Three-dimensional modelling using spatial regression machine learning and hydrogeological basement VES

Abstract

In the last decade, machine learning algorithms have shown their superior performance in the spatial interpolation of environmental properties compared to classical interpolation models. In particular, the random forest ensemble model has provided the best adjustment. In this work, we compare the performance of support vector machines (SVM), simple trees (ST), random forests (RF) and extremely random forests (ERF), using discrete depths obtained by vertical electrical sounding (VES) from the hydrogeological basement of a sedimentary basin in Argentina; the coordinates are not gridded but almost aligned. On the other hand, in different artificial intelligence applications, the ERF algorithm has surpassed several methods of machine learning, including random forests. To the best of our knowledge, we hereby report the first spatial regression application of the novel ERF algorithm, which predicted—even better than RF—values it had not been trained for with an average R2 score of 97.6%. This allowed us to obtain a satisfactory generalization of VES depths in the form of a three-dimensional approximation of the basement. The ERF algorithm also outperformed RF in computation time and smoothness of the surface generated. The primary significance of the results reported here lies in the relative independence that this technique has to offer, considering the area of application and gridding. Added to this, the nature of the method by means of which the discrete data are obtained is independent as well, as these could not only be derived from the VES technique, but also from well data or from different geophysical inversions.