Using Machine Learning and Geospatial Data to Predict Groundwater Occurrence of Arsenic in the Colorado Plateau

The presence of arsenic in groundwater poses a significant global health risk, particularly in arid and semi-arid regions where high concentrations are often found in alluvial aquifers. High arsenic levels can lead to severe health issues, including several cancers, impaired child development, and other health outcomes. Given the health impacts and potentially large number of individuals exposed to high levels of arsenic each year, there is an urgent need to accurately identify areas and populations at risk of exposure. Application of geospatial prediction models, such as spatial and aspatial regression techniques, are widely applied but limited due to the non-linear relationship among variables across large geographic areas that makes prediction challenging. Machine learning methods offer a promising solution for capturing the complex patterns of geological, hydrological, and climatic factors that influence arsenic occurrence in groundwater.

The purpose of this study was to predict arsenic groundwater concentrations throughout the Colorado Plateau using machine learning methods. A secondary objective was to identify counties that are at risk of arsenic exposure (defined as As concentrations >10 ug/L). To accomplish this goal we extracted water quality data on the Colorado Plateau (AZ, UT, NM, CO) from the Water Quality Portal, and compiled hydrologic, geologic, and soils geospatial data. Vector data were converted to a raster format and all prediction layers standardized to a 500 m resolution using ESRI ArcGIS Pro (v 3.3). Then using R (v 4.4) we compiled a dataset, and cleaned variables for use in ML models. The final dataset, which consisted of more than 200 variables, was used for use with XGBoost, Random Forest Classification, and Boosted Regression Tree models. The developed dataset, balanced by arsenic occurrence categories (C1 = 0-5 ug/L, C2 = 5.1-10.0 ug/L, C3 > 10 ug/L), was split into training (70%) and testing (30%) sets. Model performance was optimized using 10-fold cross-validation on the UArizona Puma Supercomputing Cluster. Results indicated that Random Forest Classification achieved the highest testing data accuracy at 91.3%, while XGBoost showed the highest training data accuracy at 82.6%. Among the models, elevation, annual mean precipitation, and minerals such as calcite, and kaolinite demonstrated high importance. Spatial analysis of the testing data indicated potential for arsenic exposure for communities located in New Mexico (Sandoval, Cibola), Arizona (Navajo, Mohave, Coconino), Utah (Washington, Iron) and Colorado (San Miguel, and Montrose). Next steps in this work include additional hyperparameter tuning, addressing incomplete observations in the dataset (e.g., missing observations for location elevation or well depth), and further refining the geographic areas of interest due to predictions of arsenic groundwater concentrations approaching or exceeding drinking water limits.