First-break picking is an important step for correcting long-wavelength statics anomalies from the low-velocity weathering layer. In this context, applications of deep learning in onset detection greatly improve the extraction efficiency compared to traditional methods. In deep learning algorithms, while convolutional neural networks (CNNs) are excellent for image classification tasks, typically involves a fixed-length input and a huge amount of training data for a low-resolution output. Instead, fully convolutional networks (FCNs) exceed in semantic segmentation, they handle variable-length data and performs pixel-wise classification. Consequently, the training volume is substantially reduced for a high-resolution result. Although the segmentation outperforms in image recognition, it lags in location accuracy. Therefore, the post-processing of the segmented image for finer delineation is our approach to tackle the issue. We have implemented a velocity-constrained pooling based on U-net, an FCN encoder-decoder variant, that enhances the precision of predicted pixels before the first break extraction, and an adaptive dynamic sampling to reinforce the first break prediction, as an added value. Additionally, the network performance is evaluated with sensitivity and specificity analysis. Finally, the application over a land dataset shows promising results.

Presentation Date: Tuesday, October 13, 2020

Session Start Time: 8:30 AM

Presentation Time: 8:55 AM

Location: 351F

Presentation Type: Oral

This content is only available via PDF.
You can access this article if you purchase or spend a download.