Robust Ground Roll Noise Suppression Based on Dictionary Learning and Bandpass Filtering

Abstract

Ground roll noise is a common type of seismic noise arising from land acquisition. Suppressing ground roll noise has long been a challenging problem due to the strong coupling between ground rolls and reflection signals. The easiest way to suppress ground roll noise is by a bandpass filter, which removes the low-frequency ground roll and retains the high-frequency signals. Due to the coupling issue, the bandpass filter, however, is prone to cause the removal of low-frequency signals and the existence of high-frequency noise. To combat this problem, we propose a novel dictionary learning (DL) method for decoupling the ground rolls and reflection signals in the low-frequency band. We first filter the seismic shot gather using a strong low-pass filter, by which we remove all ground rolls. Then, we apply the DL method to retrieve the leaked low-frequency reflection signals. The dictionary atoms are obtained from the high-frequency reflection signals and then used to code the low-frequency mixture between ground rolls and signals. As a result, the reflection signals in the low-frequency part are easily separated by the sparse coding process and are added back to the high-frequency signals. The final output of the proposed algorithm is the summation between the high-frequency reflections and the retrieved low-frequency reflections. We apply the proposed method to both synthetic and real shot gathers containing ground roll noise and demonstrate its promising results.