Automated design of a convolutional neural network with multi-scale filters for cost-efficient seismic data classification

Zhi Geng,Yanfei Wang

doi:10.1038/s41467-020-17123-6

Abstract

Geoscientists mainly identify subsurface geologic features using exploration-derived seismic data. Classification or segmentation of 2D/3D seismic images commonly relies on conventional deep learning methods for image recognition. However, complex reflections of seismic waves tend to form high-dimensional and multi-scale signals, making traditional convolutional neural networks (CNNs) computationally costly. Here we propose a highly efficient and resource-saving CNN architecture (SeismicPatchNet) with topological modules and multi-scale-feature fusion units for classifying seismic data, which was discovered by an automated data-driven search strategy. The storage volume of the architecture parameters (0.73 M) is only ~2.7 MB, ~0.5% of the well-known VGG-16 architecture. SeismicPatchNet predicts nearly 18 times faster than ResNet-50 and shows an overwhelming advantage in identifying Bottom Simulating Reflection (BSR), an indicator of marine gas-hydrate resources. Saliency mapping demonstrated that our architecture captured key features well. These results suggest the prospect of end-to-end interpretation of multiple seismic datasets at extremely low computational cost.

Highlights

Geoscientists mainly identify subsurface geologic features using exploration-derived seismic data
All morphology patterns in such images can be properly learned by artificial neural networks (ANNs), many of which are based on the popular convolutional neural networks (CNNs) designed for image-related tasks in computer vision
We hypothesized that key signal features embedded in exploration seismic data could be captured by an ANN with significantly fewer parameters than classic CNN architectures

Summary

Results

Receiver operator characteristic (ROC) curves (Fig. 3a) were used to assess the comprehensive performance of SeismicPatchNet and five other classic CNN architectures using the synthetic test dataset. SeismicPatchNet clearly achieved higher precision for realworld BSR prediction than the other CNN architectures (Fig. 4a), though its advantage was only 0.33% greater than the lowest performer (ResNet-50, as in the first test). A complete section of the 3D seismic data (inline 88, Fig. 6a) was chosen to test the predictive performance of the CNN architectures for characterizing the subsurface distribution of BSR (Fig. 6b–s). This case study demonstrated the robustness of SeismicPatchNet in this application

Discussion

Methods

Code availability