Core Concepts
A novel deep graph learning-based approach called DGL-FB that leverages higher-dimensional seismic data to achieve superior accuracy and stability in first break picking compared to 2D benchmark methods.
Abstract
The paper proposes a deep graph learning-based framework called DGL-FB for seismic first break (FB) picking. The key aspects of the method are:
Graph Data Preparation:
Each seismic trace is represented as a node in a large graph.
Edges are established between nodes based on the similarity of their midpoints, connecting traces with high correlation.
To manage the large graph size, a subgraph sampling technique is developed to streamline model training and inference.
Deep Graph Encoder:
A deep graph encoder is designed using the GraphSAGE method to encode the subgraphs into global features.
The encoder leverages a combination of SAGEConv layers with LSTM aggregators and TanH activation to capture the comprehensive information in the subgraphs.
ResUNet-1D Picking Header:
The global features from the graph encoder are concatenated with the local trace signals to form the combined features.
A 1D ResUNet architecture is employed to segment the combined features and detect the first breaks.
A weighted binary cross-entropy loss function is used to supervise the output segmentation, balancing the main loss and auxiliary loss.
The proposed DGL-FB framework is evaluated on a field dataset and compared to a 2D U-Net-based benchmark method. The results demonstrate that DGL-FB achieves a 7.2% higher accuracy and a 99.3% reduction in root mean square error (RMSE) compared to the benchmark, showcasing superior accuracy and stability in first break picking.
Stats
The research is supported by the National Natural Science Foundation of China under grant 12371512.
Quotes
"To model seismic data in higher dimensions, we introduce a novel modeling approach to analyze seismic data on a graph. Each seismic trace is treated as a node, and edges are established between nodes based on the near midpoints of corresponding sources and receivers."
"Consequently, graph analysis provides a more generalized methodology. The application of graph neural network theory in geophysics has gained traction."
"To address this issue, we propose a novel picking framework using the deep graph learning technique. Specifically, we first build a huge graph for the whole survey. Second, to boost the training and inference processes, the subgraphs are sampled from the huge graph. Third, a graph neural network is built to encode the global information of the subgraph. Finally, the combined information of global information and local information is fed to a ResUNet to obtain the FB."