Automatic interpretation of seismic horizon based on multi-attribute and Attention U-Net

Abstract Horizontal tracking constitutes a foundational and critical aspect of seismic data processing and interpretation. Traditional manual horizon tracking work is time-consuming and subjectively interferes with huge factors. In response, automated horizon tracking methods have been developed, albeit with the drawback of requiring extensive parameterization. Intelligent horizon tracking methods combined with deep learning are also a popular development direction in the future. Nonetheless, the efficacy of conventional convolutional neural network approaches, reliant solely on seismic amplitude data, is markedly limited in scenarios characterized by poor data quality, complex geological structures, and the prohibitive cost of acquiring extensive high-quality training labels. In addressing these challenges, we further optimize this work based on the current mainstream intelligent horizon tracking method, and propose a method based on seismic multi-attribute constraints and attention mechanism of U-Net neural network. Through the multi-attribute combined analysis, the two attributes with the most obvious features for horizon interpretation (instantaneous phase, principal component filtering) are selected to be combined with the amplitude data for seismic attribute constraints. Multi-attribute analysis furnishes extensive feature information crucial for interpreting subterranean structures and horizons. It enables the capture of multi-scale feature details throughout network training, facilitating enhanced visualization and identification of horizons. Moreover, a curve reconstruction method using waveform similarity is added to augment the training data and training labelled data to improve the generalization of the training model. The Attention U-Net network model designed in this work captures the target horizon feature information more accurately. The prediction results of multi-channel seismic attribute data with 1D input at the network end show that the horizon tracking results of this work are significantly better than those predicted by the traditional U-Net network using a single seismic amplitude data.