Seismic data are often contaminated by random noise, which reduces the signal-to-noise ratio (SNR), obscures weak reflections, and complicates interpretation and inversion. Effective noise attenuation is essential for improving the quality and reliability of seismic analysis. The Time-Frequency Peak Filter (TFPF) is an effective method for suppressing random noise without requiring prior knowledge of wavelet phase or an explicit velocity model. However, conventional TFPF uses a fixed analysis-window length, creating a trade-off between noise suppression and signal preservation. A long window can attenuate noise effectively but may distort reflection amplitudes, weaken high-frequency components, and reduce the resolution of closely spaced events. A short window preserves local signal details more effectively but can leave substantial residual noise. Because these characteristics vary throughout a seismic record, a fixed window may not be suitable for all regions.



This study proposes an adaptive TFPF framework guided by the local Hurst exponent, which measures long-range dependence and temporal persistence. Coherent seismic reflections generally show greater continuity and persistence, whereas random noise tends to exhibit less predictable behavior. The local Hurst exponent is estimated using an optimized sliding-window scheme, allowing the method to track changes in data characteristics across the record. A transitional-zone correction is also applied to address ambiguity in areas where signal and noise coexist or where local estimates fall between clearly signal-dominated and noise-dominated values. This correction supports a smoother response as local data characteristics change.



Instead of classifying data through a rigid binary threshold, the proposed method combines the outputs of short- and long-window TFPF operations using locally weighted blending. The weights vary according to the local Hurst estimates. In regions with coherent reflections and higher persistence, the blend favors the short-window output to preserve reflection amplitudes, high-frequency content, and waveform details. In regions with more random behavior, it gives greater weight to the long-window output to strengthen noise attenuation. This gradual combination is designed to reduce abrupt transitions, blocky artifacts, and discontinuities between neighboring regions.



The method was evaluated using synthetic datasets and a two-dimensional field seismic record. Synthetic tests at different SNR levels assessed noise attenuation and signal preservation under varying contamination conditions. Results showed improved output SNR and effective suppression of random noise, while difference sections indicated limited leakage of the desired signal. Application to the field record demonstrated that the method reduced random noise while preserving the structural continuity, dip, and relative amplitudes of subsurface reflections. Mean amplitude spectra were also analyzed to assess changes in frequency content, indicating that essential broadband information was maintained.



Overall, local-Hurst-guided adaptive TFPF adjusts the contribution of short- and long-window filtering according to local data characteristics. The results suggest that this approach can improve the balance between random-noise attenuation and preservation of important seismic attributes, supporting subsequent interpretation and inversion. Its performance depends on reliable local Hurst estimation and appropriate analysis-window parameters. Nevertheless, the synthetic and field tests demonstrate its potential as a practical approach for seismic data denoising while retaining features needed for geological interpretation and further quantitative analysis.