Simultaneous Automatic Picking and Manual Picking Refinement for First-Break

TL;DR

This work addresses robust first-break picking in microseismic data under outlier and label-noise conditions by introducing Simultaneous Picking and Refinement (SPR). SPR models the true first-break as a latent variable $\tilde{y}$ and jointly optimizes picking and refinement through a probabilistic framework combining a Laplace labeling prior $P(y|\tilde{y})$ and a Bernoulli prediction $P(\tilde{y}|x; W)$ with $\hat{y}=h(x; W)$. Training proceeds by alternating updates of $\tilde{y}$ and network parameters $W$, yielding two inference modes: $y^*$ for automatic first-break picking and $y^{**}$ for refining manual picks, enabling resilience to mislabeled data and outliers. Extensive experiments on Sudbury and Lalor datasets demonstrate improved accuracy and strong cross-site generalization, robustness to noisy signals, and effective noisy-label refinement, with the method shown to be architecture-agnostic. The approach provides a principled, flexible framework for robust automatic first-break picking suitable for integration with diverse deep-learning architectures in seismic data processing.

Abstract

First-break picking is a pivotal procedure in processing microseismic data for geophysics and resource exploration. Recent advancements in deep learning have catalyzed the evolution of automated methods for identifying first-break. Nevertheless, the complexity of seismic data acquisition and the requirement for detailed, expert-driven labeling often result in outliers and potential mislabeling within manually labeled datasets. These issues can negatively affect the training of neural networks, necessitating algorithms that handle outliers or mislabeled data effectively. We introduce the Simultaneous Picking and Refinement (SPR) algorithm, designed to handle datasets plagued by outlier samples or even noisy labels. Unlike conventional approaches that regard manual picks as ground truth, our method treats the true first-break as a latent variable within a probabilistic model that includes a first-break labeling prior. SPR aims to uncover this variable, enabling dynamic adjustments and improved accuracy across the dataset. This strategy mitigates the impact of outliers or inaccuracies in manual labels. Intra-site picking experiments and cross-site generalization experiments on publicly available data confirm our method's performance in identifying first-break and its generalization across different sites. Additionally, our investigations into noisy signals and labels underscore SPR's resilience to both types of noise and its capability to refine misaligned manual annotations. Moreover, the flexibility of SPR, not being limited to any single network architecture, enhances its adaptability across various deep learning-based picking methods. Focusing on learning from data that may contain outliers or partial inaccuracies, SPR provides a robust solution to some of the principal obstacles in automatic first-break picking.

Figures (10)

• Figure 1: (a) Amplitude signals. Manual picking are marked with red circles. (b) Binarized 2-D label, the position of the first-break is 1 and the rest is 0. (c) Indexed 1-D label, the index of the first-break among all the sampling points.
• Figure 2: The training process of SPR. The network parameters $W$ and potential first-break $\widetilde{y}$ are alternately updated to simultaneously learn picking as well as manually picking refinement.
• Figure 3: Comparison of picking results of different methods for trace 1005070-1005172 of dataset Sudbury. Manual picking and picking results of different methods are indicated by red circles and blue dots, respectively.
• Figure 4: Comparison of picking results of different methods for trace 2362428-2362573 of dataset Lalor. Manual picking and picking results of different methods are indicated by red circles and blue dots, respectively.
• Figure 5: Generalization experiments for first-break picking. The model is trained on data Sudbury and tested on data Lalor. The figure shows traces 298083-298257 in data Lalor. Manual picking and generalization results of different methods are indicated by red circles and blue dots, respectively.