Deep learning for pattern recognition in seismic reflection data

Pattern recognition plays an important role in analyzing seismic reflection data, which contains valuable information of the subsurface geological structures, and serves as a powerful method for hydrocarbon exploration. Conventional seismic pattern recognition methods commonly involve handcrafted seismic attributes or filters that do not apply to seismic data with complex structures. On the other hand, with new seismic acquisition techniques and equipment providing an increasing amount of data, conventional methods tend to be inefficient in processing large-scale and high-dimensional datasets. Over the past decade, the improvement of computer powers and software development has promoted deep learning as an efficient and effective tool for pattern recognition, which extracts features directly from data without relying on assumptions. This dissertation presents deep learning methods for pattern recognition in seismic reflection data from various perspectives. First, I introduce a semi-supervised learning framework for salt segmentation to alleviate the burden of preparing a large amount of labeled training data. The unsupervised consistency loss enforces the convolutional neural network (CNN) to extract essential information from labeled and unlabeled data, leading to more accurate segmentation results and better generalization ability on different datasets than the supervised learning baseline. Second, I formulate relative geologic time (RGT) estimation as a regression problem and design a U-shape CNN to solve this problem. The encoder-decoder architecture with skip connections results in accurate RGT predictions directly from seismic images. Although trained on a synthetic dataset, the network generalizes well to complex field data. Third, I propose an unsupervised learning method for seismic random noise attenuation. In the proposed method, a convolutional autoencoder is trained to reconstruct clean images from noisy seismic images without supervision from labeled data. The network training is constrained by local orthogonalization loss for better signal and noise separation. Next, I apply CNNs to reconstruct subsurface velocity models from common-image gathers (CIGs), which involves depth-to-depth mapping. The focuses or the flatness of seismic events in CIGs contain valuable information about the surface velocity model. Trained with synthetic dataset migrated using wrong velocity models, the network learns the relationship between the incorrect positioning of seismic energy in CIGs and the corresponding correct velocity update. In the next chapter, I explore the possibility of employing a different network architecture, Transformers, for velocity model building. In the proposed method, velocity models are directly estimated from raw recorded seismic reflection data using a variant of Vision Transformers specially tailored for FWI (FWIT), consisting of an encoder and a decoder. The encoder of FWIT learns to extract high-level information from input shot gathers, which is further analyzed by the decoder to estimate the velocities based on the attention mechanism. The ability to learn long-dependency and the flexibility of predicting variable-length output make Transformers a more suitable architecture for FWI than CNNs. Finally, I discuss and suggest possible future research topicsGeological Science