3D seismic geometry quality control and corrections by applying machine learning

Abstract

Seismic geometry quality control (QC) and corrections are crucial but labor-intensive steps in seismic data preprocessing. Current methods to estimate the correct positions of sources and receivers are usually based on the first-break traveltimes, which may contain large errors, thereby affecting the accuracy of the results. We have applied a deep convolutional neural network to identify shots and receivers that have position error, and we searched for the correct position. Once an error in position is identified by scanning data, a grid search for the correct location is conducted and the result is evaluated by the system until an optimal position is found. The network is trained on 3200 training sets from real data that have been corrected by the traditional method. Through cross validation on 800 sets, the classifier achieves a precision of 99.5% and a recall rate of 1. The final errors between the true positions and corrected positions are less than 10% of the shot spacing. An uncorrected real data experiment reveals that the proposed machine-learning method for geometry QC and correction provides similar results to the conventional manual correction approach but without human interference. Because the wavefield pattern of the training data for this purpose is global, there is no need to train the system again when applying the method to correct receiver position or process another data set. This claim is verified with different real data.