A fine-tuning workflow for automatic first-break picking with deep learning

TL;DR

This paper introduces a deep learning workflow using a fine-tuned U-Net with residual blocks for automatic first-break picking in seismic data, emphasizing the impact of weight initialization and dataset diversity on accuracy.

Contribution

It demonstrates that transfer learning with pretrained networks and careful dataset creation significantly improve automatic first-break picking performance.

Findings

Pretrained networks require fewer labeled samples for fine-tuning.

Using a general dataset can improve accuracy if carefully curated.

Velocity models from pretrained network picks are closer to expert results.

Abstract

First-break picking is an essential step in seismic data processing. First arrivals should be picked by an expert. This is a time-consuming procedure and subjective to a certain degree, leading to different results for different operators. In this study, we used a U-Net architecture with residual blocks to perform automatic first-break picking based on deep learning. Focusing on the effects of weight initialization on this process, we conduct this research by using the weights of a pretrained network that is used for object detection on the ImageNet dataset. The efficiency of the proposed method is tested on two real datasets. For both datasets, we pick manually the first breaks for less than 10% of the seismic shots. The pretrained network is fine-tuned on the picked shots and the rest of the shots are automatically picked by the neural network. It is shown that this strategy allows to reduce the size of the training set, requiring fine tuning with only a few picked shots per survey. Using random weights and more training epochs can lead to a lower training loss, but such a strategy leads to overfitting as the test error is higher than the one of the pretrained network. We also assess the possibility of using a general dataset by training a network with data from three different projects that are acquired with different equipment and at different locations. This study shows that if the general dataset is created carefully it can lead to more accurate first-break picking, otherwise the general dataset can decrease the accuracy. Focusing on near-surface geophysics, we perform traveltime tomography and compare the inverted velocity models based on different first-break picking methodologies. The results of the inversion show that the first breaks obtained by the pretrained network lead to a velocity model that is closer to the one obtained from the inversion of expert-picked first breaks.