Identification of Karst caves in seismic data based on deep convolutional neural network

Abstract

Karst cave identification is significant for the exploration and development of fracture-cavity oil and gas reservoirs. Conventional identification methods are multi-solution and inefficient. Therefore, a deep learning method with strong feature learning and generalization capabilities is introduced into Karst cave identification. However, it is still a challenging task to identify Karst caves by deep learning due to the complex response characteristics of Karst caves to the seismic wavefield, the small sizes of anomalies, and the difficulties in obtaining training samples. Faced with this pro-blem, we propose a "two-step" deep learning me-thod for identifying Karst caves in seismic data. Spe-cifically, the U-Net model is used to identify the "bead-shaped" anomalous reflection on the seismic section. Then, according to the identification results of the "bead-shaped" anomalies, seismic data are cropped into small patches and input into the deep residual network to implement the prediction of the actual Karst cave profile. Considering the difficulties in obtaining training data for actual Karst cave prediction, we propose implementing wave equation forward modeling to generate seismic Karst cave data with accurate labels. The application of field seismic data shows that the me-thod is accurate in Karst cave identification, has strong noise resistance, and can greatly save the cost of manual interpretation.