Comparative analysis of the application of 2D/3D segmentation models in the task of identifying seismic horizons

DOI: 10.24412/2076-6785-2022-8-36-39

Abstract
In this paper, two approaches to solving the problem of identifying seismic horizons by partially specified positions of horizons in a cube of seismic amplitudes are compared. The first approach is to train a neural network of the U-Net architecture for two-dimensional data on longitudinal and transverse sections of the cube. The second approach is to transform the standard U-Net architecture into a three-dimensional one and train it on 3D cubes of seismic data. Both approaches have been tested on publicly available data on the F3 Netherlands field. A three-dimensional model gives a more accurate forecast, and a two-dimensional model learns much faster and requires fewer computing resources.

Materials and methods
A comparative analysis of the quality of solving the problem of seismic horizons allocation for 2D and 3D architectures of the segmentation convolutional neural network U-Net is performed. The advantages and disadvantages of each of the approaches are determined. Practical recommendations for evaluating and training models on new data are given.

Keywords
seismic exploration, reflected wave method, geology, segmentation, computer vision, U-Net, seismic horizons, neural networks