Building a global seismic texture interpretation network

The primary goal of seismic interpretation is to understand seismic signals, categorize them into various patterns, connect each pattern with a specific depositional event, and finally reconstruct the geologic history. Therefore, the emerging machine learning techniques, particularly the convolutional neural networks, appear most suitable for tackling the problem of annotating various patterns existing in a seismic dataset.. Therefore, implementing deep neural networks into 3D seismic interpretation is the current research focus in the community. However, most of the work is performed for certain interpretation challenges in certain study areas, and thus the neural network building and training need to be repeated. Meanwhile, we identify great potentials for building a global network for identifying and interpreting the important seismic patterns in any given seismic dataset.

Scope: One model, all seismic patterns in all seismic datasets

Here, we initialize it by building both a dataset (StData-12) and a network (StNet). We hope it could be supportive for machine learning-assisted seismic interpretation. For example, the StData-12 could be used for testing the performance of new ML algorithms on seismic interpretation, whereas the StNet could be applied for quick seismic interpretation. Apparently, it is a long-term project and requires collaborations between geologists, geophysicists, and seismic interpreters.so that the StData-12 and SpiNet can be further upgraded to be more comprehensive and powerful.

StData-12

We are targeting a comprehensive dataset that covers all the seismic patterns. However, the subsurface geology is complicated, and it is difficult to build such a dataset alone. Here, StData-12 is built by utilizing the open-source seismic datasets available for us, including

(a) the F3 block over the Netherlands North Sea

(b) the Great South Basin in New Zealand

In the current StData-12 (Di et al., 2019), a total of 12 seismic patterns are tentatively categorized from them, including 7 types of horizons (flat, dipping, deformed, weak, strong, and chaotic), 2 types of stratigraphic sequences (regression and transgression), and 3 types of structures (fault, saltbody, and gas chimney).

StNet

The current StNet (Di et al., 2019) is a deconvolutional neural network, which is is capable of automatically recognizing and annotating all the seismic patterns defined in StData-12 in real time. It starts with 3 convolutional blocks, which contain 3, 2, and 1 convolutional layers, respectively. Correspondingly, 3 deconvolutional blocks are placed in the end for size recovery to ensure that the annotations are provided at the correct locations in the output image. In the middle is a 1x1 block of 2 layers and 1024 features for connecting the convolutional and deconvolutional blocks.

Your support is needed

Due to our limited dataset resources and geoscientific knowledge, our annotation may be not comprehensive enough for covering all important seismic features or undesirably mixture a few features into one pattern. Correspondingly, the current architecture of the StNet may be considered not efficient and deep enough particularly when the training dataset StData-12 is increasing and/or more seismic patterns are added into the target list in the future. Therefore, this is a long-term project, and collaboration is greatly in need. Therefore, your support could be, but not limited to

1. redefining the seismic patterns in the StData-12

2. expanding the StData-12 size

3. redesigning and fine-tuning the StNet network

Copyright

For the convenience of application, both the StData-12 and StNet is open-source under no license and available at Here. Please consider acknowledging the contributors listed below.

Contributors:

1. Di, H., D. Gao, and G. AlRegib, 2019, Developing a seismic texture analysis neural network for machine-aided seismic pattern recognition and classification: Geophysical Journal International, 218, 1262-1275 [Link]