The noise removal, wave separation and de-convolution present an important issue in seismic processing. All those seismic processing tasks deal with an observed seismic signal which contains acquisition noises, different kind of wave fields, multiples and the goal is to improve the seismic imaging. Most of the standard processing techniques are based on frequency characteristics of observed signal and the algorithms are almost ‘data driven’. What we propose in this paper is a generalized probabilistic approach based on a geostatistical estimator: the kriging. The kriging provides a general mathematical model known as the best linear estimator. Why using kriging instead of standard processing methods? The answer consists in the fact that the kriging provides a good linear unbiased estimator but also an estimation error which enables to control the quality of the operator and its parameter choices. We illustrate this approach in the case of filtering of seismic gathers, separation of wave fields on vertical seismic profiles and finally an example of application of the de-convolution.