Training is done in two steps, considering only three available well-logs: gamma-ray, density and photoelectric effect. In the first step, we model each lithology by an ergodic and gaussian continous Hidden Markov Model. Labelled data is mandatory to train the HMM of each lithology; as core data from the Marcoule site is not available, we use labels resulting from a previous research work obtained with a Kohonen map on the Marcoule site. We use the corresponding well-log to train the HMMs of each lithology, after cutting the sequence of measures of the whole well into pieces corresponding to each lithology. Then, in a second step, training is pursued in a contextual way, that is on longer sequences of the same well in which several lithologies are present, and by concatenating HMMs of such lithologies. The Viterbi algorithm is used during both steps to reestimate the HMMs parameters.

Classification is performed in two ways on the other well, also using the Viterbi algorithm: first, in sub-sequences of a fixed short length; second, on the whole well using this way the Viterbi algorithm to perform segmentation of different lithologies. Results for the Marcoule site are presented.

Supplementary URL: http://