Atrazine, a triazine-class herbicide widely used in U.S. Corn and soybean production is highly soluble, persistent and prone to runoff. It poses risks such as reduced aquatic productivity and biological endocrine disruption. Observation-driven geostatistical models can extend sparse monitoring networks by estimating contaminant levels across space and time. We evaluated all available Iowa surface water atrazine measurements from 1986 to 2025 ( n = 13,247), identifying sparse and inconsistent sampling before 2000 and after 2004, with a period of sustained high sampling intensity from 2000 to 2004 (>1,000 samples per year). In this study, we applied a geostatistical, Bayesian Maximum Entropy (BME) framework to model log-transformed atrazine concentrations in Iowa surface waters from 2000 to 2004 using 6,478 observations from 673 monitoring sites. We compared a traditional spatial-only geostatistical framework with a fully spatiotemporal BME model incorporating both spatial and temporal data. Spatiotemporal dependence was modeled using an additive, nested space–time, exponential covariance structure. This allows spatial and temporal autocorrelation to jointly inform prediction while preserving localized contamination patterns. Model performance was evaluated using leave-one-out cross-validation. Results show spatiotemporal BME significantly outperformed spatial-only approaches, improving predictive power ( R 2 = 0.73 vs. 0.64) and reducing mean squared error (0.318 vs. 0.432) log µg 2 /L. Performance gains were largest for error metrics sensitive to extreme values, with spatiotemporal BME substantially reducing mean squared error relative to spatial-only models, indicating improved representation of episodic high-concentration events. Our findings demonstrate atrazine exposure risk appears to be predictable and likely aligned with seasonal application, chemical half-life, and delayed runoff. These processes likely produce localized hotspots often exceeding the U.S. EPA’s 3 µg/L drinking water standard. Atrazine correlation showed strong spatial (mobility 25–150 km) and temporal persistence (~20–50 days). This approach complements mechanistic models, offering a flexible, data-driven framework for environmental health monitoring where data are sparse or irregular.
Jat et al. (Wed,) studied this question.