Continuous denoising level adjustment of seismic data through filter modification

Abstract

Noise reduction is an important step in seismic data processing. Noise levels display a significant degree of spatial and temporal variability owing to complicated geologic conditions and acquisition environments, making noise reduction extremely challenging. Existing denoising models for seismic data based on deep learning are typically developed using a training set with a single constant noise level or with numerous discrete noise levels within a given range. These models cannot be applied for the recovery of seismic data with continuous noise levels. If the noise level of the seismic data to be processed does not match the denoising model, noise reduction is likely to be incomplete or signal details may be lost. Notably, the filters (convolution kernels) in the models trained by data sets with different noise levels are very similar in terms of visual patterns; only the statistics of their weights differ, such as the mean and variance. Based on this principle, we have designed a generic denoising network to artificially adjust the denoising level. The filter modification layer (FML) in this generic denoising network modifies the filter channel-by-channel. A continuous change in the denoising level between the beginning and final levels can be conducted by altering the FML, thereby preventing over- or underdenoising.