Abstract
Identifying geological structures from seismic images is critical in understanding the Earth's evolutionary history, assessing geological disasters, and predicting resource distribution. Researchers have used synthetic seismic data with structure labels generated by numerical simulation to train deep learning models to avoid the extensive effort and resources required to manually label seismic data. However, most existing studies are task-oriented and often simplify or even ignore some common and important geological features during the simulation process, which limits the performance of the trained model on real datasets. To further improve the generality and flexibility of deep learning models for structure identification on datasets with complex structural features, we propose an improved parametric 3D structure modeling framework for numerical simulation of synthetic seismic data that include dip, fold, unconformity, channel, cave, salt body, and composite fault structure. Moreover, we revise some simulation solutions to make the simulated geological structures better resemble the real ones. To the best of our knowledge, our synthetic seismic dataset has the richest structural features to date. The results of comparative experiments on multiple synthetic and field datasets reveal that a deep learning model trained on datasets with limited geological features is prone to learn biased representations, which causes the model to misclassify unseen geological structures with similar seismic responses to the target structure. On the contrary, the same deep learning model trained on the proposed dataset provides significantly reduced false predictions of structures with similar seismic responses and improves the identification accuracy of datasets with complex geological features. Although this study only takes fault and channel identification as case studies, the proposed synthetic seismic dataset can be easily adapted to identify other seismic structures.