Advanced geostatistical modeling of piezometric data in Northeastern Italy

Geostatistical models are often employed within various hydrological and earth sciences applications, enabling objective-oriented estimations and the accurate quantification of local uncertainty. This work proposes to power geostatistical methods via Machine Learning algorithms to be used into their pre-processing and post-processing phases. The preprocessing algorithm drives the variogram analysis by looking for the optima variogram structure, lag distance among, and correlation scale for fitting experimental data. In the post-processing phase, once the kriging-based estimation of the target variable is acquired, the algorithm attempts to iteratively correct the field of residuals and juxtapose it to the estimated field for the best possible validation. The exit criterion is based on the Mean Average Error for the set of validation wells. These algorithms are applied to achieve piezometric reconstructions in Northeastern Italy, leveraging yearly water table data. The results show that the algorithm can learn the best spatial structure to fit experimental data, taking to an abrupt residuals drop after a few post-processing iterations within the post-processing phase. These results outperform any previous ones, leading to unprecedently accurate spatial estimations of the water table to unravel large-scale patterns in Northeastern Italy.