Automatic geologic fault identification from seismic data using 2.5D channel attention U-net

Abstract

Fault identification plays a vital role in geologic structure interpretation, reservoir characterization, and reservoir evaluation. The traditional process for fault identification is label intensive. Deep learning-based automatic fault detection from seismic data has attracted extensive attention in recent years. However, using a 2D neural network (i.e., a single seismic data slice) leads to the loss of correlation among seismic slices, and using a 3D neural network (i.e., a 3D seismic data cube) increases the workload of fault labeling and the difficulty of training. To overcome the drawbacks of 2D and 3D neural networks, we have introduced a channel attention mechanism in the original U-net architecture. The new model, which is referred to as the 2.5D channel attention U-net (2.5D CAU-net), takes four adjacent seismic slices as inputs instead of the entire 3D seismic data cube. The efficacy of 2.5D CAU-net for fault detection is demonstrated using publicly available synthetic seismic data sets and real seismic field data sets. We explored the influence of cropping the original seismic data into various sizes on fault identification. Experimental results find that the 2.5D CAU-net can effectively use the correlation information among adjacent seismic slices to improve fault detection performance. For different seismic data cropping strategies, results indicate that the more extensive size cropping method allows the network to obtain more contextual information and more complete fault features, improving the efficacy of fault picking. However, the cropping approach with a larger size leads to a smaller number of labeled examples. When the amount of training data is not large enough, the larger size cropping approach tends to lead to model overfitting, which will decrease the performance of the fault recognition, especially in a 3D neural network. Overall, 2.5D CAU-net can improve the efficiency of fault recognition with less annotation effort and computational resources.