UPNet: A Robust First Break Picking Deep Learning Network with Uncertainty Quantification

Core Concepts

UPNet is a novel uncertainty-based picking deep learning network that not only estimates the uncertainty of network output but also can filter the pickings with low confidence, achieving higher accuracy and robustness than deterministic DNN-based models in first break picking tasks.

Abstract

The content discusses a novel deep learning-based approach called UPNet for robust first break (FB) picking in seismic exploration.
Key highlights:

FB picking is a crucial step in determining subsurface velocity models and well placement, but current deep neural network (DNN)-based automatic picking methods still face challenges in ensuring robustness, especially in low signal-to-noise ratio (SNR) scenarios.
To address this, the authors introduce uncertainty quantification into the FB picking task and propose UPNet, which consists of a Bayesian segmentation network (BSN), a multi-information regression network (MIRN), and an uncertainty-based decision method (UDM).
BSN estimates a posterior distribution of the FB picking map, MIRN infers the accurate FB time based on the segmentation map sampled from the posterior distribution, and UDM decides the final FB picking based on the uncertainty of MIRN regression results.
Experiments on four open-source field datasets show that UPNet exhibits higher accuracy and robustness than deterministic DNN-based models, achieving state-of-the-art performance.
The measurement uncertainty provided by UPNet can also serve as a reference for human decision-making.

Stats

"Accurate FB picking can provide precise static correction results, significantly improving the performance of subsequent seismic data processing [1], e.g., velocity analysis, stratigraphic imaging, etc."
"The current inefficient manual FB picking cannot satisfy the requirement because of the increase in the density of seismic data acquisition."

Quotes

"To analyze the uncertainty of the FB pickings of the DL-based method, we propose a novel uncertainty-based picking network named UPNet, which includes a Bayesian segmentation network (BSN), a multi-information regression network (MIRN), and an uncertainty-based decision method (UDM)."
"Experiments on four open-source field datasets show that UPNet exhibits higher accuracy and robustness than deterministic DNN-based models, achieving state-of-the-art performance."

Key Insights Distilled From

UPNet

by Hongtao Wang... at arxiv.org 04-09-2024

https://arxiv.org/pdf/2305.13799.pdf

Deeper Inquiries

How can the uncertainty quantification techniques used in UPNet be extended to other seismic data processing tasks beyond first break picking

The uncertainty quantification techniques used in UPNet can be extended to other seismic data processing tasks by incorporating them into various stages of the data processing pipeline. For example, in tasks like seismic event detection or seismic phase identification, uncertainty quantification can help in assessing the reliability of the detected events or phases. By incorporating uncertainty estimates, the system can provide more reliable and trustworthy results to the end-users. Additionally, in tasks like seismic imaging or velocity model building, uncertainty quantification can help in understanding the confidence levels associated with the generated models. This information can be crucial in decision-making processes related to exploration or resource extraction.

What are the potential limitations of the uncertainty-based decision method in UPNet, and how could it be further improved

One potential limitation of the uncertainty-based decision method in UPNet could be the threshold selection for filtering unstable pickings. The threshold value for variance (tvar(p)) may need to be carefully chosen to balance between removing unreliable picks and retaining potentially valid ones. If the threshold is set too high, it may filter out valid picks, leading to a loss of important information. On the other hand, if the threshold is too low, it may not effectively remove unstable picks, impacting the overall robustness of the results. To improve this aspect, a more adaptive thresholding mechanism based on the specific characteristics of the data could be implemented. Additionally, incorporating feedback mechanisms to adjust the threshold dynamically based on the data distribution could enhance the performance of the uncertainty-based decision method.

How might the insights from UPNet's robust first break picking be applied to improve other types of signal detection and processing in geophysical or other scientific domains

The insights from UPNet's robust first break picking can be applied to improve other types of signal detection and processing in geophysical or other scientific domains by focusing on the principles of uncertainty quantification and robust decision-making. For example, in applications like medical imaging or remote sensing, where accurate detection of specific features is crucial, incorporating uncertainty quantification techniques can help in assessing the reliability of the detected features. By understanding the uncertainty associated with the detected signals, decision-makers can have more confidence in the results and make informed decisions. Additionally, the concept of preferring to miss rather than overuse, as demonstrated in UPNet, can be applied to various signal processing tasks to prioritize accuracy and reliability over quantity, ensuring high-quality results even in challenging conditions.