Preserving Signal: Automatic Rank Determination for Noise Suppression
Author: Stewart Trickett
Presented at SEG 2015
New Orleans, LA, USA
Summary: Matrix rank reduction filters such as Cadzow / SSA applied to constant-frequency slices have become popular for suppressing random noise on seismic data sets. A critical parameter for such methods is the matrix rank. A small rank gives strong noise suppression that may damage signal, and is best suited for noisy data and simple geological structures. A large rank gives weak noise suppression that preserves signal well, and is best suited for cleaner data and complex geological structures. Even within a single seismic survey, however, conditions change with time, space, and frequency, making a fixed rank inappropriate. Here I describe how the rank can be automatically determined for each matrix, allowing the filter to adapt to changing conditions. Examples are given on synthetic and real data. The result is an easy-to-use noise suppressor that finds a reasonable balance between signal preservation and noise removal throughout the section.