Hydrocarbon-Bearing Information Mining of Prestack Seismic Gather Image Based on Prior-Guided Attention Mechanism

At present, there is still no effective data mining methods to purposely and robustly extract oil and gas information from the prestack seismic gather image (SGI). The traditional clustering methods fail to achieve this due to the lack of appropriate distance metric and difficulty of computational complexity. To solve this problem, we propose an “end-to-end” deep clustering method for hydrocarbon-bearing information mining, embedding a novel module which integrates the geoscientific prior information and attention mechanism (AM). We first construct a novel AM guided by the prior information of hydrocarbon regarding global position and horizontal gradient on SGI. Second, based on the clustering loss and reconstruction loss, we perform an autoencoder convolutional neural network that is suitable for SGI. Further, the prior-guided AM (PIAM) is to be embedded in the unsupervised deep clustering network, which is capable of selectively extracting deep features of oil and gas information. The theoretical and practical experiments show that the proposed workflow can robustly and targetedly achieve the accurate mining of oil and gas information for SGI.