recordings are being fed in at widening bandwidths with the lowest frequencies having lowest signal to noise. The initial match is very poor and the trace fit accuracy increases as the iterations proceed.

Start

4Hz AWI

5Hz AWI

7Hz AWI

11Hz AWI

22Hz FWI

XWI^TM model evolution

Generating long length-scale corrections from multiple starting model initialisations. Here we show the result of applying reflection-driven bulk changes from two different starting positions followed by AWI to converge towards the optimal minimiser in both cases. 

Scenario 1

Scenario 2

Iter 0 model

Iter 20 update (RWI)

Iter 0 model

Iter 20 update (RWI)

Iter 30 update (AWI)

Iter 50 update (AWI)

Iter 30 update (AWI)

Iter 50 update (AWI)

Scenario 1

Scenario 2

Iter 0 model

Iter 20 model (RWI)

Iter 0 model

Iter 20 model (RWI)

Iter 30 model (AWI)

Iter 50 model (AWI)

Iter 30 model (AWI)

Iter 50 model (AWI)

XWI no RWI. XWI no AWI.

Local Minima Trap. 

Pixel values at borehole location are shown in red and compared to the sonic log. At a depth of 2.2km the estimated velocity shifts in the wrong direction away from ground truth.

Start

FWI

RWI then AWI.

Global Minimum Misfit Minimiser.

In two separate cases, the background values of velocity is updated towards the basin of attraction of loss function minimiser. 

Scenario I

Scenario II

Start

Start

RWI+AWI

RWI+AWI

Final XWI

XWI^TM model evolution

Random initialisations are made possible using RWI+AWI converging towards the true answer without succumbing to the known limitations of FWI alone. FWI alone leads to a spurious result with an erroneous low velocity band seen to cut across model. The entire chain is performed with zero human intervention and the data remains on the cloud throughout. 

RWI+AWI+FWI

FWI alone