Synchrosqueezing Voices Through Deep Neural Networks for Horizon Interpretation

Abstract

Horizon picking stands as a crucial element in reservoir characterisation, yet it remains a labour-intensive process. The manual interpretation of horizons across thousands of vertical seismic slices in a 3D seismic survey significantly further amplifies the time and effort invested in this task. While several automatic methods have been developed for extracting horizons in seismic images, their effectiveness can be compromised in the presence of interruptions in lateral continuity, such as faults and noise. Additionally, closely spaced horizons pose a challenge, making it even more difficult to accurately depict their exact locations. For tracking the horizon surfaces through a 3D seismic volume, it is necessary to exploit other seismic attributes extracted from the 3D seismic data. We proposed to use spectral voice components together with the original seismic amplitudes to track target horizon surfaces. We generated the time-frequency spectrum using a high-resolution method namely the synchrosqueezing wavelet transform (SWT) method, and the real part of the complex SWT spectrum is the voice component. We imported the spectral voice component and the seismic amplitude into a neural network. A framework of deep convolutional neural network (dCNN) was adopted for tracking horizon surfaces within a 3D seismic volume. We demonstrated this application on a field seismic dataset where closely spaced thin layers are within a complex faulted formation with noisy and low signal to noise ratio seismic data. The integration of amplitude and phase within the voice component attribute demonstrates its efficacy in enhancing the quality of the generated horizons, particularly when compared to using only seismic amplitude for this task. A field data example from the F3 dataset showcases the capability of our method in accurately delineating horizons across fault surfaces and in close proximity to unconformities. This surpasses the current limitations of existing horizon-picking methods.