STABILIZATION OF NORMAL‐INCIDENCE SEISMOGRAM INVERSION REMOVING THE NOISE‐INDUCED BIAS*

A main problem in computing reflection coefficients from seismograms is the instability of the inversion procedure due to noise. This problem is attacked for two well‐known inversion schemes for normal‐incidence reflection seismograms. The crustal model consists of a stack of elastic, laterally homogeneous layers between two elastic half‐spaces. The first method, which directly computes the reflection coefficients from the seismogram is called “Dynamic Deconvolution”. The second method, here called “Inversion Filtering”, is a two‐stage procedure. The first stage is the construction of a causal filter by factorization of the spectral function via Levinson‐recursion. Filtering the seismogram is the second stage. The filtered seismogram is a good approximation for the reflection coefficients sequence (unless the coefficients are too large).

In the non‐linear terms of dynamic deconvolution and Levinson‐recursion the noise could play havoc with the computation. In order to stabilize the algorithms, the bias of these terms is estimated and removed. Additionally incorporated is a statistical test for the reflection coefficients in dynamic deconvolution and the partial correlation coefficients in Levinson‐recursion, which are set to zero if they are not significantly different from noise.

The result of stabilization is demonstrated on synthetic seismograms. For unit spike source pulse and white noise, dynamic deconvolution outperforms inversion filtering due to its exact nature and lesser computational burden. On the other hand, especially in the more realistic bandlimited case, inversion filtering has the great advantage that the second stage acts linearly on the seismogram, which allows the calculation of the effect of the inversion procedure on the wavelet shape and the noise spectrum.