Abstract
<strong class="journal-contentHeaderColor">Abstract.</strong> Deep learning has been widely used for various kinds of data mining tasks but not much for seismic stratigraphic interpretation due to the lack of labeled training datasets. We present a workï¬ow to automatically generate numerous synthetic training datasets and take the seismic clinoform delineation as an example to demonstrate the effectiveness of using the synthetic datasets for training. In this workï¬ow, we ï¬rst perform stochastic stratigraphic forward modeling to generate numerous stratigraphic models of clinoform layers and corresponding porosity properties by randomly but properly choosing initial topographies, sea level curves, and thermal subsidence curves. We then convert the simulated stratigraphic models into impedance models by using the velocity-porosity relationship. We further simulate synthetic seismic data by convolving reï¬ectivity models (converted from impedance models) with Ricker wavelets (with various peak frequencies) and adding real noise extracted from ï¬eld seismic data. In this way, we automatically generate a total of 3000 diverse synthetic seismic data and the corresponding stratigraphic labels such as relative geologic time models and facies of clinoforms, which are all made publicly available. We use these synthetic datasets to train a modiï¬ed encoder-decoder deep neural network for clinoform delineation in seismic data. Within the network, we apply a preconditioning process of structure-oriented smoothing to the feature maps of the decoder neural layers, which is helpful to avoid generating holes or outliers in the ï¬nal output of clinoform delineation. Multiple 2D and 3D synthetic and ï¬eld examples demonstrate that the network, trained with only synthetic datasets, works well to delineate clinoforms in seismic data with high accuracy and efï¬ciency. Our workï¬ow can be easily extended for other seismic stratigraphic interpretation tasks such as sequence boundary identiï¬cation, synchronous horizon extraction, shoreline trajectory identiï¬cation and so on.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have