Related Experiment Video
Updated: Dec 16, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
890
Automated design of a convolutional neural network with multi-scale filters for cost-efficient seismic data
Zhi Geng1,2, Yanfei Wang3,4,5
1Key Laboratory of Petroleum Resources Research, Institute of Geology and Geophysics, Chinese Academy of Sciences, 100029, Beijing, P. R. China. gengzhi@mail.iggcas.ac.cn.
Nature Communications
|July 5, 2020
Summary
A new deep learning model, SeismicPatchNet, efficiently classifies seismic data for subsurface geology. This resource-saving CNN significantly speeds up the identification of Bottom Simulating Reflection (BSR), crucial for gas hydrate exploration.
Area of Science:
- Geoscience and Artificial Intelligence
- Deep Learning for Seismic Data Analysis
Background:
- Subsurface geologic feature identification relies on seismic data, often using conventional deep learning for image analysis.
- Complex seismic signals create high-dimensional, multi-scale data, making traditional Convolutional Neural Networks (CNNs) computationally expensive.
Purpose of the Study:
- To develop a highly efficient and resource-saving CNN architecture for seismic data classification.
- To improve the computational efficiency and accuracy of seismic image interpretation, particularly for Bottom Simulating Reflection (BSR) detection.
Main Methods:
- Proposed SeismicPatchNet, a novel CNN architecture incorporating topological modules and multi-scale-feature fusion units.
- Discovered the architecture using an automated data-driven search strategy.
- Evaluated performance against established models like VGG-16 and ResNet-50, including parameter storage and prediction speed.
Main Results:
- SeismicPatchNet exhibits a minimal parameter storage volume (~2.7 MB), significantly smaller than VGG-16.
- Achieved prediction speeds approximately 18 times faster than ResNet-50.
- Demonstrated superior performance in identifying Bottom Simulating Reflection (BSR), a key indicator for marine gas hydrate resources.
Conclusions:
- SeismicPatchNet offers a computationally efficient solution for seismic data interpretation.
- The architecture effectively captures key seismic features, as shown by saliency mapping.
- Presents a promising approach for end-to-end interpretation of diverse seismic datasets with minimal computational cost.
