Predictive modeling of elevated groundwater nitrate in a karstic spring-contributing area using random forests and regression-kriging

Many of Florida’s large springs have seen an order of magnitude increase in nitrate concentration since the mid-twentieth century, which has contributed to the proliferation of nuisance algae and alteration of spring ecosystems. Cost-effective strategies to limit nitrate inputs require identification of contributing land areas within springs (springsheds) where surficial nitrogen sources are most likely to be transported to the underlying aquifer. To address spatial variability in vulnerability to nitrogen loading, spatial models specific to nitrate were developed for the Silver Springs springshed (Florida, USA). Random forest classification models were trained using an extensive (1554 wells) groundwater nitrate dataset assembled from public water system and agency monitoring data. Spatial layers representing soil hydrology, subsurface geology, recharge potential, and nitrogen sources were used as predictor variables. Random forest models produced out-of-bag error estimates of 21% or less, and variable importance plots indicated that a subset of subsurface geological predictors was the most important contributors to overall model accuracy. Although predictors representing land use and nitrogen sources contributed less to overall model accuracy, they were still important in the final spatial discrimination of the most vulnerable areas. Random forest model accuracy was further improved by kriging of model residuals, and kriged residuals were added to model estimates to produce final prediction maps. The models developed are well suited for a management decision framework for environmental restoration, as it informs the manager with maps of probabilistic information. Recognizing the potential for legacy nitrate impacts, we recommend the current models be adopted as a part of a tiered approach to restoration projects that first prioritizes critical areas using the models presented herein and subsequently uses site-specific information to verify local impacts.