Mamba-Driven and Feature-Fused U-Net for Automatic Seismic Horizon Interpretation

Seismic horizon interpretation is a crucial yet time-consuming task in seismic data processing and interpretation. As it can be formulated as a classification or segmentation task, numerous machine learning and deep learning methods have been developed for seismic horizon delineation. However, effectively extracting local and global seismic features remains a significant challenge. We propose the Mamba-driven and feature-fused U-Net (MDFF-U-Net) for seismic horizon interpretation in this study, employing U-Net as the backbone architecture. The model incorporates a Mamba-driven module to exploit sequential dependencies and capture intrinsic global relationships within seismic traces. Additionally, a feature-fusion module based on the Kolmogorov–Arnold Network (KAN) is integrated into the U-Net encoder-decoder framework to effectively combine local and global seismic features, enabling precise horizon segmentation. We validate the proposed approach on the F3 field dataset through comprehensive ablation studies, evaluating the effectiveness of each module. Comparative experiments against several state-of-the-art deep learning models further demonstrate the superior performance and robustness of MDFF-U-Net in seismic horizon interpretation.