Automatic Fault Interpretation Based on Semantic Segmentation of Deep Learning method

Faults identification is an important part in seismic data interpretation, which will play a significant guiding role in later crustal development structural interpretation and the research of oil and gas exploitation and distribution. With the development of seismic exploration technology, the amount of seismic data becomes larger and larger, and the requirement of fault interpretation accuracy is also getting higher and higher. However, conventional faults interpretation methods, such as manual faults picking or seismic attribute analysis, have more or less the disadvantages of low efficiency, long time-consuming and low accuracy. In recent years, with the rise of AI, the deep-learning methods have been applied in many research and aspects, and it also has bring some innovative ideas to seismology. In seismic interpretation, we can let computers to learn faults recognition patterns instead of interpreters for automatic fault interpretation, which will save time and improve efficiency and also keep high accuracy if appropriate data are given. Therefore, we propose a deep-learning method based on semantic segmentation in image recognition to realize automatic interpretation of seismic faults. Semantic segmentation can perform intensive segmentation tasks for seismic images, and each pixel of seismic images can be classified to a specified category (fault or non-fault). We adopt the deeplabv2 method and modify the model on the basis of ResNet network to realize the semantics segmentation of seismic image. Compared with the conventional convolutional neural network, which takes a small number of pixels around faults points as samples, this method uses the whole seismic image as training samples to learn the global characteristics of seismic amplitude image, and achieves the faults detection of each pixel of seismic image. We also use the Atrous Spatial Pyramid Pooling method in the deep learning network to obtain a better segmentation effect and the CRFs in the back-end to refine the results. In order to get the feature pattern in actual situation of seismic faults, we utilize a part of the real seismic data as training samples. After obtaining the optimal model, we apply it to the remaining data for fault automatic interpretation, and can get a better result.