A Recurrent Deep Learning Framework for Seismic Wavefield Separation in Ocean-Bottom Node Surveys

Accurate separation of upgoing and downgoing seismic wavefields is essential for improving subsurface imaging in ocean-bottom node (OBN) surveys, where recorded wavefields may overlap and complicate seismic data interpretation. This study proposes a deep learning framework based on stacked long short-term memory (LSTM) networks augmented with batch normalization (BN) and dropout regularization for the automated, data-driven of reflected P-wave wavefields recorded by OBseparation N sensors. The network is trained on a synthetic dataset generated from a realistic 1-D velocity model, with supervised labels derived from one-way wave-equation modelling. Training converges at epoch 38 with a best validation mean squared error (MSE) of 0.049293, corresponding to an R² of 0.9508. Quantitative evaluation demonstrates consistent superiority over conventional f-k filtering and competing deep learning baselines including multi-layer perceptrons (MLP), one-dimensional convolutional neural networks (CNN-1D), and gated recurrent unit (GRU) networks. An ablation study confirms the additive contribution of each architectural component. Three-dimensional visualization of predicted reflected P-wave amplitudes reveals physically coherent wavefield geometry consistent with theoretical expectations. The proposed methodology offers a computationally efficient and geologically interpretable alternative to established separation workflows, with direct applicability to real-world OBN processing pipelines.