Predicting Arsenic and Nitrate in California: a spatial and temporal Machine Learning analysis

Aquifers are essential for supporting agriculture, industry, and public water systems. However, the accumulation of chemical contaminants in groundwater is a significant concern, posing threats to human health and the environment. In California, as in many other regions, groundwater is of crucial importance and its contamination is therefore of particular concern. The coexistence of nitrate and arsenic contamination in many aquifers has created a complex situation requiring urgent attention. Understanding the interconnections between these contaminants and their stressors is crucial for effective water resources management, even though this aspect remains partially unexplored despite available extensive data in California. The aim of our study is to predict nitrate and arsenic contamination in California’s aquifer using two Machine Learning (ML) algorithms: Random Forest (RF) and Dense feed-forward Neural Network (DNN). The temporal and spatial variability of these contaminants was modelled using an extensive dataset from various sources spanning over 40 years, from 1980 to 2022. Three different configurations of dataset partitioning were evaluated: a random grouping of data; the first 38 years of data for calibration, and the last 5 years for the test set; the 70% of the monitoring sites for training and the 30% as the test set. Furthermore, the importance features assessment in the prediction was performed. The RF algorithm led to better results on average than the DNN, which, however, outperformed RF in predicting the highest values. Both algorithms achieved Nash Sutcliffe Efficiency (NSE) coefficients exceeding 0.78 for the test set in the first partitioning configuration. The sensitivity analysis of the features revealed that geographic information, along with soil hydraulic properties, had the most significant impact on the prediction. The achieved results open up new possibilities for predicting future values or reconstructing time series of contaminants and also investigating the capability to simulate contamination in ungauged sites. By employing such techniques, researchers and policymakers can proactively safeguard water quality and public health, tackling contamination challenges in California and other regions facing similar concerns.