Predictive models for first arrivals on seismic reflection data, Manitoba, New Brunswick, and Ontario / B. Patwa, P.L. St-Charles, G. Bellefleur, and B. Rousseau.: M183-2/8877E-PDF

"First arrivals are the primary waves picked and analyzed by seismologists to infer properties of the subsurface. Here we try to solve a problem in a small subsection of the seismic processing workflow: first break picking of seismic reflection data. We formulate this problem as an image segmentation task. Data is preprocessed, cleaned from outliers and extrapolated to make the training of deep learning models feasible. We use Fully Convolutional Networks (specifically UNets) to train initial models and explore their performance with losses, layer depths, and the number of classes. We propose to use residual connections to improve each UNet block and residual paths to solve the semantic gap between UNet encoder and decoder which improves the performance of the model. Adding spatial information as an extra channel helped increase the RMSE performance of the first break predictions. Other techniques like data augmentation, multitask loss, and normalization methods, were further explored to evaluate model improvement"--Abstract, page 1.

Permanent link to this Catalogue record:
publications.gc.ca/pub?id=9.910026&sl=0

• MARC XML format
• MARC HTML format

Request alternate formats