A big data-driven deep-learning method for seismic wavefield separation of vertical seismic profile data

ABSTRACT Traditional vertical seismic profile (VSP) wavefield separation methods, such as transform domain methods and median filtering, struggle with spatial aliasing and the picking of complex wave events. Many deep-learning models trained on data sets from these traditional methods or synthetic data sets often face limitations in applicability and accuracy. Therefore, high-quality data sets are crucial for deep learning. We introduce an efficient iterative strategy designed to incrementally increase the diversity of data sets and their consistency with field data. This strategy uses a pretrained network to iteratively predict VSP data that were not included in its training. The process continues until clean up- and downgoing waves are extracted. Then, the obtained clean up- and downgoing waves are recombined to form new data sets. Subsequently, we continue training the pretrained network on the new data sets. This enables the final network to adapt to complex VSP data effectively. With this approach, data sets have been gradually upgraded from a simple convolutional seismic data model to more comprehensive wave equation-based data sets, ultimately including a massive amount of field data. To improve model generalization and address real production needs, we generate a large-scale data set totaling over one terabyte, enabling a truly big data-driven approach. Using these comprehensive data sets, we train a universal network for wavefield separation. Through tests with synthetic and field data not included in the training data sets, we show that the separation effect of the universal network is significantly superior to that of traditional and other deep-learning methods.