Adaptive steering kernel regression for prestack seismogram denoising

• Gray distance is considered as another determinant of regression function weights.
• Kernel shape varies with characteristics of data samples in different regions.
• The edge information is preserved more effectively and completely.
• Seismic reflection events are recovered more completely and continuously accordingly.

Noise attenuation is a necessary and persistent problem during prestack seismic data processing. In this paper, we discuss a developed nonlinear method called adaptive steering kernel regression (ASKR) and apply it to seismic random noise attenuation. In classical kernel regression (KR), spatial distance is considered as the only determinant of regression function weights. A kernel with fixed shape is used for all samples, which results in severe distortion of edge information. In the discussed ASKR, the weights are estimated based upon another important determinant — gray distance. The shape of kernel varies with the characteristics of data samples in different regions. Thus, the edge information, which corresponds to reflection events in seismic records, is preserved more effectively and completely. Results on both synthetic records and real seismic data show its feasibility and effectiveness. Moreover, we verify its better performance than classical KR in the aspect of amplitude preservation and noise attenuation.