Modeling And Classification Of Lithofacies Using The Continuous Wavelet Transform And Neural Network: A Case Study From The Berkine Basin (algeria)

We use a neural statistical method combined with the multiscale wavelet analysis for lithofacies classification from well-logs data. This approach has been applied to constrain the lithofacies boundaries by parameterezing five sets of well log data which are density, neutron porosity, gamma ray, sonic velocity and photoelectric cross section, obtained from two boreholes located in the Berkine sedimentary basin. This basin is located, in the North-East of the Saharan Platform. It is considered as a vast Palaeozoic depression in which the crystalline basement is covered by an important sedimentary series and presents a considerable interest in terms of oil. First, we analyze the fluctuations of these five sets of well log data using the continuous wavelet transform, power law exponents are then derived for the 2838.5-3082 m depth interval. The power law exponent allows to establish the Hölder exponent of each well-log data.Second, a Self Organizing Map (SOM) neural network model was generated in an unsupervised feed-forward mode for training the set of Hölder exponents. We can then infer the lithology sampled by the drill in two different ways: firstly by direct measurements and secondly by their corresponding Hölder exponents. The current data analysis suggests that this approach is able to emulate the pattern of all five sets of borehole data and correctly identify lithofacies. Indeed, we observe, that the spectral exponents derived logs are more efficient than the direct downhole measurements. Moreover, the results demonstrate that our approach presents a robust and powerful tool for the classification of complex lithofacies successions from the sedimentary borehole log data. This method may provide useful guide/information for understanding the petrophysical properties and structural discontinuities in other areas.