Learning to label seismic structures with deconvolution networks and weak labels

Abstract

Recently, there has been increasing interest in using deep learning techniques for various seismic interpretation tasks. However, unlike shallow machine learning models, deep learning models are often far more complex and can have hundreds of millions of free parameters. This not only means that large amounts of computational resources are needed to train these models, but more critically, they require vast amounts of labeled training data as well. In this work, we show how automatically-generated weak labels can be effectively used to overcome this problem and train powerful deep learning models for labeling seismic structures in large seismic volumes. To achieve this, we automatically generate thousands of weak labels and use them to train a deconvolutional network for labeling fault, salt dome, and chaotic regions within the Netherlands F3 block. Furthermore, we show how modifying the loss function to take into account the weak training labels helps reduce false positives in the labeling results. The benefit of this work is that it enables the effective training and deployment of deep learning models to various seismic interpretation tasks without requiring any manual labeling effort. We show excellent results on the Netherlands F3 block, and show how our model outperforms other baseline models.