Author(s): V. Gomez-Escalonilla; P. Martinez-Santos; A. De La Hera-Portillo; S. Diaz-Alcaide; E. Montero Gonzalez; M. Martin-Loeches

Linked Author(s):

Keywords: Groundwater vulnerability; DRASTIC; Nitrate; Monitoring; GIS; Duero; Spain

Abstract: Groundwater vulnerability is a manifold concept that ultimately represents the ease with which groundwater can get contaminated by human activities. The result of groundwater vulnerability studies is typically presented in map form, as vulnerability stems from the combination of a series of spatially distributed variables, including depth to the water table, soil type, lithology, and land use, among others. Multiple approaches to determine groundwater vulnerability have been developed over the years. These mostly include GIS-based methods and numerical methods. While information is often lacking to perform numerical modelling, GIS approaches typically suffer from two major issues, namely, the reliance on static variables and coefficients, and the dependence on aprioristic knowledge, without necessarily implying actual validation. The advent of machine learning techniques could represent a major breakthrough in groundwater vulnerability studies by providing a validation-based method to update coefficients and explanatory variables in each given case. This research aims to improve upon the classic DRASTIC approach by combining what is actually known about groundwater contamination in a given aquifer with artificial intelligence approaches. A large number of machine learning algorithms from different families was trained on groundwater monitoring data for a series of aquifers in central Spain. This served the purpose of identifying which of the DRASTIC explanatory variables (depth to the water table, recharge, aquifer media, soil type, topography, impact of the vadose zone, hydraulic conductivity, land use) were more relevant in each part of the study region. Furthermore, machine learning was used to adjust the weight of each coefficient to render a calibrated representation of groundwater vulnerability in the study region. Overall, tree-based algorithms are observed to outperform other supervised classification approaches on a regular basis. This is likely attributable to the conditional logic underlying tree algorithms, which is akin to that behind DRASTIC. Certain ensemble methods are also adept at depicting groundwater vulnerability. This approach is versatile enough to cater to other classic vulnerability methods and can be readily exported to other settings.

DOI: https://doi.org/10.3850/iahr-hic2483430201-378

Year: 2024