Neural Network Seismic Imaging with Atrous Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Seismic imaging in complex subsurface geology faces challenges such as poor illumination, noise, and scale variation, leading to poorly imaged regions and inefficient feature encoding in hydrocarbon reservoir identification.
Innovation Solution
The use of a neural network with atrous convolution and varying rectangular atrous rates, concatenated and convolved with 3x3x3 kernels, along with an encoder-decoder architecture in three dimensions, trained on separate seismic images with different frequencies and grid sizes, and applied with different down-sampling rates to enhance seismic images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional seismic imaging methods are used, then the imaging process is simple and fast, but the image quality is poor in complex subsurface geology with regions of poor illumination and noise
Solution Approach 1:
A trained neural network is introduced as an intermediary between the raw seismic data and the final seismic image. The neural network processes the input seismic data through multiple layers including convolutional layers, atrous convolution layers, and skip connections to enhance image quality, reduce noise, and improve illumination in complex subsurface geology while maintaining computational efficiency
Solution Approach 2:
The imaging system transitions from static traditional methods to a dynamic neural network-based approach. The neural network adaptively adjusts its processing based on the input data characteristics, using variable atrous rates and skip connections to dynamically handle different geological complexities and improve image quality across varying subsurface conditions
2Adaptability or versatility
If conventional imaging methods are used, then the processing is computationally efficient, but the ability to handle multi-scale features and noise is insufficient
Solution Approach 1:
The neural network architecture is segmented into multiple specialized layers including convolutional layers for feature extraction, atrous convolution layers for multi-scale feature capture, and skip connections for preserving spatial information. Each layer segment performs a specific function to collectively handle multi-scale features and noise effectively
Solution Approach 2:
The system incorporates atrous convolution with variable atrous rates that operate at different spatial scales, effectively adding a dimensional aspect to feature extraction. This allows the network to capture features at multiple scales simultaneously, enhancing the ability to handle complex geological structures with varying dimensions
3Measurement precision
If standard convolutional layers are used, then the computational load is manageable, but the encoding of features at different scales is inefficient
Solution Approach 1:
The neural network applies partial convolution operations through atrous convolution layers with selective atrous rates, processing only the necessary spatial frequencies and scales required for accurate feature encoding. This partial action approach maintains computational efficiency while achieving superior multi-scale feature representation compared to full convolutional processing
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method is described for seismic imaging including image enhancement using a trained neural network. The neural network may receive training pairs of low signal-to-noise ratio 3D seismic images and high signal-to-noise ratio 3D seismic images; train a neural network on the training pairs wherein the training uses atrous convolution; receive a seismic image representative of a subsurface volume of interest; apply the neural network to the seismic image to generate a second seismic image; and display the second seismic image on a graphical user interface. The method is executed by a computer system.