Skip to main content
NCBI home page
ACS AuthorChoice logoLink to ACS AuthorChoice
. 2024 Mar 7;58(11):5079–5092. doi: 10.1021/acs.est.3c07576

Predicting Redox Conditions in Groundwater at a National Scale Using Random Forest Classification

Anthony J Tesoriero †,*, Susan A Wherry †, Danielle I Dupuy ‡, Tyler D Johnson §
PMCID: PMC10956438  PMID: 38451152

Abstract

graphic file with name es3c07576_0008.jpg

Redox conditions in groundwater may markedly affect the fate and transport of nutrients, volatile organic compounds, and trace metals, with significant implications for human health. While many local assessments of redox conditions have been made, the spatial variability of redox reaction rates makes the determination of redox conditions at regional or national scales problematic. In this study, redox conditions in groundwater were predicted for the contiguous United States using random forest classification by relating measured water quality data from over 30,000 wells to natural and anthropogenic factors. The model correctly predicted the oxic/suboxic classification for 78 and 79% of the samples in the out-of-bag and hold-out data sets, respectively. Variables describing geology, hydrology, soil properties, and hydrologic position were among the most important factors affecting the likelihood of oxic conditions in groundwater. Important model variables tended to relate to aquifer recharge, groundwater travel time, or prevalence of electron donors, which are key drivers of redox conditions in groundwater. Partial dependence plots suggested that the likelihood of oxic conditions in groundwater decreased sharply as streams were approached and gradually as the depth below the water table increased. The probability of oxic groundwater increased as base flow index values increased, likely due to the prevalence of well-drained soils and geologic materials in high base flow index areas. The likelihood of oxic conditions increased as topographic wetness index (TWI) values decreased. High topographic wetness index values occur in areas with a propensity for standing water and overland flow, conditions that limit the delivery of dissolved oxygen to groundwater by recharge; higher TWI values also tend to occur in discharge areas, which may contain groundwater with long travel times. A second model was developed to predict the probability of elevated manganese (Mn) concentrations in groundwater (i.e., ≥50 μg/L). The Mn model relied on many of the same variables as the oxic/suboxic model and may be used to identify areas where Mn-reducing conditions occur and where there is an increased risk to domestic water supplies due to high Mn concentrations. Model predictions of redox conditions in groundwater produced in this study may help identify regions of the country with elevated groundwater vulnerability and stream vulnerability to groundwater-derived contaminants.

Keywords: redox reactions, machine learning, dissolved oxygen, nitrate, manganese, groundwater

Short abstract

Redox conditions in groundwater may markedly affect the fate and transport of contaminants but are not often known. This study uses random forest classification to predict redox conditions in groundwater in the contiguous United States.

1. Introduction

Redox processes influence groundwater contaminant transport and potential toxicity either by directly transforming contaminants to other species1 or by causing the precipitation or dissolution of compounds that contain or sorb contaminants in the vadose and saturated zones.2−4 These processes have been shown to markedly affect the transport of nutrients,5 volatile organic compounds,6 and trace metals7,8 and have significant implications for human health. For example, redox-sensitive constituents are among the contaminants most likely to exceed health-based screening levels for drinking water from both domestic and public supply wells.9−12 While redox conditions are typically established by natural factors, anthropogenic activities can affect redox conditions and alter contaminant mobilization. For example, denitrification of nitrate from agricultural sources may lead to the mobilization of some trace metals13,14 and irrigation induced changes in redox conditions may mobilize arsenic.15 Similarly, managed aquifer recharge may shift redox conditions, causing the release of arsenic and other geogenic contaminants.16 Redox gradients in groundwater can also have a significant effect on the fate and lag times of contaminants transported from the landscape to groundwater and streams.17,18

The influence of redox reactions on contaminant transport underscores the importance of defining redox reaction zones in groundwater (Figure 1). Redox reactions may be abiotically or biotically mediated;19 however, the overall redox conditions of most aquatic environments are determined by the terminal electron accepting processes of microbial metabolism.20 Microbial metabolism depends on the oxidation of organic or inorganic (e.g., FeS2) species to generate energy for growth and maintenance, with the metabolic reaction that yields the most energy typically dominating over competing reactions.21 The preference for the most energetically favorable reactions results in a predictable sequence of reactions, often termed the redox ladder.20 Microbes will first oxidize organic carbon or reduced minerals using O2 as an electron acceptor through aerobic respiration. After O2 is depleted, facultative anaerobes will begin to use nitrate (NO3–) as an electron acceptor during denitrification. The reduction of Mn(IV), Fe(III), and sulfate will typically occur next, followed by methanogenesis. A redox classification scheme for groundwater was established using the concentrations of products and reactants in key redox reactions.22 A similar approach has been utilized in several aquifers across the world to assess contaminant transport, both in groundwater and streams.23−25 Redox conditions have also been mapped at the local scale using electrodes26 and at the small catchment scale (101 km2) using modeling.27

Figure 1.

Figure 1

Open in a new tab

Redox conditions depend on natural and anthropogenic factors that affect the delivery of oxygen in recharging groundwater and/or the reactivity and amount of electron donors. For example, redox conditions may change as the depth below the water table increases or as organic-rich riparian zones are encountered. [Figure adapted from Wherry et al.,28 not subject to U.S. copyright. Published by the American Chemical Society].

The spatial variability of redox reaction rates has made the spatially continuous determination of redox conditions at regional scales (e.g., >100,000 km2) problematic. As a result, until recently, studies that have estimated redox conditions on the regional scale have been rare. This has changed within the past decade as machine learning and statistical techniques have been increasingly used to predict redox conditions in aquifers around the world.29−35 The advancement in machine learning techniques and the availability of a national water quality data set36 have created an opportunity to accurately predict redox conditions in groundwater across the contiguous United States, a much larger scale than previously assessed. In this study, machine learning methods were used to relate explanatory variables representing geology, soil characteristics, hydrology, land use, and many other factors to redox-active constituent concentrations to predict redox conditions in groundwater in the contiguous United States. These predictions may be important inputs for both process-based and statistical water quality models that are designed to assess the vulnerability of groundwater and streams to contamination.28

2. Methods

Redox conditions in groundwater were predicted using measured water quality data, explanatory variables, and machine learning methods. First, redox conditions in groundwater were determined for groundwater from each well in an existing database. Redox conditions in groundwater were then related to explanatory factors by using machine learning methods. Gradient boosting machines (GBM),37 extreme gradient boosting (XGBOOST),38 and random forest39 were initially employed using all variables to inform our method selection process. Model performance was similar (<2% difference in percent accuracy) for all three methods. Random forest was selected for this application primarily because of the ease with which it can handle categorical variables. The water quality data set, redox classification, explanatory variables, and the steps used to apply the random forest classification (RFC) method are described below.

2.1. Water Quality Data and Redox Classification

Chemical data from over 43,000 wells from a previous study36 were examined for use as an indicator of redox conditions in groundwater in the contiguous United States (Figures SI-1 and SI-2). Sample years range from 1988 to 2017, with the most recent or most complete data set used when wells were sampled more than once. Dissolved oxygen (O2), nitrate or nitrite plus nitrate (NO3–), Mn, and iron (Fe) data from this data set were used to establish redox conditions in groundwater. Nitrate, Mn, and Fe samples were filtered in the field (0.45 μm, typically by using acrylic polymer or glass fiber filters).

Redox conditions have often been defined using individual concentrations, such as defining oxic conditions as occurring if dissolved oxygen concentrations exceed a certain threshold.30 However, a more informative classification can be attained if the redox classification is defined using multiple constituents.40 In this study, thresholds of Mn < 50 μg/L and Fe < 100 μg/L were deemed consistent with oxic conditions based on a previous classification system.40 A threshold for dissolved O2 of 2 mg/L, rather than 0.5 mg/L as defined previously,40 was used in this study to define oxic conditions based on laboratory and field data. Laboratory studies have indicated that denitrification may occur at O2 concentrations over 1 mg/L.41 Field assessments of redox conditions may be complicated by pumping and long well screens that can mix groundwater of different redox compositions.42,43 A compilation of field studies suggested that the threshold for denitrification of 0.5 mg/L may be too low; the relation between denitrification reaction progress and O2 concentrations in groundwater from over 400 wells across the United States indicated that water from nearly all wells with O2 concentrations below 2 mg/L showed evidence of denitrification.5 A threshold of 2 mg/L was also consistent with a classification tree analysis of a larger data set (872 wells) where O2 concentrations less than 1.93 mg/L were found to be predictive of denitrifying conditions.44

For the oxic/suboxic model, water from wells were classified as oxic if the following criteria were met: dissolved O2 ≥ 2 mg/L, Mn < 50 μg/L, and Fe < 100 μg/L. Water from wells were classified as suboxic if either nitrate-reducing (i.e., dissolved O2 < 2 mg/L, nitrate ≥0.5 mg/L, Mn < 50 μg/L, and Fe < 100 μg/L), Mn-reducing (i.e., dissolved O2 < 2 mg/L, nitrate <0.5 mg/L, Mn ≥ 50 μg/L, and Fe < 100 μg/L), or Fe-reducing (i.e., dissolved O2 < 2 mg/L, nitrate <0.5 mg/L, and Fe ≥ 100 μg/L) conditions were indicated. Samples with mixed redox classifications were dropped; this included samples with such low dissolved oxygen, nitrate, Mn, and Fe concentrations that a precise redox classification could not be made. Seventy percent of the data (13,723 wells) were randomly selected to construct a training data set that was used to build the oxic/suboxic model developed for this study. Data from the remaining 30% of the wells (5881 wells) formed the hold-out data set. The hold-out data set was not used to construct the model. As a result, comparing model predictions for the hold-out data set with measured observations allowed for an independent evaluation of model performance.

A second model was constructed to evaluate the presence or absence of Mn concentrations ≥50 μg/L; concentrations at or above this level are consistent with Mn-reducing conditions.40 No other constituents were used to indicate Mn-reducing conditions for this model to maximize the number of samples available. Mn concentration and explanatory data were available from a total of 36,515 wells, with 70% of these used for the training data set (25,482) and the remaining 30% (11,033) used for the hold-out data set.

2.2. Explanatory Variables

Over 200 explanatory variables representing geology, hydrology, land use, soil hydrology and chemistry, and nitrogen inputs were compiled for the random forest models that predicted redox conditions in groundwater (Table S1). Predictors were attributed to each well either by point extraction or by calculating the mean or median value of the attribute within a 500 m radius around each well. While the optimum buffer size to assess the impact of watershed characteristics on water quality may vary,45 the 500 m buffer size has been shown to adequately represent watershed characteristics in groundwater assessments using similar data sets.46,47 Variables were considered based on their potential to describe the sources of electron donors or acceptors in the vadose (e.g., soil organic carbon content, soil drainage) and saturated zones (e.g., surficial geology and subsurface lithology), reaction rates (e.g., temperature), or travel time (e.g., depth below the water table and lateral position (LP) in the watershed). Many variables are also indicators of hydrologic conditions that could identify conducive conditions for the onset of suboxic conditions such as depth to water, soil hydrology, recharge, and topographic wetness index (TWI). Several redox-sensitive constituent concentrations in soils (e.g., Fe and Mn) were also considered as indicators of soil redox environments.

Land use practices were also included as variables, as these perturbations can alter natural redox conditions in a groundwater system. For example, increased irrigation with water containing high levels of dissolved oxygen and nitrate may lead to a more oxic condition in groundwater than would otherwise occur, potentially mobilizing contaminants.48,49 Similarly, an increase in the amount of nitrate applied to the land surface can lead to a more oxidized environment than would occur naturally.50 A complete list of the variables considered for model inclusion is provided in Table S1.

2.3. Random Forest Classification

Random forest classification (RFC) has proven to be an effective method for predicting the probability of a classification event for a binary dependent variable in environmental applications.33,51,52 RFC analysis applies the random forest algorithm to classification tree analysis, resulting in many classification trees and more accurate classifications.39,51 The random forest algorithm combines “bagging” with random selection of variables for each partition of a classification tree. The number of variables to consider at each partition, termed mtry, and the number of trees, were varied until model performance was optimized. In bagging, decision trees are created from randomly selected subsets of the training samples, with each tree yielding a prediction. These predicted values are averaged to yield the best prediction. Observations that are in the training data set but are not part of the randomly selected subset of samples used to train the model are called out-of-bag (OOB) samples and are used to test the performance of the model. Models were also evaluated using the hold-out data set.

Models were initially constructed by using all variables. However, removing variables from machine learning models has been shown to improve interpretability of the model with little loss in predictive performance.28,33,53 The relative importance of each variable was determined to aid in the selection of variables and interpretation of the final model. Variable importance was determined using the mean decrease in accuracy (MDA) observed between model results and results determined by randomly permuting a selected variable.39 MDA is calculated by determining the prediction error rate for classification on the out-of-bag portion of the data in the full model and after permutation of an individual predictor variable. The differences between these error rates are then averaged over all trees and normalized by the standard deviation of the differences to calculate the MDA. These steps are repeated for each remaining variable. Higher MDA values suggest that a variable has greater importance in comparison to variables with lower MDA values. To minimize the computer processing time, only the top 25 variables based on MDA were evaluated for consideration in the final model. If any of these 25 variables were highly correlated (r ≥ 0.7), the lower ranked correlated variable was removed from consideration for model inclusion. After removing correlated variables, the top 20 variables were evaluated for model inclusion using recursive feature elimination (RFE) in the R caret package;54 the USGS Tallgrass supercomputer was used to run these simulations.55 RFE uses backward feature elimination within a 10-fold cross validation (CV) routine performed on the training data to reduce the number of variables. During RFE, the least important variable is removed, a new model is constructed, and variable importance is reranked. This process was continued until a 10-variable model was generated. Models were then tuned using the caret package54 to determine the optimum values for mtry and the number of trees. The final parameters for the oxic/suboxic model were mtry = 5 and the number of trees = 4000. For the Mn model, the final parameters were mtry = 6 and the number of trees = 3000. The final 10-variable models were used to predict redox conditions in groundwater at 1 km resolution across the contiguous United States (CONUS). All input variables, model output, and the R scripts used to construct the models are provided in the data release associated with this article.56

Partial dependence plots were constructed to illustrate the marginal effect a variable has on the probability of an event (e.g., oxic conditions occur). Partial dependence plots for a target variable were constructed by holding all other variables at their average value and using the random forest classification model to calculate the marginal effect that the target variable has on class probabilities. Marginal effects have proven to be useful in describing the average effect of changes in explanatory variables on the change in the probability of outcomes in nonlinear models.51,57

Model performance was evaluated in the OOB and hold-out data sets by determining the percentage of samples that were correctly classified and using Cohen’s κ statistic. For both measures, it is necessary to classify each prediction as either oxic or suboxic. Samples were classified as oxic if the predicted probability of oxic water was ≥50% and suboxic if the predicted probability of oxic water was <50%. Predictions were then compared to observed conditions in both the OOB and hold-out data sets to determine the percent of observations that were correctly classified. Model sensitivity was determined by calculating the percentage of oxic observations that were correctly classified as oxic. Model specificity was determined by calculating the percentage of suboxic observations that the model correctly classified as suboxic. Performance measures were similarly calculated for the Mn model.

Cohen’s κ statistic, κ, is a statistical measure of agreement that corrects for agreement due to chance and is an effective measure for comparing presence/absence models.58 A κ value of <0.2 indicates that there is poor or only slight agreement between predictions and observations other than what would be expected by chance, values between 0.2 and 0.4 indicate fair agreement, values between 0.4 and 0.6 indicate moderate agreement, and values above 0.6 indicate substantial or near perfect agreement.59

3. Results and Discussion

3.1. Oxic/Suboxic Model

The 10 variables included in the oxic/suboxic model describe four general characteristics: hydrology, geology, soil characteristics, and hydrologic position (Table 1). Surficial geology and subsurface lithology were both included in the model, likely due to two factors: (1) the influence of geology on hydraulic conductivity and its effect on recharge and (2) the amount of electron donors present in the geologic deposits. For example, partial dependence plots suggest that volcanic deposits, carbonate residual materials, and glaciofluvial deposits were the most favorable environments for oxic conditions, likely due to the relatively rapid infiltration and low concentration of electron donors in these environments. Conversely, silty and clayey glacial till, alluvial deposits, and proglacial deposits were among the least likely to have oxic conditions. These deposits are relatively young and either limit recharge, like glacial till, or are associated with settings (e.g., stream deposition) that are associated with long groundwater residence times and greater reactivity and abundance of electron donors. These factors may also affect the downward migration of a natural weathering front.25 The lithology variable described deeper sediments than did the surficial geology variable. Unconsolidated sand and gravel aquifers, igneous and metamorphic aquifers, and volcanic aquifers were related to an increased likelihood of oxic conditions. Conversely, sedimentary aquifers were associated with suboxic conditions, suggesting an increased source of electron donors in this environment.

Table 1. Variables Included in the Oxic/Suboxic Modela.

variable mean decrease in accuracy source
Geology    
surficial geology 359 Soller et al.60
subsurface lithology 295 Kauffman et al.61
Soil Hydrology    
vadose zone water content 358 Zell and Sanford62
well-drained soils 330 Wieczorek63
soil grain size (#10 sieve) 240 Wieczorek63
poorly drained soils 214 Wieczorek63
Watershed Hydrology    
base flow index (BFI) 309 Wolock64
topographic wetness index 196 Wolock65
Hydrologic Position    
lateral position within eighth-order watershed 289 Belitz, Moore, Arnold, Sharpe, and Starn66
depth below the water table 235 Zell and Sanford62
Open in a new tab
a

The relative influence of each variable, based on mean decrease in accuracy (MDA), is also provided. Variables with higher MDA values have greater importance for model performance than variables with lower MDA values. More information on variables is provided in Table S1.

Four of the 10 variables in the oxic/suboxic model described soil properties, specifically drainage, grain size, and water content. A sharp increase in the likelihood of oxic conditions was observed as the percentage of well-drained soils increased, while a decrease in the likelihood of oxic conditions was observed as the percentage of poorly drained soils increased (Figure 2). The likelihood of oxic conditions also decreased as the percentage of fine-grained sediment increased (Figure 2). Fine-grained sediments often have lower hydraulic conductivity than coarser sediments, so differences in recharge may have been responsible for these relations. The likelihood of oxic conditions also generally decreased as vadose zone water content increased. The relations between oxic conditions and these four variables were likely driven by the interplay between recharge and redox transformations. Recharge provides aquifers with dissolved oxygen, which may be consumed by microbial processes. If recharge rates are high, oxic conditions are more likely.

Figure 2.

Figure 2

Open in a new tab

Partial dependence plots for soil variables showing the marginal effect of a single variable on the predicted probability of oxic groundwater. Soil variables in the model are (a) % of well-drained soils within buffer surrounding a well, (b) % of soil passing through a number 10 sieve, (c) % of poorly drained soils within buffer, and (d) vadose zone water content. Tick marks represent deciles of data in the training data set.

Four variables describing hydrology and/or hydrologic position were included in the oxic/suboxic model. The base flow index (BFI) is the percentage of annual streamflow that is derived from base flow. Oxic conditions were more likely as BFI increased, particularly when BFI values exceeded 60% (Figure 3). This finding is consistent with streambed measurements of dissolved oxygen in a previous study, where streambed samples in streams with low to moderate BFI values (<60%) had median dissolved oxygen concentrations near or below 2 mg/L.67 In contrast, the dissolved oxygen concentrations in streams with high BFI values typically had median streambed dissolved oxygen concentrations above 6 mg/L.67 High BFI watersheds are less likely to have confining layers in the shallow subsurface and more likely to have a stronger connection between aquifers and streams than low BFI watersheds. As a result, in areas with similar topography and climate, high BFI watersheds are expected to have higher recharge than low BFI watersheds.

Figure 3.

Figure 3

Open in a new tab

Partial dependence plots for hydrology and hydrologic position variables showing the marginal effect of a single variable on the predicted probability of oxic groundwater. Hydrology and hydrologic position variables in the model are (a) base flow index, (b) topographic wetness index, (c) lateral position in eighth-order watershed, and (d) depth below the water table. Tick marks represent deciles of data in the training data set. Some outlier values are not shown.

The second hydrology variable in the oxic/suboxic model was the topographic wetness index (TWI). TWI is a measure of the topographic control of hydrological processes.68 Areas with high values of TWI are typically more susceptible to saturated land surfaces and are more likely to produce overland flow. TWI has been used to predict soil moisture content,69 assess flood risk,70,71 predict mean transit times of surface water and groundwater,72,73 and explain ecosystem conditions.74 In this model, higher values of TWI typically resulted in a lower likelihood of oxic conditions (Figure 3). High TWI values are more likely to be in near-stream environments that may be rich in electron donors75 and have longer transit times than areas with low TWI values. Both of these factors favor reducing conditions. The TWI metric differs from BFI in that TWI relies on the topography of a watershed, while BFI values are based on measured streamflow during varying flow regimes.

Hydrologic position is expected to have a significant influence on redox conditions, since longer travel times allow for more time for redox reactions to occur. In fact, changes in redox-sensitive constituents as a function of age or depth have been used to estimate redox reaction rates in groundwater systems.5,25,76 Depth below the water table was included in the oxic/suboxic model with the likelihood of oxic conditions decreasing as depth below the water table increased (Figure 3). For areas with similar amounts of recharge, groundwater age is expected to increase as depths below the water table increase.73,77 While depth below the water table provides a key variable to explain the hydrologic position of a groundwater sample, it does not consider the relative position of a sample location within a watershed. Due to the convergence of flow paths in a discharge area, water from wells with similar depths below the water table will tend to be much older near the stream than near the watershed divide. A relatively new nationally available data set, the lateral position in a watershed, addresses this issue.66 Lateral position (LP) is defined as the relative position of a point in a watershed. LP is calculated for each stream order by dividing the shortest horizontal distance to the stream by the shortest horizontal distance from the stream to the divide, with this value then multiplied by 10,000.78 As a result, LP is dimensionless, with values varying from 0 at the stream to 10,000 at the divide. While LP values from first-order (LP1) to ninth-order (LP9) streams were considered, lateral position for eighth-order streams (LP8) was the only LP variable that was one of the top 10 predictors for the oxic/suboxic model and therefore the only LP variable included in the final model. The likelihood of oxic conditions was lowest closest to the stream, with oxic conditions increasing quickly to about a fifth of the way up the watershed (LP = 2000) and then increasing more slowly (Figure 3). A lower likelihood of oxic conditions when LP8 values were low is consistent with previous work showing the increased likelihood of suboxic conditions as streams are approached79,80 due to an increase in electron donors in riparian zones.81 However, it should be noted that these studies often documented suboxic conditions near lower-order streams than eighth-order. We speculate that LP8 was selected for model inclusion because suboxic zones adjacent to eighth-order streams were typically wider than lower-order streams and/or because the effects of lower-order streams were captured by other variables. An increase in groundwater travel times may also contribute to the relation between LP8 and redox conditions since it may be expected that groundwater ages may increase as discharge areas are approached. Interestingly, LP8 was a weak predictor of groundwater ages in a recent study in the Great Lakes, with LP values for lower-order streams better at predicting age.82

3.2. Manganese Model

The variables included in the Mn model describe the same four general characteristics as those in the oxic/suboxic model: hydrology, geology, soil characteristics, and hydrologic position (Table 2). Similar to the oxic/suboxic model, surficial geology and subsurface lithology were both included in the Mn model, likely due to their influence on hydraulic conductivity and electron donor abundance. For example, partial dependence plots suggest that sedimentary deposits were one of the most favorable lithologies for Mn-reducing conditions, likely due to a higher concentration of electron donors in this environment. For the surficial deposits, partial dependence plots suggest that fine-grained sediments and biological sediments (e.g., calcareous materials) were more likely to have Mn-reducing conditions.

Table 2. Variables Included in the Mn Modela.

variable mean decrease in accuracy source
Geology    
surficial geology 398 Soller, Reheis, Garrity, and Van Sistine60
subsurface lithology 225 Kauffman, Degnan, Belitz, Stackelberg, and Erickson61
Soil Hydrology    
somewhat poorly drained soils 171 Wieczorek63
Watershed hydrology    
base flow index 412 Wolock64
air temperature 329 PRISM Climate Group83
depth to water 337 Zell and Sanford62
topographic wetness index 229 Wolock65
Hydrologic Position    
depth below the water table 478 Zell and Sanford62
distance to drainage eighth-order watershed 311 Belitz, Moore, Arnold, Sharpe, and Starn66
lateral position within eighth-order watershed 304 Belitz, Moore, Arnold, Sharpe, and Starn66
Open in a new tab
a

The relative influence of each variable, based on mean decrease in accuracy (MDA), is also provided. Variables with higher MDA values have greater importance for model performance than variables with lower MDA values.More information on variables is provided in Table S1.

Several variables in the Mn model were factors that may be related to transport or reactions occurring between the land surface and the water table. Proximity to streams is an important predictor of elevated Mn, with two variables representing this feature in the model: lateral position in an eighth-order watershed and distance to an eighth-order watershed. Partial dependence plots for both of these variables suggest that the probability of elevated Mn concentrations increases as streams are approached (Figure 4). In fact, 26% of wells with high Mn concentrations (>300 μg/L) in this data set were found within 500 m of a river.36 Manganese reduction coupled with the oxidation of organic carbon has been well documented,84 with previous studies finding high Mn concentrations in near-stream environments, likely due to the presence of organic-rich sediments in alluvial aquifers.85,86 Higher probabilities of elevated Mn in groundwater at shallow depth to water (Figure 4) and depth below the water table (Figure 5) are also consistent with the transport of organic carbon or other processes that may occur in the near water table environment, such as the reductive dissolution of Mn-oxides that accumulate in sediments near the water table.87

Figure 4.

Figure 4

Open in a new tab

Partial dependence plots showing the marginal effect of a single variable on the predicted probability of a Mn concentration in groundwater ≥50 μg/L. All of these plots illustrate that close proximity to shallow water table environments is related to an increased probability of elevated Mn concentrations. Tick marks represent deciles of data in the training data set.

Figure 5.

Figure 5

Open in a new tab

Partial dependence plots showing the marginal effect of a single variable on the predicted probability of a Mn concentration in groundwater ≥50 μg/L. These plots illustrate the effect of hydrologic position, watershed drainage properties, and temperature on the predicted probability of elevated Mn concentrations. Tick marks represent deciles of data in the training data set. Some outlier values are not shown.

Three of the remaining variables are related to drainage conditions within 500 m of a well (Figures 4 and 5). Partial dependence plots show that the probability of elevated Mn generally increases as topographic wetness index values and the percentage of somewhat poorly drained soils increase, which may be expected since poor drainage conditions often lead to suboxic conditions and the reductive dissolution of Mn-oxides. The relation between elevated Mn and BFI may also be related to watershed drainage conditions. The decrease in the probability of elevated Mn as BFI increases may be because high BFI watersheds are more likely to have oxic water due to the well-drained conditions.67 The tendency for the probability of elevated Mn to decrease as air temperature increases (Figure 5) is not expected since O2 depletion in groundwater and the onset of Mn-reducing conditions have been shown to occur more quickly at higher temperatures.88−90 While a definitive explanation of this relation is not possible, less vegetation and organic carbon in hot, arid parts of the country may contribute to the tendency for lower Mn concentrations in these environments.

No variables representing nitrogen loading or land use were included in the oxic-suboxic or Mn models. This suggests that natural factors, not anthropogenic factors, are the dominant influences on redox conditions, at least at the scale at which this study was done. There are many notable examples of anthropogenic influences on redox conditions in the literature from both point24,91,92 and nonpoint sources.13,16 Some of these anthropogenic factors likely affected redox conditions in some samples; however, the effects were not sufficient to be selected as a major driver of redox conditions in this large data set.

3.3. Performance and Limitations

The oxic/suboxic model was tested by comparing predicted probabilities in the out-of-bag data (OOB) and hold-out data sets with observed classifications. The oxic/suboxic model correctly predicted the redox classification in 78 and 79% of samples in the OOB and hold-out data sets, respectively. κ values of 0.52 and 0.54 for the OOB and hold-out data sets, respectively, indicate moderate agreement. Sensitivity and specificity rates in the OOB data set were 86 and 65%, respectively. Similar sensitivity and specificity rates were observed in the hold-out data set (i.e., 88 and 65%). The high sensitivity values suggest that model predictions of oxic conditions have a high degree of accuracy (i.e., low false-positive rate). Conversely, predictions of suboxic conditions were more likely to be incorrect (i.e., moderate false-negative rate) than were predictions of oxic conditions.

The Mn model correctly classified whether samples were ≥50 μg/L or <50 μg/L in 78 and 80% of the samples in the OOB and hold-out data sets, respectively. κ values of 0.48 and 0.49 for the OOB and hold-out data sets, respectively, indicate moderate agreement. Sensitivity and specificity rates in the OOB data set were 56 and 89%, respectively. Similar sensitivity and specificity rates were observed in the hold-out data set (i.e., 58 and 89%). The high specificity values indicate that model predictions of Mn concentrations in groundwater <50 μg/L have a high degree of accuracy (i.e., low false-negative rate). Lower sensitivity rates for the Mn model than in the oxic/suboxic model were expected because events (i.e., Mn ≥ 50 μg/L for the Mn model and oxic classification for the oxic/suboxic model) were a smaller fraction of the sample population in the Mn data set than in oxic/suboxic data set.

There are several key limitations to the models developed to predict redox conditions in the CONUS. First, models were created by using available data that were not evenly distributed across the nation (Figures SI-1 and SI-2). Predictions may be less accurate where data are sparse and in all cases, model predictions should not take precedence over local knowledge of redox conditions in groundwater. The primary utility of model predictions is to provide guidance in areas where local assessments of redox studies do not exist. Second, redox conditions can change markedly as depth below the water table increases but the rate of change varies spatially.18,76 Depth below the water table is based on well depth and previously modeled estimates of the depth to water. Well depth was used instead of depth to the midpoint of the screened interval because screen length was not available for many wells in our data set. Relying on well depth instead of the depth to the midpoint of a screened interval may result in model predictions that overestimate the depth of a redox condition, particularly if screen lengths are large. Inaccuracies in modeled estimates of depth to water may also affect model predictions developed in this paper. Third, temporal changes in redox-sensitive concentrations were not examined, but previous studies suggest that temporal changes are likely to be small in most cases.18,93 Further, data94 from a recently completed groundwater trends study95 suggest that changes in redox classification over a decadal time scale are not common. Fourth, models may be improved if the buffer radii of variables describing surface properties are optimized and if more information is available on conditions at depth. Most variables in this study describe surface or near-surface properties. The presence of confining layers or layers rich in electron donors at depth will affect redox conditions but could not be characterized with currently available data on the CONUS scale. Some examples of data that could improve predictions include CONUS-scale estimates of apparent groundwater age, depth to a confining layer, and the location of formations that are rich in electron donors. Last, while redox-sensitive concentration thresholds used in this study have been widely applied,40,96 these thresholds vary as a function of pH, a feature that is not included in this classification system. As a result, predictions of redox conditions may be less accurate under extreme pH conditions (e.g., acid mine drainage sites).

3.4. Spatial Distribution of Redox Conditions and Implications for Vulnerability Assessments

The spatial distribution of oxic/suboxic conditions provides vital information for assessing the susceptibility of areas to redox-sensitive contaminants (Figure 6). The Mn model is not discussed in detail for brevity and because much of the discussion of the oxic–suboxic model also applies to the Mn model. Changes in redox conditions with depth were assessed by examining partial dependence plots (Figures 3d and 5a) and by comparing maps of predicted redox conditions at different depths below the water table. Partial dependence plots suggest that the probability of oxic conditions (Figure 3d) and elevated Mn decreases with depth (Figure 5a). When examined spatially, changes in redox conditions with depth were not dramatic, with areas with a high probability of oxic water generally decreasing as the depth below the water table increased (e.g., Figures 6 and SI-3). Similarly, areas with a high probability of a Mn concentration above 50 μg/L also decreased as the depth below the water table increased (Figures 7 and SI-4). Redox predictions were compared to selected field-based redox characterizations that were conducted as part of a national water quality program.95,97 Most of these studies followed contaminants along a transect from recharge in upland areas to discharge in streams. Comparing these field-based redox characterizations to model predictions provides insight into the performance of the models and potential applications of model results.

Figure 6.

Figure 6

Open in a new tab

Map depicting the predicted probability of oxic conditions in shallow groundwater. Predictions were made by assuming depth below the water table is 5 m. Selected study sites are examples of areas where detailed assessments of redox conditions were conducted using water quality data. The relations between predictions and observations at these sites are discussed in the text. A map of the predicted probability of oxic water in deeper groundwater is provided in the Supporting Information.

Figure 7.

Figure 7

Open in a new tab

Map depicting the predicted probability of a manganese concentration ≥50 μg/L in shallow groundwater. Predictions were made by assuming depth below the water table is 5 m. Selected study sites are examples of areas where detailed assessments of redox conditions were conducted using water quality data. The relations between predictions and observations at these sites are discussed in the text. A map of the predicted probability of a manganese concentration ≥50 μg/L in deeper groundwater is provided in the Supporting Information.

Redox conditions in shallow groundwater within glacial deposits in the Midwestern United States were highly variable. Suboxic areas are more likely in portions of the Midwestern United States, including large portions of Indiana, Illinois, and Iowa. Many of these areas have poorly drained soils and glacial till, with many agricultural areas using tile drains to allow for crop production. Transect studies conducted in areas with a low probability of oxic conditions in Indiana,98 Iowa,18 and Minnesota25,99 confirm the presence of suboxic conditions in shallow groundwater in these regions (see Eastern Iowa in Figure 6). Even with high nitrogen loading in these areas, suboxic conditions often lead to low nitrate concentrations in groundwater,18 but may provide conducive conditions for the mobilization of phosphorus in groundwater.100 Conversely, portions of the upper Midwest (see Central Wisconsin in Figure 6) and on the Delmarva Peninsula (see Figure 6) have a higher probability of oxic conditions due to more well-drained sediments; this finding is supported by the deep penetration of oxic conditions in these areas.101−103 These areas are much more likely to provide legacy nitrate to streams.67 Exceptions to these predictions are expected when localized confining areas are present.104

Suboxic conditions were predicted for the North Carolina coastal plain, an area that often has poor drainage and an ample supply of organic carbon. Predicted probabilities of oxic conditions were lower in the outer coastal plain than in the inner coastal plain. This is consistent with a previous assessment of redox conditions that suggested that suboxic conditions are prevalent in this area, particularly in the outer coastal plain where dissolved organic carbon contents were high.105 Prevalent suboxic conditions on the coastal plain of North Carolina likely contributes to the low risk of nitrate contamination of private wells in this area.106

The model correctly identified the Central Columbia Plateau and the High Plains aquifer as areas that are likely to be oxic. Both of these areas have extensive nitrate contamination in shallow groundwater and little denitrification due to the prevalence of oxic conditions.107 Conversely, discharge areas are more likely to be suboxic and the model correctly predicts that suboxic conditions are likely in the Mississippi Embayment108,109 and in the trough of the Central Valley of California.110 Similarly, the high probability of Mn concentrations ≥50 μg/L in the Mississippi Embayment (Figure 7) is consistent with previous work.111

Predicted probabilities of oxic conditions should prove useful for assessing the vulnerability of groundwater to redox-sensitive contaminants. Redox-sensitive contaminants, such as nitrate and arsenic, are among the most common constituents that exceed maximum contaminant levels (MCLs) in public drinking water supplies.11 Sampling of domestic wells for water quality constituents in the United States generally occurs less frequently than public water supplies because domestic wells are not regulated at the federal level. A nationwide sampling of 2167 domestic wells in the United States suggested that two redox-sensitive contaminants, arsenic and Mn, were among the most likely to exceed maximum contaminant levels (MCLs) or secondary MCLs.10 In California, people served by water from domestic wells may face greater water quality concerns than those served by public water supplies, with nitrate, arsenic, and chromium(VI) as the major contaminants.112 Given the prevalence of elevated concentrations of redox-sensitive contaminants, predictions of redox conditions provided in this study should be useful in prioritizing monitoring efforts, particularly for domestic wells, since they may pose a greater risk to human health and are often not regularly monitored. Predictions of redox conditions in groundwater may also prove useful as an explanatory variable in statistical or machine learning models that predict concentrations for a specific constituent (e.g., nitrate) and may also help estimate reaction rates for process-based models.

Predictions of redox conditions in groundwater may also inform estimates of redox-sensitive concentrations in stream base flow; lack of information on reaction rates is considered a significant data gap for predicting nitrate inputs to streams from groundwater.113 Model results presented here can help predict whether denitrification will occur prior to discharge. For example, streams that are in watersheds that are dominated by oxic water may be more likely to have high nitrate concentrations during base flow conditions.28 These oxic watersheds may also be more likely to discharge nitrate to streams that recharged the aquifer a decade or more ago (i.e., legacy nitrate). As a result, redox predictions may help with spatial targeting of conservation measures, a key strategy for improving water quality when legacy issues are present.114

Acknowledgments

This research was funded by the U.S. Geological Survey’s Water Quality Processes Program and the National-Extent Groundwater Quality Prediction Project. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government. The authors thank the journal associate editor and anonymous reviewers for many insightful comments on this manuscript. Peter McMahon, Christopher Green, and J.K. Böhlke are also acknowledged for their reviews of an earlier version of this manuscript. Erin Poor’s help with the data release that accompanies this paper is also appreciated. This paper is dedicated to the memory of our coauthor Tyler Johnson.

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.est.3c07576.

  • Table that lists all of the variables considered in the models constructed in this paper; maps showing sample locations; and maps showing predictions of redox conditions at depth (PDF)

The authors declare no competing financial interest.

Author Status

∥ Deceased.

Notes

Model input and output files, R scripts, and other information on model construction are provided in a data release associated with this publication.56

Supplementary Material

es3c07576_si_001.pdf (1.2MB, pdf)

References

  1. Borch T.; Kretzschmar R.; Kappler A.; Cappellen P. V.; Ginder-Vogel M.; Voegelin A.; Campbell K. Biogeochemical redox processes and their Impact on contaminant dynamics. Environ. Sci. Technol. 2010, 44 (1), 15–23. 10.1021/es9026248. [DOI] [PubMed] [Google Scholar]
  2. Carlyle G. C.; Hill A. R. Groundwater phosphate dynamics in a river riparian zone: effects of hydrologic flowpaths, lithology and redox chemistry. J. Hydrol. 2001, 247 (3–4), 151–168. 10.1016/S0022-1694(01)00375-4. [DOI] [Google Scholar]
  3. Hering J. G.; Hug S. J.; Farnsworth C.; O’Day P. A.. Role of coupled redox transformations in the mobilization and sequestration of arsenic. Aquatic Redox Chemistry; American Chemical Society, 2011; Vol. 1071, pp 463–476. [Google Scholar]
  4. Nghiem A. A.; Shen Y.; Stahl M.; Sun J.; Haque E.; DeYoung B.; Nguyen K. N.; Thi Mai T.; Trang P. T. K.; Pham H. V.; Mailloux B.; Harvey C. F.; van Geen A.; Bostick B. C. Aquifer-scale observations of iron redox transformations in arsenic-impacted environments to predict future contamination. Environ. Sci. Technol. Lett. 2020, 7 (12), 916–922. 10.1021/acs.estlett.0c00672. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Tesoriero A. J.; Puckett L. J. O2 reduction and denitrification rates in shallow aquifers. Water Resour. Res. 2011, 47, W12522 10.1029/2011WR010471. [DOI] [Google Scholar]
  6. Essaid H. I.; Bekins B. A.; Cozzarelli I. M. Organic contaminant transport and fate in the subsurface: Evolution of knowledge and understanding. Water Resour. Res. 2015, 51 (7), 4861–4902. 10.1002/2015WR017121. [DOI] [Google Scholar]
  7. Smedley P. L.; Kinniburgh D. G. A review of the source, behaviour and distribution of arsenic in natural waters. Appl. Geochem. 2002, 17 (5), 517–568. 10.1016/S0883-2927(02)00018-5. [DOI] [Google Scholar]
  8. Pi K.; Wang Y.; Xie X.; Ma T.; Su C.; Liu Y. Role of sulfur redox cycling on arsenic mobilization in aquifers of Datong Basin, northern China. Appl. Geochem. 2017, 77, 31–43. 10.1016/j.apgeochem.2016.05.019. [DOI] [Google Scholar]
  9. Toccalino P. L.; Norman J. E.; Hitt K. J.. Quality of Source Water from Public-Supply Wells in the United States, 1993–2007, U.S. Geological Survey Scientific Investigations Report 2010–5024; U.S. Geological Survey, 2010.
  10. DeSimone L. A.Quality of Water from Domestic Wells in Principal Aquifers of the United States, 1991–2004, U.S. Geological Survey Scientific Investigations Report 2008–5227; U.S. Geological Survey, 2009. https://pubs.usgs.gov/sir/2008/5227/ (last accessed Feb 08, 2024).
  11. Allaire M.; Wu H.; Lall U. National trends in drinking water quality violations. Proc. Natl. Acad. Sci. U.S.A. 2018, 115 (9), 2078. 10.1073/pnas.1719805115. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Siegel H. G.; Soriano M. A.; Clark C. J.; Johnson N. P.; Wulsin H. G.; Deziel N. C.; Plata D. L.; Darrah T. H.; Saiers J. E. Natural and anthropogenic processes affecting domestic groundwater quality within the northwestern Appalachian Basin. Environ. Sci. Technol. 2022, 56 (19), 13761–13773. 10.1021/acs.est.2c04011. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Riedel T.; Kübeck C.; Quirin M. Legacy nitrate and trace metal (Mn, Ni, As, Cd, U) pollution in anaerobic groundwater: Quantifying potential health risk from “the other nitrate problem. Appl. Geochem. 2022, 139, 105254 10.1016/j.apgeochem.2022.105254. [DOI] [Google Scholar]
  14. Westrop J. P.; Yadav P.; Nolan P. J.; Campbell K. M.; Singh R.; Bone S. E.; Chan A. H.; Kohtz A. J.; Pan D.; Healy O.; Bargar J. R.; Snow D. D.; Weber K. A. Nitrate-stimulated release of naturally occurring sedimentary uranium. Environ. Sci. Technol. 2023, 57 (10), 4354–4366. 10.1021/acs.est.2c07683. [DOI] [PubMed] [Google Scholar]
  15. Azam M. S.; Shafiquzzaman M.; Haider H. Arsenic release dynamics of paddy field soil during groundwater irrigation and natural flooding. J. Environ. Manage. 2023, 343, 118204 10.1016/j.jenvman.2023.118204. [DOI] [PubMed] [Google Scholar]
  16. Fakhreddine S.; Prommer H.; Scanlon B. R.; Ying S. C.; Nicot J.-P. Mobilization of arsenic and other naturally occurring contaminants during managed aquifer recharge: A critical review. Environ. Sci. Technol. 2021, 55 (4), 2208–2223. 10.1021/acs.est.0c07492. [DOI] [PubMed] [Google Scholar]
  17. Zachara J. M.; Long P. E.; Bargar J.; Davis J. A.; Fox P.; Fredrickson J. K.; Freshley M. D.; Konopka A. E.; Liu C.; McKinley J. P.; Rockhold M. L.; Williams K. H.; Yabusaki S. B. Persistence of uranium groundwater plumes: Contrasting mechanisms at two DOE sites in the groundwater–river interaction zone. J. Contam. Hydrol. 2013, 147, 45–72. 10.1016/j.jconhyd.2013.02.001. [DOI] [PubMed] [Google Scholar]
  18. Tesoriero A. J.; Stratton L. E.; Miller M. P. Influence of redox gradients on nitrate transport from the landscape to groundwater and streams. Sci. Total Environ. 2021, 800, 150200 10.1016/j.scitotenv.2021.150200. [DOI] [PubMed] [Google Scholar]
  19. Falkowski P. G.; Fenchel T.; Delong E. F. The microbial engines that drive Earth’s biogeochemical cycles. Science 2008, 320 (5879), 1034–1039. 10.1126/science.1153213. [DOI] [PubMed] [Google Scholar]
  20. Grundl T. J.; Haderlein S.; Nurmi J. T.; Tratnyek P. G.. Introduction to Aquatic Redox Chemistry. Aquatic Redox Chemistry; American Chemical Society, 2011; Vol. 1071, pp 1–14. [Google Scholar]
  21. Pankow J. F.Aquatic Chemistry Concepts, 2nd ed.; CRC Press, 2019. [Google Scholar]
  22. Chapelle F. H.; McMahon P. B.; Dubrovsky N. M.; Fujii R. F.; Oaksford E. T.; Vroblesky D. A. Deducing the distribution of terminal electron-accepting processes in hydrologically diverse groundwater systems. Water Resour. Res. 1995, 31 (2), 359–371. 10.1029/94WR02525. [DOI] [Google Scholar]
  23. Postma D.; Larsen F.; Hue N. T. M.; Duc M. T.; Viet P. H.; Nhan P. Q.; Jessen S. Arsenic in groundwater of the Red River floodplain, Vietnam: Controlling geochemical processes and reactive transport modeling. Geochim. Cosmochim. Acta 2007, 71 (21), 5054–5071. 10.1016/j.gca.2007.08.020. [DOI] [Google Scholar]
  24. Christensen T. H.; Bjerg P. L.; Banwart S. A.; Jakobsen R.; Heron G.; Albrechtsen H. J. Characterization of redox conditions in groundwater contaminant plumes. J. Contam. Hydrol. 2000, 45 (3–4), 165–241. 10.1016/S0169-7722(00)00109-1. [DOI] [Google Scholar]
  25. Böhlke J. K.; Wanty R.; Tuttle M.; Delin G.; Landon M. Denitrification in the recharge area and discharge area of a transient agricultural nitrate plume in a glacial outwash sand aquifer, Minnesota. Water Resour. Res. 2002, 38 (7), 10-1–10-26. 10.1029/2001WR000663. [DOI] [Google Scholar]
  26. Naudet V.; Revil A.; Bottero J. Y.; Begassat P. Relationship between self-potential (SP) signals and redox conditions in contaminated groundwater. Geophys. Res. Lett. 2003, 30 (21), 2091 10.1029/2003GL018096. [DOI] [Google Scholar]
  27. Hansen A. L.; Christensen B. S. B.; Ernstsen V.; He X.; Refsgaard J. C. A concept for estimating depth of the redox interface for catchment-scale nitrate modelling in a till area in Denmark. Hydrogeol. J. 2014, 22 (7), 1639–1655. 10.1007/s10040-014-1152-y. [DOI] [Google Scholar]
  28. Wherry S. A.; Tesoriero A. J.; Terziotti S. Factors affecting nitrate concentrations in stream base flow. Environ. Sci. Technol. 2021, 55 (2), 902–911. 10.1021/acs.est.0c02495. [DOI] [PubMed] [Google Scholar]
  29. Knoll L.; Breuer L.; Bach M. Nation-wide estimation of groundwater redox conditions and nitrate concentrations through machine learning. Environ. Res. Lett. 2020, 15 (6), 064004 10.1088/1748-9326/ab7d5c. [DOI] [Google Scholar]
  30. Tesoriero A. J.; Terziotti S.; Abrams D. B. Predicting redox conditions in groundwater at a regional scale. Environ. Sci. Technol. 2015, 49 (16), 9657–9664. 10.1021/acs.est.5b01869. [DOI] [PubMed] [Google Scholar]
  31. Koch J.; Stisen S.; Refsgaard J. C.; Ernstsen V.; Jakobsen P. R.; Højberg A. L. Modeling depth of the redox interface at high resolution at national scale using random forest and residual gaussian simulation. Water Resour. Res. 2019, 55 (2), 1451–1469. 10.1029/2018WR023939. [DOI] [Google Scholar]
  32. Close M. E.; Abraham P.; Humphries B.; Lilburne L.; Cuthill T.; Wilson S. Predicting groundwater redox status on a regional scale using linear discriminant analysis. J. Contam. Hydrol. 2016, 191, 19–32. 10.1016/j.jconhyd.2016.04.006. [DOI] [PubMed] [Google Scholar]
  33. Tesoriero A. J.; Gronberg J. A.; Juckem P. F.; Miller M. P.; Austin B. P. Predicting redox-sensitive contaminant concentrations in groundwater using random forest classification. Water Resour. Res. 2017, 53 (8), 7316–7331. 10.1002/2016WR020197. [DOI] [Google Scholar]
  34. Wilson S. R.; Close M. E.; Abraham P. Applying linear discriminant analysis to predict groundwater redox conditions conducive to denitrification. J. Hydrol. 2018, 556, 611–624. 10.1016/j.jhydrol.2017.11.045. [DOI] [Google Scholar]
  35. Erickson M. L.; Elliott S. M.; Brown C. J.; Stackelberg P. E.; Ransom K. M.; Reddy J. E. Machine learning predicted redox conditions in the glacial aquifer system, northern continental United States. Water Resour. Res. 2021, 57 (4), e2020WR028207 10.1029/2020WR028207. [DOI] [Google Scholar]
  36. McMahon P. B.; Belitz K.; Reddy J. E.; Johnson T. D. Elevated manganese concentrations in United States groundwater, role of land surface–soil–aquifer connections. Environ. Sci. Technol. 2019, 53 (1), 29–38. 10.1021/acs.est.8b04055. [DOI] [PubMed] [Google Scholar]
  37. Elith J.; Leathwick J. R.; Hastie T. A working guide to boosted regression trees. J. Anim. Ecol. 2008, 77 (4), 802–813. 10.1111/j.1365-2656.2008.01390.x. [DOI] [PubMed] [Google Scholar]
  38. Chen T.; Guestrin C. In XGBoost: A Scalable Tree Boosting System, KDD ’16Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016; pp 785–794. 10.1145/2939672.2939785 (accessed Dec 14, 2023). [DOI]
  39. Breiman L. Random forests. Mach. Learn. 2001, 45 (1), 5–32. 10.1023/A:1010933404324. [DOI] [Google Scholar]
  40. McMahon P. B.; Chapelle F. H. Redox processes and water quality of selected principal aquifer systems. Ground Water 2008, 46 (2), 259–271. 10.1111/j.1745-6584.2007.00385.x. [DOI] [PubMed] [Google Scholar]
  41. Chen F.; Xia Q.; Ju L. K. Aerobic denitrification of Pseudomonas aeruginosa monitored by online NAD(P)H fluorescence. Appl. Environ. Microbiol. 2003, 69 (11), 6715–6722. 10.1128/AEM.69.11.6715-6722.2003. [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. McMahon P. B.; Chapelle F. H.; Bradley P. M.. Evolution of Redox Processes in Groundwater. Aquatic Redox Chemistry; American Chemical Society, 2011; Vol. 1071, pp 581–597. [Google Scholar]
  43. Green C. T.; Böhlke J. K.; Bekins B. A.; Phillips S. P. Mixing effects on apparent reaction rates and isotope fractionation during denitrification in a heterogeneous aquifer. Water Resour. Res. 2010, 46 (8), W08525 10.1029/2009WR008903. [DOI] [Google Scholar]
  44. Hinkle S. R.; Tesoriero A. J. Nitrogen speciation and trends, and prediction of denitrification extent, in shallow US groundwater. J. Hydrol. 2014, 509, 343–353. 10.1016/j.jhydrol.2013.11.048. [DOI] [Google Scholar]
  45. Rasool U.; Yin X.; Xu Z.; Faheem M.; Rasool M. A.; Siddique J.; Hassan M. A.; Senapathi V. Evaluating the relationship between groundwater quality and land use in an urbanized watershed. Environ. Sci. Pollut. Res. 2023, 30 (31), 77107–77126. 10.1007/s11356-023-27775-8. [DOI] [PubMed] [Google Scholar]
  46. Johnson T. D.; Belitz K. Assigning land use to supply wells for the statistical characterization of regional groundwater quality: Correlating urban land use and VOC occurrence. J. Hydrol. 2009, 370 (1), 100–108. 10.1016/j.jhydrol.2009.02.056. [DOI] [Google Scholar]
  47. Bawa R.; Dwivedi P. Impact of land cover on groundwater quality in the Upper Floridan Aquifer in Florida, United States. Environ. Pollut. 2019, 252, 1828–1840. 10.1016/j.envpol.2019.06.054. [DOI] [PubMed] [Google Scholar]
  48. Moon H. S.; Komlos J.; Jaffé P. R. Biogenic U(IV) oxidation by dissolved oxygen and nitrate in sediment after prolonged U(VI)/Fe(III)/SO42– reduction. J. Contam. Hydrol. 2009, 105 (1), 18–27. 10.1016/j.jconhyd.2008.10.014. [DOI] [PubMed] [Google Scholar]
  49. Xie X.; Wang Y.; Li J.; Yu Q.; Wu Y.; Su C.; Duan M. Effect of irrigation on Fe(III)–SO42– redox cycling and arsenic mobilization in shallow groundwater from the Datong basin, China: Evidence from hydrochemical monitoring and modeling. J. Hydrol. 2015, 523, 128–138. 10.1016/j.jhydrol.2015.01.035. [DOI] [Google Scholar]
  50. van Berk W.; Fu Y. Redox roll-front mobilization of geogenic uranium by nitrate input into aquifers: risks for groundwater resources. Environ. Sci. Technol. 2017, 51 (1), 337–345. 10.1021/acs.est.6b01569. [DOI] [PubMed] [Google Scholar]
  51. Cutler D. R.; Edwards T. C.; Beard K. H.; Cutler A.; Hess K. T.; et al. Random forests for classification in ecology. Ecology 2007, 88 (11), 2783–2792. 10.1890/07-0539.1. [DOI] [PubMed] [Google Scholar]
  52. Pennino M. J.; Leibowitz S. G.; Compton J. E.; Hill R. A.; Sabo R. D. Patterns and predictions of drinking water nitrate violations across the conterminous United States. Sci. Total Environ. 2020, 722, 137661 10.1016/j.scitotenv.2020.137661. [DOI] [PMC free article] [PubMed] [Google Scholar]
  53. Díaz-Uriarte R.; de Andres S. A. Gene selection and classification of microarray data using random forest. BMC Bioinf. 2006, 7, 3 10.1186/1471-2105-7-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  54. Kuhn M.; Wing J.; Weston S.; Williams A.; Keefer C.; Engelhardt A.; Cooper T.; Mayer Z.; Kenkel B.; Benesty M.; Lescarbeau R.; Ziem A.; Scrucca L.; Tang Y.; Candan C.; Hunt T.. caret: Classification and Regression Training. 2022. https://CRAN.R-project.org/package=caret (accessed Dec 19, 2023).
  55. Falgout J. T.; Gordon J.; Davis M. J.. USGS Advanced Research Computing, USGS Tallgrass Supercomputer; U.S. Geological Survey, 2023. 10.5066/P9XE7ROJ (accessed Nov 27, 2023). [DOI]
  56. Wherry S. A.; Tesoriero A. J.; Dupuy D. I.. Input and Results from a Random Forest Classification (RFC) Model That Predicts Redox Conditions in Groundwater in the Contiguous United States, U.S. Geological Survey Data Release; U.S. Geological Survey, 2023. 10.5066/P9DVPJIX (accessed Feb 02, 2024). [DOI]
  57. Norton E. C.; Dowd B. E.; Maciejewski M. L. Marginal effects—quantifying the effect of changes in risk factors in logistic regression models. JAMA 2019, 321 (13), 1304–1305. 10.1001/jama.2019.1954. [DOI] [PubMed] [Google Scholar]
  58. Manel S.; Ceri Williams H.; Ormerod S. J. Evaluating presence-absence models in ecology: The need to account for prevalence. J. Appl. Ecol. 2001, 38 (5), 921–931. 10.1046/j.1365-2664.2001.00647.x. [DOI] [Google Scholar]
  59. Landis J. R.; Koch G. G. The measurement of observer agreement for categorical data. Biometrics 1977, 33, 159–174. 10.2307/2529310. [DOI] [PubMed] [Google Scholar]
  60. Soller D. R.; Reheis M. C.; Garrity C. P.; Van Sistine D. R.. Map Database for Surficial Materials in the Conterminous United States, U.S. Geological Survey Data Series 425; U.S. Geological Survey, 2009. https://pubs.usgs.gov/ds/425/ (accessed April 15, 2018).
  61. Kauffman L. J.; Degnan J. R.; Belitz K.; Stackelberg P. E.; Erickson M. L.. Data for Depth of Groundwater Used for Drinking-Water Supplies in the United States, U.S. Geological Survey Data Release; U.S. Geological Survey, 2021. 10.5066/P94640EM (accessed Aug 09, 2022). [DOI]
  62. Zell W. O.; Sanford W. E. Calibrated simulation of the long-term average surficial groundwater system and derived spatial distributions of its characteristics for the contiguous United States. Water Resour. Res. 2020, 56 (8), e2019WR026724 10.1029/2019WR026724. [DOI] [Google Scholar]
  63. Wieczorek M. E.Area- and Depth-Weighted Averages of Selected SSURGO Variables for the Conterminous United States and District of Columbia, U.S. Geological Survey Data Series 866; U.S. Geological Survey, 2014.
  64. Wolock D. M.Base-Flow Index Grid for the Conterminous United States, U.S. Geological Survey Data Release; U.S. Geological Survey, 2003. 10.5066/P9MCTH3J (last accessed Jan 12, 2024). [DOI]
  65. Wolock D. M.Saturation Overland Flow Estimated by TOPMODEL for the Conterminous United States, U.S. Geological Survey Data Release; U.S. Geological Survey, 2003. 10.5066/P98MA1KO (last accessed Jan 12, 2024). [DOI]
  66. Belitz K.; Moore R. B.; Arnold T. L.; Sharpe J. B.; Starn J. J. Multiorder hydrologic position in the conterminous United States: A set of metrics in support of groundwater mapping at regional and national scales. Water Resour. Res. 2019, 55 (12), 11188–11207. 10.1029/2019WR025908. [DOI] [Google Scholar]
  67. Tesoriero A. J.; Duff J. H.; Saad D. A.; Spahr N. E.; Wolock D. M. Vulnerability of streams to legacy nitrate sources. Environ. Sci. Technol. 2013, 47 (8), 3623–3629. 10.1021/es305026x. [DOI] [PubMed] [Google Scholar]
  68. Sørensen R.; Zinko U.; Seibert J. On the calculation of the topographic wetness index: evaluation of different methods based on field observations. Hydrol. Earth Syst. Sci. 2006, 10 (1), 101–112. 10.5194/hess-10-101-2006. [DOI] [Google Scholar]
  69. Buchanan B. P.; Fleming M.; Schneider R. L.; Richards B. K.; Archibald J.; Qiu Z.; Walter M. T. Evaluating topographic wetness indices across central New York agricultural landscapes. Hydrol. Earth Syst. Sci. 2014, 18 (8), 3279–3299. 10.5194/hess-18-3279-2014. [DOI] [Google Scholar]
  70. Choubin B.; Moradi E.; Golshan M.; Adamowski J.; Sajedi-Hosseini F.; Mosavi A. An ensemble prediction of flood susceptibility using multivariate discriminant analysis, classification and regression trees, and support vector machines. Sci. Total Environ. 2019, 651, 2087–2096. 10.1016/j.scitotenv.2018.10.064. [DOI] [PubMed] [Google Scholar]
  71. Pourali S. H.; Arrowsmith C.; Chrisman N.; Matkan A. A.; Mitchell D. Topography wetness index application in flood-risk-based land use planning. Appl. Spat. Anal. Policy 2016, 9 (1), 39–54. 10.1007/s12061-014-9130-2. [DOI] [Google Scholar]
  72. Hrachowitz M.; Soulsby C.; Tetzlaff D.; Dawson J. J. C.; Malcolm I. A. Regionalization of transit time estimates in montane catchments by integrating landscape controls. Water Resour. Res. 2009, 45 (5), W05421 10.1029/2008WR007496. [DOI] [Google Scholar]
  73. Green C. T.; Ransom K. M.; Nolan B. T.; Liao L.; Harter T. Machine learning predictions of mean ages of shallow well samples in the Great Lakes Basin, USA. J. Hydrol. 2021, 603, 126908 10.1016/j.jhydrol.2021.126908. [DOI] [Google Scholar]
  74. Bader M. Y.; Ruijten J. J. A. A topography-based model of forest cover at the alpine tree line in the tropical Andes. J. Biogeogr. 2008, 35 (4), 711–723. 10.1111/j.1365-2699.2007.01818.x. [DOI] [Google Scholar]
  75. Hayakawa A.; Funaki Y.; Sudo T.; Asano R.; Murano H.; Watanabe S.; Ishida T.; Ishikawa Y.; Hidaka S. Catchment topography and the distribution of electron donors for denitrification control the nitrate concentration in headwater streams of the Lake Hachiro watershed. Soil Sci. Plant Nutr. 2020, 66 (6), 906–918. 10.1080/00380768.2020.1827292. [DOI] [Google Scholar]
  76. Kolbe T.; de Dreuzy J.-R.; Abbott B. W.; Aquilina L.; Babey T.; Green C. T.; Fleckenstein J. H.; Labasque T.; Laverman A. M.; Marçais J.; Peiffer S.; Thomas Z.; Pinay G. Stratification of reactivity determines nitrate removal in groundwater. Proc. Natl. Acad. Sci. U.S.A. 2019, 116 (7), 2494–2499. 10.1073/pnas.1816892116. [DOI] [PMC free article] [PubMed] [Google Scholar]
  77. McMahon P. B.; Plummer L. N.; Böhlke J. K.; Shapiro S. D.; Hinkle S. R. A comparison of recharge rates in aquifers of the United States based on groundwater-age data. Hydrogeol. J. 2011, 19 (4), 779–800. 10.1007/s10040-011-0722-5. [DOI] [Google Scholar]
  78. Ransom K. M.; Nolan B. T.; Stackelberg P. E.; Belitz K.; Fram M. S. Machine learning predictions of nitrate in groundwater used for drinking supply in the conterminous United States. Sci. Total Environ. 2022, 807, 151065 10.1016/j.scitotenv.2021.151065. [DOI] [PubMed] [Google Scholar]
  79. Hill A. R. Groundwater nitrate removal in riparian buffer zones: a review of research progress in the past 20 years. Biogeochemistry 2019, 143 (3), 347–369. 10.1007/s10533-019-00566-5. [DOI] [Google Scholar]
  80. Steiness M.; Jessen S.; van’t Veen S. G. M.; Kofod T.; Højberg A. L.; Engesgaard P. Nitrogen-loads to streams: Importance of bypass flow and nitrate removal processes. J. Geophys. Res.: Biogeosci. 2021, 126 (5), e2020JG006111 10.1029/2020JG006111. [DOI] [Google Scholar]
  81. Tesoriero A. J.; Liebscher H.; Cox S. E. Mechanism and rate of denitrification in an agricultural watershed: Electron and mass balance along groundwater flow paths. Water Resour. Res. 2000, 36 (6), 1545–1559. 10.1029/2000WR900035. [DOI] [Google Scholar]
  82. Green C. T.; Ransom K. M.; Nolan B. T.; Liao L.; Harter T. Machine learning predictions of mean ages of shallow well samples in the Great Lakes Basin, USA. J. Hydrol. 2021, 603 (Part B), 126908 10.1016/j.jhydrol.2021.126908. [DOI] [Google Scholar]
  83. PRISM Climate Group . United States Average Annual Precipitation and Temperature Data, 1981–2010 (800m; ASCII GRID); Oregon State University, 2014. https://prism.oregonstate.edu/normals (accessed Aug 11, 2014).
  84. Wang X.; Xie G.-J.; Tian N.; Dang C.-C.; Cai C.; Ding J.; Liu B.-F.; Xing D.-F.; Ren N.-Q.; Wang Q. Anaerobic microbial manganese oxidation and reduction: A critical review. Sci. Total Environ. 2022, 822, 153513 10.1016/j.scitotenv.2022.153513. [DOI] [PubMed] [Google Scholar]
  85. McArthur J. M.; Sikdar P. K.; Nath B.; Grassineau N.; Marshall J. D.; Banerjee D. M. Sedimentological control on Mn, and other trace elements, in groundwater of the Bengal Delta. Environ. Sci. Technol. 2012, 46 (2), 669–676. 10.1021/es202673n. [DOI] [PubMed] [Google Scholar]
  86. de Meyer C. M. C.; Rodríguez J. M.; Carpio E. A.; García P. A.; Stengel C.; Berg M. Arsenic, manganese and aluminum contamination in groundwater resources of Western Amazonia (Peru). Sci. Total Environ. 2017, 607–608, 1437–1450. 10.1016/j.scitotenv.2017.07.059. [DOI] [PubMed] [Google Scholar]
  87. Gillispie E. C.; Austin R. E.; Rivera N. A.; Bolich R.; Duckworth O. W.; Bradley P.; Amoozegar A.; Hesterberg D.; Polizzotto M. L. Soil weathering as an engine for manganese contamination of well water. Environ. Sci. Technol. 2016, 50 (18), 9963–9971. 10.1021/acs.est.6b01686. [DOI] [PubMed] [Google Scholar]
  88. Burke V.; Greskowiak J.; Asmuß T.; Bremermann R.; Taute T.; Massmann G. Temperature dependent redox zonation and attenuation of wastewater-derived organic micropollutants in the hyporheic zone. Sci. Total Environ. 2014, 482–483, 53–61. 10.1016/j.scitotenv.2014.02.098. [DOI] [PubMed] [Google Scholar]
  89. Munz M.; Oswald S. E.; Schäfferling R.; Lensing H.-J. Temperature-dependent redox zonation, nitrate removal and attenuation of organic micropollutants during bank filtration. Water Res. 2019, 162, 225–235. 10.1016/j.watres.2019.06.041. [DOI] [PubMed] [Google Scholar]
  90. Prommer H.; Stuyfzand P. J. Identification of temperature-dependent water quality changes during a deep well injection experiment in a pyritic aquifer. Environ. Sci. Technol. 2005, 39 (7), 2200–2209. 10.1021/es0486768. [DOI] [PubMed] [Google Scholar]
  91. Abiriga D.; Vestgarden L. S.; Klempe H. Long-term redox conditions in a landfill-leachate-contaminated groundwater. Sci. Total Environ. 2021, 755, 143725 10.1016/j.scitotenv.2020.143725. [DOI] [PubMed] [Google Scholar]
  92. Lønborg M. J.; Engesgaard P.; Bjerg P. L.; Rosbjerg D. A steady state redox zone approach for modeling the transport and degradation of xenobiotic organic compounds from a landfill site. J. Contam. Hydrol. 2006, 87 (3), 191–210. 10.1016/j.jconhyd.2006.05.004. [DOI] [PubMed] [Google Scholar]
  93. Degnan J. R.; Levitt J. P.; Erickson M. L.; Jurgens B. C.; Lindsey B. D.; Ayotte J. D. Time scales of arsenic variability and the role of high-frequency monitoring at three water-supply wells in New Hampshire, USA. Sci. Total Environ. 2020, 709, 135946 10.1016/j.scitotenv.2019.135946. [DOI] [PubMed] [Google Scholar]
  94. Lindsey B. D.; Dondero A. M.; Watson E.; Johnson T. D.. Data from Decadal Change in Groundwater Quality Web Site, 1988–2022, U.S. Geological Survey Data Release; U.S. Geological Survey, 2023. 10.5066/P9YEB7FS (accessed Dec 05, 2023). [DOI]
  95. Lindsey B. D.; Fleming B. J.; Goodling P. J.; Dondero A. M. Thirty years of regional groundwater-quality trend studies in the United States: Major findings and lessons learned. J. Hydrol. 2023, 627, 130427 10.1016/j.jhydrol.2023.130427. [DOI] [Google Scholar]
  96. Jurgens B. C.; McMahon P. B.; Chapelle F. H.; Eberts S. M.. An Excel Workbook for Identifying Redox Processes in Ground Water, U.S. Geological Survey Open-File Report 2009–1004; U.S. Geological Survey, 2009.
  97. Gilliom R. J.; Alley W. M.; Gurtz M. E.. Design of the National Water-Quality Assessment Program; Occurrence and Distribution of Water-Quality Conditions, USGS Circular 1112; USGS, 1995. https://pubs.usgs.gov/circ/circ1112/ (last accessed Feb 08, 2024).
  98. Fenelon J. M.; Moore R. C. Transport of agrichemicals to ground and surface water in a small central Indiana watershed. J. Environ. Qual. 1998, 27 (4), 884–894. 10.2134/jeq1998.00472425002700040024x. [DOI] [Google Scholar]
  99. Puckett L. J.; Cowdery T. K.; McMahon P. B.; Tornes L. H.; Stoner J. D. Using chemical, hydrologic, and age dating analysis to delineate redox processes and flow paths in the riparian zone of a glacial outwash aquifer-stream system. Water Resour. Res. 2002, 38 (8), 9-1–9-20. 10.1029/2001WR000396. [DOI] [Google Scholar]
  100. Schilling K. E.; Jacobson P. J.; St Clair M.; Jones C. S. Dissolved phosphate concentrations in Iowa shallow groundwater. J. Environ. Qual. 2020, 49 (4), 909–920. 10.1002/jeq2.20073. [DOI] [PubMed] [Google Scholar]
  101. Saad D. A. Agriculture-related trends in groundwater quality of the glacial deposits aquifer, central Wisconsin. J. Environ. Qual. 2008, 37 (5), S-209–S-225. 10.2134/jeq2007.0053. [DOI] [PubMed] [Google Scholar]
  102. Browne B. A.; Guldan N. M. Understanding long-term baseflow water quality trends using a synoptic survey of the ground water-surface water interface, central Wisconsin. J. Environ. Qual. 2005, 34 (3), 825–835. 10.2134/jeq2004.0134. [DOI] [PubMed] [Google Scholar]
  103. Denver J. M.; Tesoriero A. J.; Barbaro J. R. Trends and transformation of nutrients and pesticides in a coastal plain aquifer system, United States. J. Environ. Qual. 2010, 39 (1), 154–167. 10.2134/jeq2009.0107. [DOI] [PubMed] [Google Scholar]
  104. Böhlke J. K.; Denver J. M. Combined use of groundwater dating, chemical, and isotopic analyses to resolve the history and fate of nitrate contamination in two agricultural watersheds, Atlantic Coastal Plain, Maryland. Water Resour. Res. 1995, 31 (9), 2319–2339. 10.1029/95WR01584. [DOI] [Google Scholar]
  105. Tesoriero A. J.; Spruill T. B.; Eimers J. L. Geochemistry of shallow ground water in coastal plain environments in the southeastern United States: implications for aquifer susceptibility. Appl. Geochem. 2004, 19 (9), 1471–1482. 10.1016/j.apgeochem.2004.01.021. [DOI] [Google Scholar]
  106. Messier K. P.; Wheeler D. C.; Flory A. R.; Jones R. R.; Patel D.; Nolan B. T.; Ward M. H. Modeling groundwater nitrate exposure in private wells of North Carolina for the Agricultural Health Study. Sci. Total Environ. 2019, 655, 512–519. 10.1016/j.scitotenv.2018.11.022. [DOI] [PMC free article] [PubMed] [Google Scholar]
  107. Tesoriero A. J.; Burow K. R.; Frans L. M.; Haynes J. V.; Hobza C. M.; Lindsey B. D.; Solder J. E. Using age tracers and decadal sampling to discern trends in nitrate, arsenic, and uranium in groundwater beneath irrigated cropland. Environ. Sci. Technol. 2019, 53 (24), 14152–14164. 10.1021/acs.est.9b03459. [DOI] [PubMed] [Google Scholar]
  108. Welch H. L.; Green C. T.; Coupe R. H. The fate and transport of nitrate in shallow groundwater in northwestern Mississippi, USA. Hydrogeol. J. 2011, 19 (6), 1239–1252. 10.1007/s10040-011-0748-8. [DOI] [Google Scholar]
  109. Barlow J. R. B.; Coupe R. H. Groundwater and surface-water exchange and resulting nitrate dynamics in the Bogue Phalia Basin in northwestern Mississippi. J. Environ. Qual. 2012, 41 (1), 155–169. 10.2134/jeq2011.0087. [DOI] [PubMed] [Google Scholar]
  110. Rosecrans C. Z.; Nolan B. T.; Gronberg J. M. Prediction and visualization of redox conditions in the groundwater of Central Valley, California. J. Hydrol. 2017, 546, 341–356. 10.1016/j.jhydrol.2017.01.014. [DOI] [Google Scholar]
  111. Knierim K. J.; Kingsbury J. A.; Belitz K.; Stackelberg P. E.; Minsley B. J.; Rigby J. R. Mapped predictions of manganese and arsenic in an alluvial aquifer using boosted regression trees. Groundwater 2022, 60 (3), 362–376. 10.1111/gwat.13164. [DOI] [PMC free article] [PubMed] [Google Scholar]
  112. Pace C.; Balazs C.; Bangia K.; Depsky N.; Renteria A.; Morello-Frosch R.; Cushing L. J. Inequities in drinking water quality among domestic well communities and community water systems, California, 2011–2019. Am. J. Public Health 2022, 112 (1), 88–97. 10.2105/AJPH.2021.306561. [DOI] [PMC free article] [PubMed] [Google Scholar]
  113. Li L.; Sullivan P. L.; Benettin P.; Cirpka O. A.; Bishop K.; Brantley S. L.; Knapp J. L. A.; van Meerveld I.; Rinaldo A.; Seibert J.; Wen H.; Kirchner J. W. Toward catchment hydro-biogeochemical theories. WIREs Water 2021, 8 (1), e1495 10.1002/wat2.1495. [DOI] [Google Scholar]
  114. Basu N. B.; Van Meter K. J.; Byrnes D. K.; Van Cappellen P.; Brouwer R.; Jacobsen B. H.; Jarsjö J.; Rudolph D. L.; Cunha M. C.; Nelson N.; Bhattacharya R.; Destouni G.; Olsen S. B. Managing nitrogen legacies to accelerate water quality improvement. Nat. Geosci. 2022, 15 (2), 97–105. 10.1038/s41561-021-00889-9. [DOI] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

es3c07576_si_001.pdf (1.2MB, pdf)

Articles from Environmental Science & Technology are provided here courtesy of American Chemical Society

RESOURCES