[Paper Review] Active learning for seismic processing parameterisation, with an application to first break picking

Alan Richardson|arXiv (Cornell University)|Jun 17, 2021

Machine Learning and Algorithms5 references4 citations

TL;DR

This paper proposes an active learning approach for seismic first break picking that reduces manual labeling effort by intelligently selecting the most informative shot gathers for human annotation. Using a query-by-committee strategy with multiple picking algorithms, it achieves comparable accuracy to random sampling with approximately half the number of labeled gathers, particularly focusing on regions with abrupt first break shifts like geological discontinuities.

ABSTRACT

Parameter values for seismic processing steps are often chosen on a regular grid of samples and interpolated. Active learning instead attempts to optimally select the samples on which parameter values are chosen. For parameters that do not vary smoothly, this often reduces the number of samples that need to be labelled in order to achieve a desired accuracy on the whole dataset. In regression tasks this is typically achieved using a query by committee strategy that selects the samples on which a committee of models is most uncertain. I implement such a strategy for the first break picking task, where the parameters to be chosen are the centre and width of the picking window for each trace. For the committee members I use the centre of the picking window and three popular picking algorithms. Applying this to a real dataset, and with samples corresponding to shot gathers, the active learning approach primarily selects gathers near a jump in the first breaks, and achieves similar levels of accuracy on the whole dataset with about half the number samples picked as when the samples are randomly selected.

Motivation & Objective

• To reduce the manual labor and inefficiency of regular grid or random sampling in seismic processing parameterization.
• To address the challenge of non-smooth parameter variation in seismic data, especially around abrupt changes like first break jumps.
• To improve accuracy in first break picking by prioritizing the most uncertain and informative shot gathers for human labeling.
• To integrate active learning into existing seismic workflows without requiring deep learning black boxes.
• To demonstrate that active learning can significantly reduce the number of required manual picks while maintaining high accuracy.

Proposed method

• The method uses a query-by-committee active learning strategy, where a committee of models (including the picking window center and three established picking algorithms) estimates uncertainty on unlabelled data.
• Uncertainty is quantified as the variance in predictions across the committee members, with the most uncertain samples selected for human labeling.
• The algorithm selects shot gathers for manual picking based on this uncertainty metric, focusing on areas where predictions diverge the most.
• The process iteratively labels the most informative gathers, updating the model and re-evaluating uncertainty until a desired accuracy is reached.
• A 10% subset of the dataset is considered at each step to balance computational cost and selection quality.
• The method replaces regular grid or random sampling with a targeted selection strategy that prioritizes regions with complex or abrupt first break behavior.

Experimental results

Research questions

• RQ1 Can active learning reduce the number of manually picked gathers required to achieve high accuracy in first break picking?
• RQ2 Does selecting gathers based on model uncertainty outperform random or regular grid sampling in seismic parameterization?
• RQ3 Where are the most informative gathers located in a seismic dataset, and does active learning naturally identify them?
• RQ4 How does the number of candidate gathers considered by the committee affect the final accuracy and efficiency?
• RQ5 Can active learning effectively detect and prioritize regions with abrupt first break shifts, such as those caused by near-surface heterogeneity?

Key findings

• The active learning approach achieved similar accuracy to random sampling but required only about half the number of manually picked gathers.
• The method primarily selected gathers near shot 200, where a jump in first break times occurs due to geological complexity.
• The mean absolute error in predicted picks decreased more rapidly with active learning compared to random selection, especially in early iterations.
• Considering 1% of the dataset in the committee selection process was sufficient for optimal performance, with no significant improvement beyond 10%.
• The active learning method successfully identified and prioritized regions of high uncertainty, such as areas with abrupt first break shifts, which are critical for accurate processing.
• The approach maintained moderate computational cost by limiting the number of candidate gathers considered at each step, making it practical for real-world seismic workflows.