Using geophysical seismic data for gas-bearing fluid prediction is crucial for understanding gas–water distribution in reservoirs, serving as a vital basis for gas layer development decisions and reserve assessments. Although data mining algorithms have become the most popular solution for seismic reservoir characterization, numerous limitations still exist, particularly under conditions of scarce offshore gas field logging data. These challenges primarily involve insufficient training labeled samples, class imbalance, and poor spatial continuity of prediction results. To address the above issues, we propose a seismic semi-supervised and imbalanced graph learning (S3I-GNN) workflow in this paper, which is applied for reservoir gas-bearing identification. First, multi-view seismic graph data was constructed using sensitive seismic attributes and elastic parameters as inputs, innovatively incorporating graph neural networks into seismic characterization research. The graph synthetic minority over-sampling technique module was employed to address the imbalanced gas-bearing labels in seismic graph data, mitigating model prediction bias at the data level. The graph sampling and aggregation within the semi-supervised paradigm effectively facilitated information propagation across the entire seismic graph, enabling prediction results to exhibit enhanced spatial continuity and geological consistency. Comprehensive experiments utilizing real logging and seismic data from the Yinggehai Basin in the South China Sea demonstrated that the S3I-GNN workflow exhibits optimal and most balanced performance, particularly showing significant superiority in gas–water distribution and lateral continuity. The proposed framework offers insights into seismic fluid interpretation and characterization tasks with limited well logging information and class imbalance.
Zhu et al. (Mon,) studied this question.