SeisRWKV: Multi-scale Feature Interaction with Linear Complexity for Seismic Neighboring-shot Interference Mitigation

In seismic data acquisition, the independent simultaneous source method has emerged as a preferable alternative to traditionaltechniques, as it can improve acquisition efficiency while preserving data fidelity. However, the simultaneous activation ofmultiple seismic sources within a narrow time window often introduces neighboring-shot interference, which degrades thequality of acquired data—while seismic data processing is essentially a type of specialized image data processing task, it isfar more complex than natural image processing due to the non-stationary characteristics of seismic signals and the intricategeological backgrounds they reflect. In addition, the latency in single-shot data synthesis of autonomous acquisition nodesposes a major obstacle to supplementary data collection, further limiting the flexibility of subsequent data processing workflows.Traditional denoising methods are incompetent in addressing such interference, mainly due to two critical drawbacks: the highsimilarity between interfering signals and primary seismic signals, and their excessive reliance on manually tuned empiricalparameters. On the other hand, although deep learning has achieved remarkable success in natural image denoising, currentdeep learning-based denoising methods still face practical challenges when applied to seismic data processing scenarios:the scarcity of high-quality noisy-clean labeled sample pairs, the limited receptive field of convolutional neural networks thathinders long-range feature modeling, and the high computational complexity of Transformer models that prevents real-timedata processing. To overcome these issues, this study develops a specialized deep learning framework for neighboring-shotinterference removal and introduces the SeisRWKV model as its core component. The model employs a co-wkv bidirectionalattention mechanism to conduct global feature modeling with linear complexity, which efficiently reduces computational costswhile ensuring comprehensive sequence representation. Furthermore, the incorporation of a Multi-Channel Fusion (MCF)module enhances the fusion of information across different feature channels, strengthens the model’s ability to capture multi-scale features, and enables accurate extraction of contextual information for targeted interference suppression. Experimentsdemonstrate that SeisRWKV can effectively eliminate neighboring-shot interference, significantly improving the signal-to-noiseratio of seismic data, with performance surpassing other methods.