|
| | Abstract | | We propose an automatic method of detecting first arrivals of seismic signal using fully convolutional neural networks (fCNN). The method is based on the image-to-image translation approach: Our fCNN learn approximate function mapping from particular seismic attribute image, such as amplitude, to image which represent first break positon. Compared with conventional automated picking methods, it takes advantage from the global information of the entire shot record as well as the local information, because the method adopts "U-net" architecture and directly uses local information (time and offset) as input data. We apply our picking method and some conventional methods to a land seismic data acquired, Japan. Every thirtieth shot record in the dataset, which already manually picked, is used as supervised data. Our picking method marks 90% agreement with the manual picking result, which is 13 percent higher score than the result from the other method. |
|
|
|
| |
|