Nitrate spatial predictions by means of machine learning to improve groundwater monitoring networks

Recently, machine learning approaches are being explored as tools to underpin water management, encompassing applications such as groundwater level prediction and the integration of artificial intelligence combined with classical numerical models. This study introduces a method to support the design of groundwater quality monitoring networks through machine learning spatial predictions. Several supervised classification algorithms were trained to identify spatially distributed variables explaining the presence of nitrates in the groundwater of various aquifers in central Spain, including the Madrid Tertiary Detrital Aquifer. The dataset comprised over 240 nitrate concentration measurements and 20 explanatory variables related to geology, climatic factors, and pressures such as agricultural land, urban areas or intensive farming location. Subsequently, the algorithms with the best predictive capability were used to map nitrate contamination in order to locate unmonitored sites where contamination is likely to occur. Ensemble tree-based classifiers, such as random forests or gradient boosting, showed the most accurate predictions of groundwater contamination, with area under the curve scores around 0.8. The map-based output of this approach facilitates identifying new areas of interest requiring observation points. This method provides an alternative to expert-based criteria for locating new groundwater monitoring stations and is easily transferable to other environments.