Automated Seismic Velocity Picking via Deep Semantic Segmentation
عنوان دوره: ششمین همایش ژئوفیزیک اکتشافی نفت
کد مقاله : 2136-NIGS
نویسندگان
1کارشناس تحقیق و توسعه
2Reservoir Characterization Lead
3Research & Development Manager
چکیده
We developed a deep neural network model to automatically implement velocity analysis from semblance images, eliminating the need for extensive manual picking. Our method treats velocity picking as an image segmentation task on input semblance images. We train a U-Net convolutional neural network architecture using over 2000 common depth point (CDP) gathers and corresponding picked velocity profiles to segment the semblance images into distinct velocity regions. We optimize the model using techniques like sequence learning and customized loss functions. When evaluated on test CDP gathers excluded from training, the model achieved 99.3% accuracy in delineating the major velocity boundaries. This demonstrates the capability for high-quality automated velocity picking directly from seismic images using deep learning.
کلیدواژه ها
Title
Automated Seismic Velocity Picking via Deep Semantic Segmentation
Authors
Mohammad Ghasem Fakhari, Mahdi Saadat, Ehsan Salehi
Abstract
We developed a deep neural network model to automatically implement velocity analysis from semblance images, eliminating the need for extensive manual picking. Our method treats velocity picking as an image segmentation task on input semblance images. We train a U-Net convolutional neural network architecture using over 2000 common depth point (CDP) gathers and corresponding picked velocity profiles to segment the semblance images into distinct velocity regions. We optimize the model using techniques like sequence learning and customized loss functions. When evaluated on test CDP gathers excluded from training, the model achieved 99.3% accuracy in delineating the major velocity boundaries. This demonstrates the capability for high-quality automated velocity picking directly from seismic images using deep learning.
Keywords
velocity analysis, Deep learning, U-Net, Encoder-Decoder