A hybrid machine learning model to predict and visualize nitrate concentration throughout the Central Valley aquifer, California, USA.

Intense demand for water in the Central Valley of California and related increases in groundwater nitrate concentration threaten the sustainability of the groundwater resource. To assess contamination risk in the region, we developed a hybrid, non-linear, machine learning model within a statistical learning framework to predict nitrate contamination of groundwater to depths of approximately 500m below ground surface. A database of 145 predictor variables representing well characteristics, historical and current field and landscape-scale nitrogen mass balances, historical and current land use, oxidation/reduction conditions, groundwater flow, climate, soil characteristics, depth to groundwater, and groundwater age were assigned to over 6000 private supply and public supply wells measured previously for nitrate and located throughout the study area. The boosted regression tree (BRT) method was used to screen and rank variables to predict nitrate concentration at the depths of domestic and public well supplies. The novel approach included as predictor variables outputs from existing physically based models of the Central Valley. The top five most important predictor variables included two oxidation/reduction variables (probability of manganese concentration to exceed 50ppb and probability of dissolved oxygen concentration to be below 0.5ppm), field-scale adjusted unsaturated zone nitrogen input for the 1975 time period, average difference between precipitation and evapotranspiration during the years 1971-2000, and 1992 total landscape nitrogen input. Twenty-five variables were selected for the final model for log-transformed nitrate. In general, increasing probability of anoxic conditions and increasing precipitation relative to potential evapotranspiration had a corresponding decrease in nitrate concentration predictions. Conversely, increasing 1975 unsaturated zone nitrogen leaching flux and 1992 total landscape nitrogen input had an increasing relative impact on nitrate predictions. Three-dimensional visualization indicates that nitrate predictions depend on the probability of anoxic conditions and other factors, and that nitrate predictions generally decreased with increasing groundwater age.

Predicting Arsenic in Drinking Water Wells of the Central Valley, California.

Probabilities of arsenic in groundwater at depths used for domestic and public supply in the Central Valley of California are predicted using weak-learner ensemble models (boosted regression trees, BRT) and more traditional linear models (logistic regression, LR). Both methods captured major processes that affect arsenic concentrations, such as the chemical evolution of groundwater, redox differences, and the influence of aquifer geochemistry. Inferred flow-path length was the most important variable but near-surface-aquifer geochemical data also were significant. A unique feature of this study was that previously predicted nitrate concentrations in three dimensions were themselves predictive of arsenic and indicated an important redox effect at >10 µg/L, indicating low arsenic where nitrate was high. Additionally, a variable representing three-dimensional aquifer texture from the Central Valley Hydrologic Model was an important predictor, indicating high arsenic associated with fine-grained aquifer sediment. BRT outperformed LR at the 5 µg/L threshold in all five predictive performance measures and at 10 µg/L in four out of five measures. BRT yielded higher prediction sensitivity (39%) than LR (18%) at the 10 µg/L threshold-a useful outcome because a major objective of the modeling was to improve our ability to predict high arsenic areas.

Relating management practices and nutrient export in agricultural watersheds of the United States.

Relations between riverine export (load) of total nitrogen (N) and total phosphorus (P) from 133 large agricultural watersheds in the United States and factors affecting nutrient transport were evaluated using empirical regression models. After controlling for anthropogenic inputs and other landscape factors affecting nutrient transport-such as runoff, precipitation, slope, number of reservoirs, irrigated area, and area with subsurface tile drains-the relations between export and the area in the Conservation Reserve Program (CRP) (N) and conservation tillage (P) were positive. Additional interaction terms indicated that the relations between export and the area in conservation tillage (N) and the CRP (P) progressed from being clearly positive when soil erodibility was low or moderate, to being close to zero when soil erodibility was higher, to possibly being slightly negative only at the 90th to 95th percentile of soil erodibility values. Possible explanations for the increase in nutrient export with increased area in management practices include greater transport of soluble nutrients from areas in conservation tillage; lagged response of stream quality to implementation of management practices because of nitrogen transport in groundwater, time for vegetative cover to mature, and/or prior accumulation of P in soils; or limitations in the management practice and stream monitoring data sets. If lags are occurring, current nutrient export from agricultural watersheds may still be reflecting the influence of agricultural land-use practices that were in place before the implementation of these management practices.

Verifiable metamodels for nitrate losses to drains and groundwater in the Corn Belt, USA.

Nitrate leaching in the unsaturated zone poses a risk to groundwater, whereas nitrate in tile drainage is conveyed directly to streams. We developed metamodels (MMs) consisting of artificial neural networks to simplify and upscale mechanistic fate and transport models for prediction of nitrate losses by drains and leaching in the Corn Belt, USA. The two final MMs predicted nitrate concentration and flux, respectively, in the shallow subsurface. Because each MM considered both tile drainage and leaching, they represent an integrated approach to vulnerability assessment. The MMs used readily available data comprising farm fertilizer nitrogen (N), weather data, and soil properties as inputs; therefore, they were well suited for regional extrapolation. The MMs effectively related the outputs of the underlying mechanistic model (Root Zone Water Quality Model) to the inputs (R(2) = 0.986 for the nitrate concentration MM). Predicted nitrate concentration was compared with measured nitrate in 38 samples of recently recharged groundwater, yielding a Pearson's r of 0.466 (p = 0.003). Predicted nitrate generally was higher than that measured in groundwater, possibly as a result of the time-lag for modern recharge to reach well screens, denitrification in groundwater, or interception of recharge by tile drains. In a qualitative comparison, predicted nitrate concentration also compared favorably with results from a previous regression model that predicted total N in streams.