Seismic facies identification using a U-KAN network with multi-scale spatial pyramid attention mechanism

Accurate identification of seismic facies plays a critical role in characterizing subsurface structures, locating hydrocarbon reservoirs, and guiding resource exploration and development. Traditional manual interpretation methods are highly subjective and notoriously inefficient. In recent years, deep learning-based techniques have emerged as powerful alternatives to address these shortcomings. The introduction of Kolmogorov–Arnold Networks (KAN) has provided new insights into interpreting conventional network architectures, facilitating the development of hybrid models such as U-KAN, which integrates convolutional operators with KAN. In this study, we apply U-KAN to seismic facies identification and further augment its performance by incorporating a Multi-Scale Spatial Pyramid Attention (MSPA) mechanism. The proposed MSPA-UKAN architecture leverages the superior nonlinear representation and interpretability of KAN, along with the efficient multi-scale feature extraction capabilities of MSPA. This combination allows the model to capture multi-scale seismic features more effectively and represent complex facies transitions accurately. To mitigate limited generalization caused by geological variability, we introduce a model-based transfer learning strategy in which a pre-trained model is adapted to datasets from new regions, thereby enhancing recognition accuracy. The MSPA-UKAN was first trained on a public dataset from the F3 block in the North Sea, Netherlands, and subsequently transferred to and evaluated on the Parihaka block in New Zealand, where it demonstrated excellent seismic facies recognition performance.