Wavelet estimation in seismic convolved hidden Markov models

Inversion of seismic AVO-data is an important part of reservoir evaluation. These data are convolved but the convolution kernel and the associated errorvariances are largely unknown. We aim at estimating these model parameters without using calibration observations in wells. This constitutes the first step in socalled blind deconvolution. We solve the seismic inverse problem in a Bayesian setting and perform the associated model parameter estimation by an approximate marginal maximum likelihood method. A small test study indicate that bell-shaped wavelets with smooth edges are identified well, even for approximations of low orders.