Deep Learning for Efficient Prestack Seismic Internal Multiples Suppression with Global-Local Feature Fusion and Optuna Optimization

Abstract The suppression of internal multiples remains challenging in deep oil seismic exploration, as strong impedance contrasts in the shallow and intermediate layers generate multiples that obscure the weak reflections from the deeper hydrocarbon reservoir. Traditional methods, such as inverse scattering, have limited effectiveness in suppressing internal multiples. While deep learning has demonstrated potential in internal multiple suppression, its application is still limited by the availability of high-quality training labels, challenges in processing prestack seismic field data, and the requirement of extensive human intervention. To address these issues, we propose GLO-UNet, a hybrid deep learning framework based on U-Net that incorporates a novel Global-Local Feature Fusion (GLFusion) module at the bottleneck to combine Transformer-based global context with DenseNet-based local detail via channel attention. We further perform automated hyperparameter optimization via Optuna to reduce manual parameter tuning and training time. Experiments on synthetic seismic datasets (five-layer, Sigsbee2B) demonstrate that GLO-UNet achieves superior performance in multiple suppression and primary preservation, yielding higher signal-to-noise ratios than U-Net and G-UNet, as well as substantially lower computational time than adaptive virtual events (AVE) at inference. Results on field data from Tarim Basin, Northwest China, including prestack gathers and normal-moveout (NMO) corrected profiles, confirm that GLO-UNet effectively suppresses internal multiples while preserving key primary reflections. These results show the practical efficiency and robustness of our GLO-UNet for realistic seismic exploration scenarios.