Full text loading...
-
Visual Identification of Noisy Seismic Records with Machine Learning
- Publisher: European Association of Geoscientists & Engineers
- Source: Conference Proceedings, EAGE 2020 Annual Conference & Exhibition Online, Dec 2020, Volume 2020, p.1 - 5
Abstract
Geophysicists are effortlessly able to see the noise on seismic data and use this ability to make algorithm selection and parameter decisions to best tackle the noise in their data. Here we train a convolutional neural network to assess the noise-level on seismic data directly. This is achieved by feeding thousands of human-labelled training examples to an algorithm that learns to map features in the data to the “noise-level” labels provided. The network can tirelessly look at a large dataset and predict the noise-level for every single record (a task too large for any human to accomplish in a reasonable timeframe). This will be helpful to assist humans in understanding their data, by identifying where the noisy records are, both before denoise for finding test cases and after denoise for assessing the global success of a denoise algorithm. Using machine learning to understand the noise content in data is a first step towards fully automated denoise. The training approach developed here may be adapted for use in other problems, e.g., the identification of multiple in seismic data.