Physics-Guided Self-Supervised Suppression of Near-Surface Scattering in Land Seismic Data

In land seismic data, strong coherent scattering noise generated by small-scale, low-velocity near-surface heterogeneities can severely degrade reflection continuity and waveform fidelity. Unlike primary reflections, which remain approximately locally planar, such scattering noise violates the local plane-wave assumption and is characterized by irregular wavefronts and rapidly varying apparent slopes. Motivated by this physical distinction, we propose a physics-guided, self-supervised denoising framework that jointly reconstructs seismic data and predicts local slopes, enabling effective separation of scattering noise from reflection signals without requiring clean training data. The proposed framework jointly trains two neural networks: an implicit neural representation (INR) network for local slope estimation and a convolutional neural network (CNN) for seismic data reconstruction. A partial differential equation (PDE)-based loss is introduced to constrain the reconstructed data to satisfy the local plane-wave equation, while the local slopes are adaptively updated to minimize signal leakage. To improve scalability and efficiency, the INR is conditioned using shot-dependent information, while the CNN incorporates slope-based modulation derived from the estimated local slopes. This design allows the framework to learn from multiple shot gathers simultaneously while still generating shot-specific predictions, thereby avoiding redundant per-shot training. Synthetic and field data examples demonstrate that the proposed method effectively attenuates coherent scattering noise, enhances reflection continuity, and provides robust local slope estimation.