New Machine Learning Training Exercise: Create a seismic interpolator We have released a new Machine Learning exercise in which we describe how to create a seismic trace interpo...
New Machine Learning Training Exercise: Create a seismic interpolator
We have released a new Machine Learning exercise in which we describe how to create a seismic trace interpolator. The exercise is an example of a ‘Seismic Image Regression’’ workflow that is supported in OpendTect’s Machine Learning plugin.
In Seismic Image Regression, a deep learning model maps a seismic image (2D, 3D) to an image with continuous values of similar size. This is a supervised approach in which the model is trained on known examples of input data (in this case seismic with missing traces) and corresponding target data (seismic with full coverage). Application of the trained model delivers a new volume (3D seismic), or a new attribute to a line set (2D seismic).
In the exercise, we use OpendTect’s Attribute Engine to create an input cube with approx. 33% of the traces randomly blanked. From this volume we extract examples for training in a restricted area. The trained model is applied to the entire volume, whereby the area from which no examples are extracted acts as blind test area. In the exercise we train a 2D Unet, but you can equally well train a 3D Unet (a 2D model trains much faster than a 3D model but a 3D model is faster in the application phase).
This workflow can be applied to interpolate missing traces with relatively small gaps in the data. It has also proved to work well for infilling near-surface, V-shaped gaps in onshore data sets. Finally, we have successfully applied a model trained in this way to infill gaps in another data sets of a different vintage (onshore vs offshore, different bin-sizes, processing and geologic settings).
For more info, please see Machine Learning Workflow: Seismic Image Regression Unet Fill Seismic Traces. https://lnkd.in/dha7jfie
Author: Assia Lakhlifi
#machinelearning #OpendTect
