Skip to main content
NCBI home page
EPA Author Manuscripts logoLink to EPA Author Manuscripts
. Author manuscript; available in PMC: 2023 Jun 30.
Published in final edited form as: Water Res. 2022 Apr 9;218:118443. doi: 10.1016/j.watres.2022.118443

Regional patterns and drivers of total nitrogen trends in the Chesapeake Bay watershed: Insights from machine learning approaches and management implications

Qian Zhang 1,*, Joel T Bostic 2, Robert D Sabo 3
PMCID: PMC9743807  NIHMSID: NIHMS1808992  PMID: 35461100

Abstract

Anthropogenic nutrient inputs have led to nutrient enrichment in many waterbodies worldwide, including Chesapeake Bay (USA). River water quality integrates the spatial and temporal changes of watersheds and forms the foundation for disentangling the effects of anthropogenic inputs. We demonstrate with the Chesapeake Bay Non-Tidal Monitoring Network that machine learning approaches – i.e., hierarchical clustering and random forest (RF) classification – can be combined to better understand the regional patterns and drivers of total nitrogen (TN) trends in large monitoring networks, resulting in information useful for watershed management. Cluster analysis revealed regional patterns of short-term TN trends (2007–2018) and categorized the stations into three distinct trend clusters, namely, V-shape (n = 23), monotonic decline (n = 35), and monotonic increase (n = 26). RF models identified regional drivers of TN trend clusters by quantifying the effects of watershed characteristics (land use, geology, physiography) and major N sources on the trend clusters. Results provide encouraging evidence that improved agricultural nutrient management has resulted in declines in agricultural nonpoint sources, which in turn contributed to water-quality improvement in our period of analysis. Moreover, water-quality improvements are more likely in watersheds underlain by carbonate rocks, reflecting the relatively quick groundwater transport of this terrain. By contrast, water-quality improvements are less likely in Coastal Plain watersheds, reflecting the effect of legacy N in groundwater. Notably, results show degrading trends in forested watersheds, suggesting new and/or remobilized sources that may compromise management efforts. Finally, the developed RF models were used to predict TN trend clusters for the entire Chesapeake Bay watershed at the fine scale of river segments (n = 979), providing fine spatial information that can facilitate targeted watershed management, including unmonitored areas. More broadly, this combined use of clustering and classification approaches can be applied to other regional monitoring networks to address similar water-quality questions.

Keywords: water quality, anthropogenic sources, watershed management, cluster analysis, random forest, monitoring network

Graphical Abstract

graphic file with name nihms-1808992-f0001.jpg

1. Introduction

Nutrient enrichment is a major issue to many inland and coastal waterbodies worldwide (Boesch, 2019; Cloern, 2001; Malone and Newton, 2020). Excessive nutrients, primarily nitrogen and phosphorus, can originate from a variety of anthropogenic sources (e.g., wastewater treatment facilities, fertilizer, atmospheric deposition), whose relative contributions vary among watersheds (Boyer et al., 2002; Sabo et al., 2019; van Breemen et al., 2002). Rivers are integrators, connecting source zones and receiving waterbodies (e.g., lakes and estuaries). They manifest the spatial and temporal changes in watersheds and thus provide essential information for watershed management. River water-quality monitoring stations have been established in many regions to quantify nutrient loads and their temporal trends, which form the foundation for disentangling the effects of anthropogenic inputs.

One example of nutrient enrichment is Chesapeake Bay, the largest estuary in the United States. To alleviate this issue, coordinated reductions of nutrient and sediment loads from the Bay watershed began in the 1980s and are enforced through the 2010 Chesapeake Bay Total Maximum Daily Load (TMDL) (U.S. Environmental Protection Agency, 2010). Efforts have been made by the partnership to reduce nutrients from point sources (Boynton et al., 2008; Fisher et al., 2021; Zhang et al., 2021b), agricultural fertilizer and manure applications (Keisman et al., 2018), and atmospheric deposition (Burns et al., 2021; Eshleman and Sabo, 2016; Eshleman et al., 2013; Linker et al., 2013b; Sabo, 2018). Despite some documented progress in the restoration efforts (Lefcheck et al., 2018; Murphy et al., 2022; Murphy et al., 2011; Zhang et al., 2021; Zhang et al., 2018), continued nutrient reductions are imperative to achieve the TMDL goals. In this context, better understanding of regional patterns and drivers of riverine nitrogen trends is critical for effective watershed management.

Quantifying and understanding river water-quality trends have been an extensive area of research for decades, yet several typical challenges remain, which are relevant to Chesapeake Bay and elsewhere. First, water-quality trend studies often focus on one or a few monitoring locations, making conclusions difficult to generalize (Eshleman and Sabo, 2016; Eshleman et al., 2013). Watersheds are complex entities that often exhibit strong spatial heterogeneity. Consequently, river water quality is affected by a multitude of drivers related to nutrient sources, transport, and transformation, which can be challenging to uncover (Ator et al., 2020; Ator et al., 2019; Chanat and Yang, 2018). Nonetheless, much can be learned from the similarity in water-quality signals and the similarity in water-quality responses to natural and anthropogenic drivers, which is made possible by data from regional monitoring networks (Ator et al., 2020; Ator et al., 2019; Chanat and Yang, 2018). In this regard, clustering approaches can be especially useful for categorizing regional patterns of water-quality trends. Such approaches can reveal similarity in water-quality trends among monitoring stations in the network. In addition, the assigned trend clusters can be further analyzed toward identifying likely drivers of trends, leading to insights that are more generalizable from a whole watershed perspective, relative to studies focusing on one or a few locations.

Another challenge in understanding water-quality trends is that trend attribution analyses do not often evaluate all major input sources, leading to potentially inaccurate inferences from the individual studies or even contradictory inferences among studies focused on different sources. In the case of Chesapeake Bay, temporal trends of total nitrogen (TN) have been attributed to a diverse list of often disparate factors, including atmospheric deposition (Ator et al., 2020; Burns et al., 2021; Eshleman and Sabo, 2016; Eshleman et al., 2013; Hyer et al., 2021), point sources (Ator et al., 2020; Boynton et al., 2008; Fisher et al., 2021), and land-based nonpoint sources (Chanat and Yang, 2018). These disparities are related to the differences in the methods and data used (e.g., N sources, spatial scale, temporal coverage), making it challenging to uncover generalizable drivers to the entire watershed. Therefore, it will be beneficial to systematically evaluate all major sources of N for the majority of the Bay watershed with a consistent period of analysis. Such sources include atmospheric deposition, point sources, and agricultural fertilizer and manure, as well as agricultural surplus that has reportedly declined across the watershed but has not yet been incorporated into any attribution analysis (Sabo et al., 2022). In this regard, classification approaches such as random forest (RF) can be especially useful for relating TN trend clusters, which is a categorical variable, to potential drivers that include watershed characteristics and major N sources.

A third challenge is that river monitoring networks do not necessarily cover the entire watershed, resulting in missing information in ungauged subwatersheds. In this regard, RF models, once developed based on the monitored areas, can be used to predict water-quality trends in the unmonitored areas. In addition, water-quality trend analyses are often aimed at multi-decadal scales consistent with the temporal span of the monitoring network (Ator et al., 2019; Chanat and Yang, 2018; Eshleman et al., 2013; Zhang et al., 2015), which automatically disqualify recently established stations. Therefore, an analysis of short-term trends (~12 years) can not only leverage valuable data from recently established stations, but also provide current information for facilitating watershed management.

The main aim of our research was to address the above challenges and demonstrate the utility of advanced machine learning approaches -- more specifically, the combined use of hierarchical clustering and RF classification -- to reveal regional patterns and drivers of nitrogen trends. We use the Chesapeake Bay as a case study, specifically the 100+ stations comprising the Chesapeake Bay Non-Tidal Monitoring Network (NTN) (Moyer and Langland, 2020; Tango and Batiuk, 2016). This research involved three objectives that were achieved in a sequential manner: (1) use of hierarchical clustering to categorize TN trends from 2007–2018 at the NTN stations into distinct clusters, (2) development of RF models to identify the most influential explanatory variables for the TN trend clusters, and (3) application of the RF models to predict TN trend clusters for the entire Bay watershed at the fine scale of river segments (n = 979, mean area = 168 km2) (Figure S1). Overall, this research provides new information to the Chesapeake Bay management community regarding regional patterns and drivers of short-term TN trends.

In the field of water resources, clustering approaches have been used in studies of stream health (Sanchez et al., 2015), flood frequency (Sun et al., 2015), and watershed characteristics (Wolfe et al., 2019). RF and other classification approaches have been used in studies of water level (Guyennon et al., 2021), flow indices (Prieto et al., 2019), ground contamination (Pennino et al., 2020; Tesoriero et al., 2017), and river water quality (Lin et al., 2021; Sabo, 2018). However, to our knowledge, a combined use of these two types of approaches has not yet been implemented toward understanding regional patterns and drivers of water-quality trends, and this research is the first such attempt. Although this research is aimed toward the Chesapeake Bay watershed, the combined use of clustering and classification approaches can be applied to other regional monitoring networks to address similar water-quality questions.

2. Data and Methods

2.1. NTN stations and data

The Chesapeake Bay NTN contains 115 monitoring stations (Moyer and Langland, 2020; Tango and Batiuk, 2016). These NTN stations have been sampled using standardized protocols and quality-assurance procedures designed to compute pollutant loads and assess changes in pollutant loads over time (Tango and Batiuk, 2016). For TN and other major water-quality constituents, routine samples are collected monthly, and eight additional storm-event samples are collected per year to obtain at least 20 samples per year (Tango and Batiuk, 2016).

Annual flow-normalized loads (FNLs) of TN were estimated using the Weighted Regressions on Time, Discharge, and Season (WRTDS) method (Hirsch et al., 2010) and obtained from the U.S. Geological Survey (Moyer and Langland, 2020). FNLs better account for the interannual variability in streamflow and are a standard product used by the Chesapeake Bay Program partnership to communicate water-quality trends across the Chesapeake Bay watershed. The appropriateness of the WRTDS method for riverine trend analysis, specifically its flow-normalization approach, has been well documented in the literature (Chanat et al., 2016; Hirsch et al., 2010; Moyer et al., 2012) and has been used by many national and regional water-quality investigations (Dupas et al., 2020; Murphy and Sprague, 2019; Oelsner and Stets, 2019; Stets et al., 2020; Zhang and Blomquist, 2018). Our work focuses on annual FNLs of TN at the NTN stations for the period of 2007–2018, which are available at 84 stations. The year 2007 was selected because 34 stations were added to the network in 2005–2007. Annual FNLs at three example stations are presented in Figure 1 (a–c).

Figure 1.

Figure 1.

Open in a new tab

Time series of flow-normalized (FN) total nitrogen loads at three selected monitoring stations, showing (a-c) original loads, (b-f) standardized loads, and (g-i) Euclidean distances of the standardized loads for each pair of stations. The Euclidean distance between two time series of standardized load is calculated as the square root of the sum of the squared differences (dashed, red lines) for each year. Based on the Euclidean distances, Patuxent River and Opequon Creek are most similar to each other, whereas Patuxent River and Choptank River are most dissimilar.

2.2. Hierarchical cluster analysis

Clustering approaches can reveal similarity in TN trends among the monitoring stations, and the assigned trend clusters can be further analyzed toward identifying likely drivers of trends, leading to insights that are more generalizable from a whole watershed perspective. In this work, agglomerative hierarchical cluster analysis was used to categorize FNLs at the 84 NTN stations into groups that have similar temporal trends. Prior to cluster analysis, FNLs at each station were standardized to have a mean of zero and a standard deviation of one to facilitate comparison of trends as opposed to magnitudes:

FNLi,t,standardized=(FNLi,t−mean(FNLi))/sd(FNLi) (1)

where i represents a station and t represents year which ranges between 2007 and 2018. For the three example stations, the standardized FNLs show monotonic declines in Patuxent River and Opequon Creek but a monotonic increase in Choptank River (Figure 1, d–f).

Dissimilarity between the standardized FNLs at the individual stations was measured using Euclidean distance and linkage between clusters was determined using Ward’s minimum variance method (Ward, 1963). The Euclidean distance between two time series of FNLs is calculated as the square root of the sum of their squared differences in each year between 2007 and 2018 (Figure 1, g–i). Based on the Euclidean distances, Patuxent River and Opequon Creek are most similar to each other, while Patuxent River and Choptank River are most dissimilar.

The cluster analysis was performed in R using the “hclust” function in the “stats” R package with the “ward.D2” method (R Core Team, 2019). The optimal number of clusters (3) was determined by the silhouette method and visual inspection of the dendrogram (Kaufman and Rousseeuw, 1990). The dendrogram, which shows the hierarchical relationship between the stations, is presented in Figure 2. Moreover, the overall trends and spatial distribution of the clusters are presented in Figure 3 and Figure 4, respectively.

Figure 2.

Figure 2.

Open in a new tab

Dendrogram from the hierarchical cluster analysis, showing cluster assignments of total nitrogen trends for the Chesapeake Bay Non-Tidal Monitoring Network stations. Stations closer to each other on the dendrogram have more similar temporal trends of total nitrogen loads.

Figure 3.

Figure 3.

Open in a new tab

The three clusters of short-term (2007–2018) trend of total nitrogen, as obtained from the hierarchical cluster analysis. For each cluster, the cluster-level trend is shown with a thick, colored line, whereas the trends for the individual stations are shown with thin, black lines. Note that TN loads have been standardized for each station prior to the cluster analysis.

Figure 4.

Figure 4.

Open in a new tab

Maps showing (a) the Chesapeake Bay and its major basins and (b) the clusters of the Non-Tidal Monitoring Network stations. The bottom panels (c-e) are provided to better show contrasting patterns of nested watersheds. Each open circle represents one station (Figure 2). Inset shows the temporal trend of each cluster (Figure 3).

As an unsupervised method, cluster analysis cannot be validated using traditional approaches (e.g., k-fold cross validation) since the data have no predefined labels (classes). Nonetheless, validation is still important because the clustering results depend on choices on similarity measure, clustering algorithm, and linkage method. Toward that end, we adopted the “resampling and re-classifying procedure” of Wolfe et al. (2019) to assess the robustness of our clustering results (Figure S2). Specifically, the 84 stations were divided into 12 independent subsets, each containing seven stations. Subsequently, 12 iterations of the cluster analysis were run. For each iteration, one subset of the stations (n = 7) was removed from the original data and the remaining stations (n = 77) were reanalyzed using the same clustering procedure. The number of clusters was set at three to be consistent with the full-data analysis. The cluster assignments from these 12 iterations were then compared with those from the full-data analysis to assess agreement of the assigned clusters (Figure S2).

2.3. RF model features

RF was chosen to identify factors that affect the TN trend clusters for four reasons. First, RF has been successful as a general-purpose classification and regression approach (Biau and Scornet, 2016) and has been used in many water resources studies (see references listed above). It is appropriate for our work since the trend clusters are a categorical variable. Second, RF is comparatively less sensitive to outliers and can efficiently handle non-linear variables (Breiman, 2001; Hastie et al., 2009). Third, compared with other approaches, e.g., Classification and Regression Tree (Breiman, 1984), RF can produce generally more stable results due to its bagging algorithm that builds multiple trees and glues them together (Hastie et al., 2009). Fourth, RF can reveal additional insights on the effects of the explanatory variables by quantifying the relative importance of the variables and the marginal effects of the variables (Breiman, 2001; Hastie et al., 2009).

A total of 16 explanatory variables (“features”) were considered for predicting the trend clusters, i.e., the response variable (Table 1). These candidate features fall into the categories of land use (n = 4), geology (n = 1), physiography (n = 5), and N source trends (n = 6). These variables were selected because their effects on N loads and trends have been documented for the Chesapeake Bay watershed and elsewhere (Ator et al., 2020; Ator et al., 2019; Chanat and Yang, 2018; Clune et al., 2020; Djodjic et al., 2021; Dupas et al., 2016; Eshleman et al., 2013; Hyer et al., 2021; Terziotti et al., 2017). We note that the TN trend clusters may be further explained by other variables such as management practice implementation (Lintern et al., 2020; Sekellick et al., 2019), legacy N (Chang et al., 2021; Sanford and Pope, 2013), riparian buffers (Golden et al., 2019; Lowrance et al., 1997), and landscape configuration (Casquin et al., 2021). However, these variables were not considered in this research due to the lack of data for the NTN stations and the river segments.

Table 1.

Summary of the explanatory variables and their statistical difference among the three clusters of monitoring stations. For land use, geology, and physiography variables, the median of annual values in 2007–2018 was used. Each feature was separately evaluated against the response variable using the one-way analysis of variance (ANOVA) test. A feature is statistically different among the three clusters if the reported p-value is less than 0.1 (bold).

Explanatory variable Definition ANOVA p-value
Land use (n = 4)
 Natural_pct Fraction of natural land, percent 0.00013
 Crop_pct Fraction of cropland, percent 0.023
 Pasture_pct Fraction of pastureland, percent 0.82
 Hay_pct Fraction of hay land, percent 0.27
Geology (n = 1)
 Carb_pct Fraction of carbonate geology, percent 0.089
Physiography (n = 5)
 Appalachian_pct Fraction of Appalachian physiography, percent 0.12
 BlueRidge_pct Fraction of Blue Ridge physiography, percent 0.78
 ValleyRidge_pct Fraction of Valley and Ridge physiography, percent 0.69
 Piedmont_pct Fraction of Piedmont physiography, percent 0.13
 Coastal_pct Fraction of Coastal Plain physiography, percent 0.61
Mann-Kendall (MK) trends of N source (n = 6)
 PointSource MK MK trend of point source for the period of 2007–2018, scaled by the period-of-record median 0.33
 Deposition MK MK trend of atmospheric deposition for the period of 1997–2018, scaled by the period-of-record median 0.038
 Fertilizer MK MK trend of agricultural fertilizer for the period of 1997–2018, scaled by the period-of-record median 0.036
 Manure MK MK trend of manure for the period of 1997–2018, scaled by the period-of-record median 0.28
 AgInput_MK MK trend of agricultural input for the period of 1997–2018, scaled by the period-of-record median 0.81
 AgSurplus_MK MK trend of agricultural surplus for the period of 1997–2018, scaled by the period-of-record median 0.14
Open in a new tab

For each station, land use variables, expressed in percent, were quantified using the Chesapeake Bay Program’s Chesapeake Assessment Scenario Tool (Chesapeake Bay Program, 2017). Geology variables (in percent) were obtained from the U.S. conterminous wall-to-wall anthropogenic land use trends dataset (Falcone, 2015). Physiographic region variables (in percent) were obtained from the U.S. Geological Survey (Wieczorek et al., 2018). For each variable above, the median of annual values in the period of 2007–2018 was used for each station. In addition, for each station, N source data were obtained from the Chesapeake Assessment Scenario Tool for applicable river segments, which were summed up to obtain the annual mass rate (in kg N/year) for each major source. These annual time series were scaled by their respective long-term medians to facilitate comparison of interannual trends. Then, Mann-Kendall trends (Kendall, 1975) and Sen’s slopes (Sen, 1968) were computed for point source (“PointSource_MK”), atmospheric deposition (“Deposition_MK”), manure (“Manure_MK”), agricultural fertilizer (“Fertilizer_MK”), agricultural input (“AgInput_MK”; AgInput = agricultural fertilizer + agricultural manure + legume fixation), and agricultural surplus (“AgSurplus_MK”; AgSurplus = AgInput + atmospheric deposition to agricultural land − crop removal), using the “mannKen” function in the “wq” R package (Jassby and Cloern, 2016). For point source, MK trends were computed for the 2007–2018 period, consistent with the period of the response variable. For the other sources, MK trends were computed for the 1997–2018 period to allow for a 10-year extension that may account for lag effects (Chang et al., 2021; Chen et al., 2014; Sabo, 2018). Among three options of lag tested (i.e., 10-year, 5-year, and 0-year), the 10-year lag was determined to be most effective in modeling the response variable (i.e., clusters).

Due to missing values in some features, the number of records for the RF analysis reduced from 84 to 78 stations (see Table S1 for full data). Each candidate feature was separately evaluated against the response variable using the one-way analysis of variance (ANOVA) test to examine if each feature had significant differences among the three clusters (Figure S3). Moreover, all candidate features were used to build a base model using the “train” function in the “caret” R package (Kuhn et al., 2020) to quantify the relative importance of these features on the clusters (Figure S4a). Furthermore, the Spearman’s correlation was calculated for each pair of features (Figure S4b), which is a nonparametric measure of rank correlation (Spearman, 1904). Lastly, for the base RF model (Figure S5) and subsequently selected RF models (Figure 5), partial dependence plots were developed using the “partialPlot” function in the “randomForest” R package (Liaw and Wiener, 2002) to show the marginal effect of each feature in the model on the probability of predicted cluster (i.e., the relationship between the predicted response and each feature while holding all other features constant).

Figure 5.

Figure 5.

Open in a new tab

Partial dependence plots for nine features in the optimal random forest models, showing their marginal effect on the probability of cluster 2. Models A-C include five, four, and six features, respectively. See Tables 1–2 for the selected features and their units.

2.4. Exhaustive search algorithm

To determine the optimal RF model(s), an exhaustive search algorithm was developed and implemented in R to evaluate all possible combinations of the candidate features. To constrain model complexity, this algorithm was set to allow no more than six features in the RF model. Experimentation showed that the model accuracy reached plateau at about six features and additional features failed to improve the model accuracy (see Figure S6 for an example). For each candidate model, RF was run using the “randomForest” function in the “randomForest” R package (Liaw and Wiener, 2002) with the same parameterization (i.e., ntree = 1,000, nodesize = 1), and the model performance was evaluated using the out-of-bag (OOB) accuracy (Hastie et al., 2009):

OOBaccuracy=100−OOBerror (2)

Similar to hold-out validation techniques, the OOB accuracy is a powerful validation technique used especially for RF models (Hastie et al., 2009). Specifically, the original data are resampled to form bootstrap samples, which are used for model training. Records that are not contained in the bootstrap samples are called OOB samples, which are used for model validation by calculating the OOB error (i.e., the proportion of incorrect classifications) (Hastie et al., 2009).

Three optimal models with the highest OOB accuracies were selected by the exhaustive search algorithm (Table 2), all of which showed better performance than the base model (Figure S7a). To further evaluate these optimal models, leave-one-out cross-validation (LOOCV) was conducted, which also showed better performance by the optimal models than the base model (Figure S7b).

Table 2.

The three optimal random forest models, as selected by the exhaustive search algorithm. Models A, B, and C had the highest out-of-bag (OOB) accuracy for cluster 3, 2, and 1, respectively.

Model Model form OOB accuracy, percent
Overall Clusterl Cluster2 Cluster3
A Class ~ Natural_pct + Fertilizer_MK + ValleyRidge_pct + Deposition_MK + Carb_pct 70.5 66.7 68.8 76.0
B Class ~ AgSurplus_MK + Fertilizer_MK + Deposition_MK + Natural_pct 70.5 66.7 75.0 68.0
C Class ~ BlueRidge_pct + Deposition_MK + Coastal_pct + Crop_pct + Fertilizer_MK + Natural_pct 69.2 81.0 65.6 64.0
Open in a new tab

2.5. Ensemble model approach

The optimal RF models were applied to predict the cluster assignment for each NTN station. In addition to the predicted classes, RF also reported the probability associated with each class, as determined by the proportion of votes of the trees in the ensemble. The probability associated with the cluster was used as a measure of likelihood. Specifically, a predicted cluster is deemed to have a high likelihood (if probability ≥ 0.667), a medium likelihood (if 0.5 ≤ probability < 0.667), or a low likelihood (if probability < 0.5).

An ensemble model approach was adopted to combine the strengths of the selected optimal models (Zhang et al., 2021a). Specifically, predictions from the optimal models were compared for each station and the prediction with the highest likelihood was always selected by the ensemble model approach (Figure S8a; see Table S2 for full data).

2.6. RF prediction for the entire watershed

The optimal RF models were applied to the entire Chesapeake Bay watershed to predict TN trend clusters at a finer resolution that is directly relevant for watershed management. For this purpose, we used the river segments (n = 979) of the Chesapeake Bay Program’s Watershed Model (Chesapeake Bay Program, 2017), which are subwatersheds of Chesapeake Bay. Unlike the NTN stations, the river segments have similar areas (mean = 168 km2) and they do not overlap with each other (Figure S1). For each river segment, the TN trend cluster was predicted in two steps. First, the model features were prepared in the same manner as the NTN stations -- see Figure 6 and Figure S9 for maps of these features. Second, the three optimal RF models were applied to these features to predict the trend cluster and compute the likelihood associated with each model’s prediction. The prediction with the highest likelihood was recorded by the ensemble model approach (Figure S8b; see Table S3 for full data). Note that these model predictions assume that subwatersheds should show similar riverine TN trends if they have similar watershed characteristics (land use, geology, physiography) and similar trends of N sources. This is reasonable because many of the river segments are encompassed by the NTN watersheds.

Figure 6.

Figure 6.

Open in a new tab

Map of four selected features at the scale of river segments. See Figure S9 for the other selected features and Table 1 for feature definition.

3. Results and Discussion

3.1. Cluster analysis revealed distinct TN trend clusters (Objective 1)

Three clusters of annual flow-normalized TN loads were identified by the cluster analysis (Figure 2; see Table S1 for full data). These clusters show unique short-term (2007–2018) trends of TN load (Figure 3). Cluster 1 (n = 23) has a V-shape trend, with an inflection point around 2011. Because this inflection point coincides with an extremely wet year (2011) due to Hurricane Irene and Tropical Storm Lee, one possible explanation is that increasing amounts of N were mobilized and exported to streams in the following years (Vidon et al., 2018). Cluster 2 (n = 35) shows a monotonic decline, whereas cluster 3 (n = 26) shows a monotonic increase. These two clusters represent improving and degrading trends, respectively. Note that these clusters focus on TN temporal trends as opposed to magnitudes. For each pair of NTN stations, the clustering approach quantified the similarity of their full time series using the Euclidean distance (e.g., Figure 1, g–i), thereby dampening the influence of any extreme values in the time series. Moreover, the sensitivity analysis showed that cluster assignment was consistent among all iterations for almost all watersheds, lending confidence in the clustering results (Figure S2).

The spatial distribution of the three clusters is shown in Figure 4. Perhaps the most striking pattern is that cluster 2 watersheds tend to be major agricultural regions in and around the Great Appalachian Valley, including the Shenandoah Valley, as well as the lower Susquehanna River region. In comparison, cluster 1 watersheds tend to be distributed in more forested areas in the western and northern parts of the watershed, whereas cluster 3 watersheds are more scattered.

3.2. RF models identified regional drivers of TN trend clusters (Objective 2)

The exhaustive search algorithm identified three optimal RF models with the similar overall OOB accuracy, i.e., 70.5%, 70.5%, and 69.2% (Table 2, Figure S7). These three models had the highest accuracy for cluster 1 (76%), cluster 2 (75%), and cluster 3 (81%), respectively, indicating that each of the three models settled on a specific set of features that are most effective in explaining one of the three clusters (Table 2). These models have similar forms, as they all include Natural_pct, Deposition_MK, Fertilizer_MK, and a few other variables that represent agriculture, geology, and/or physiography.

The ensemble model approach presents an advancement that combines the strength of all three optimal RF models (Figure S8a). Overall, the predicted clusters for the NTN stations by the ensemble model (see Table S2 for full data) is quite consistent with the direct result of the cluster analysis (Figure 4).

Several candidate features showed statistically significant differences among the three clusters based on the ANOVA analysis, namely, Natural_pct, Deposition_MK, Fertilizer_MK, Crop_pct, and Carb_pct (Table 1). In addition, the base RF model with all candidate features showed that the most important features were Natural_pct, Deposition_MK, Fertilizer_MK, PointSource_MK, AgSurplus_MK, Manure_MK, and Crop_pct (Figure S4a). According to the base model, watersheds with declines in nitrogen loads (i.e., cluster 2) generally correspond to watersheds with a lower percentage of natural cover, a higher percentage of cropland, a smaller decline in atmospheric N deposition, a greater decline in agricultural fertilizer, and a larger percentage of carbonate geology (Figure S3). Moreover, Spearman’s correlation between all the predictive features is generally low, suggesting that these features provide useful and unique information toward explaining the clusters (Figure S4b).

Compared with the ANOVA analysis above, partial dependence plots from the RF models provided a simultaneous and more rigorous assessment of the marginal effects of each feature, thereby enhancing our capability to capture the effects of regional drivers of TN trends (Figure 5). For brevity, we focus on cluster 2 (i.e., declining TN loads) and we note that inference of feature effects is generally consistent between cluster 2 and cluster 3. As shown by the partial dependence plots (Figure 5), cluster 2 is more likely to occur in watersheds with a small value of Nature_pct, a large value of Crop_pct, a large value of Carb_pct, or a small value of Coastal_pct. In terms of sources, cluster 2 is more likely to occur in watersheds with large declines in Fertilizer_MK or AgSurplus_MK. However, the opposite effect is surprisingly observed with Deposition_MK. These relationships are consistent with the base RF model in terms of marginal effects (Figure S5). These results are particularly useful for making informed management decisions when choosing priority watersheds to achieve water-quality goals as well as identifying emerging risks, which are summarized below.

Message 1. Agricultural nutrient management contributed to water-quality improvement.

Our results show that declining TN trends (cluster 2) are more likely to occur in watersheds with large declines of agricultural nonpoint sources (i.e., AgSurplus_MK and Fertilizer_MK) (Figure 5). It has been previously documented that declining point sources and atmospheric deposition improved water quality across the Chesapeake Bay watershed, but agricultural nonpoint sources were thought to have limited progress. However, we note that these studies focused on multi-decadal scales (i.e., 1980s – 2010s) and did not explicitly account for agricultural surpluses (AgSurplus_MK), which indeed have been declining since the 1980s (Sabo et al., 2022). Besides being more pertinent for communicating progress to decision makers and stakeholders (Sabo et al., 2021), agricultural surplus is likely a more relevant and temporally variable metric that captures the impacts of agriculture on N export compared to the more temporally static metrics (e.g., land use, geology, physiography) (Lin et al., 2021; Sabo, 2018). By explicitly incorporating agricultural surpluses into the RF models, this research provides a new finding toward explaining TN trends. Specifically, our results associated with AgSurplus_MK suggest that TN loads in agricultural basins, with historically high surpluses, are more responsive to declines in agricultural surpluses than previously described (Sanford and Pope, 2013; Van Meter et al., 2018). Overall, these results provide encouraging evidence to watershed managers that efforts to reduce agricultural nonpoint sources (e.g., nutrient management plans) contributed to water-quality improvement in our period of analysis. However, more recent short-term trend analysis indicates that this promising change in agricultural surplus is beginning to reverse in some areas of the Bay watershed (Sabo et al., 2022). The findings of our study are therefore relevant because river water-quality depends on the progress of reducing agricultural surpluses. In this regard, efforts aimed toward enhancing nutrient use efficiency in agricultural production will be instrumental for further the restoration efforts (Chang et al., 2021; Sabo et al., 2022).

Message 2. Water-quality improvements are more likely in carbonate areas but less likely in Coastal Plain areas.

Our results show that cluster 2 is more likely to occur in watersheds with a large value of Carb_pct (Figure 5). This is consistent with recent findings that TN loads have declined from cropland to streams in carbonate areas of the Chesapeake Bay watershed (Ator et al., 2020; Ator et al., 2019). Due to solution cavities, carbonate terrain tends to have relatively quick infiltration and faster groundwater transport than other geologic settings, resulting in relatively young groundwater ages (Ator et al., 2020; Terziotti et al., 2017). Consequently, historic N loads in groundwater from carbonate terrain were much higher than other geologic settings during periods of greater agricultural N inputs and surpluses. Conversely, as agricultural inputs and surpluses decrease, water-quality improvements can be achieved more quickly (i.e., a shorter lag) in watersheds underlain by carbonate. By contrast, cluster 2 is less likely to occur in watersheds with a large value of Coastal_pct (Figure 5). This is consistent with recent findings that the low-lying Coastal Plain areas have large accumulations of legacy N in the groundwater due to historical agricultural inputs and surpluses combined with longer groundwater residence times (Sanford and Pope, 2013). Consequently, watersheds in these areas are expected to require a much longer time to achieve water-quality improvements (Chang et al., 2021; Sanford and Pope, 2013; Zhang et al., 2015).

Message 3. Degrading trends in forested areas are concerning and suggest new and/or remobilized N sources.

Our results show that improving trends (cluster 2) are less likely to occur in watersheds with larger proportions of forest (Natural_pct) and/or greater declines in atmospheric deposition (Deposition_MK) (Figure 5). Consistent with this pattern, degrading trends (cluster 3) are more likely to occur in watersheds with larger Natural_pct. This finding is unexpected considering recent reports of declining nitrate export in some predominantly forested watersheds from the 1980s to 2012 (Eshleman and Sabo, 2016; Eshleman et al., 2013) and given recent evidence of N limitation in forests in similar periods (McLauchlan et al., 2017; Sabo et al., 2020). A key distinction between our work and the studies cited above is that we focus on the more recent period (i.e., 2007–2018). Therefore, the degrading trends in the forested watersheds is likely a recent phenomenon, which may impede the progress of watershed management toward the Bay restoration. While the exact mechanisms driving the degrading trends cannot be inferred from the RF models, these trends can likely be explained by recent increases in N sources other than atmospheric deposition, and additional data and literature provide clues to at least two explanations. One apparent explanation is increasing N inputs to non-forest regions of the predominantly forested watersheds. In this regard, independent data show that forests have generally been lost at the expense of development across the Bay watershed, with natural area reduced by ~2,400 km2 but developed area expanded by ~8,100 km2 between 1985 – 2018 (Chesapeake Bay Program, 2017). Such changes can result in increased N loads because developed lands typically export much more N than forested lands on a per-unit-area basis (Ator et al., 2011; Jordan et al., 1997; Zhang et al., 2016). Associated with these land-use trends, recently published nutrient inventories revealed increasing inputs of septic N and urban fertilizer N – sources generally associated with human development (Sabo et al., 2022). Therefore, expansion of human development into predominantly forested areas can at least partially explain the increased TN trend in the forested watersheds. Another more speculative explanation is mobilization of N from internal pools, which is associated with deacidification of forest soils previously impacted by acid deposition (Groffman et al., 2018; Newcomer et al., 2021). Recent work suggests that decomposition in some forests is increasing as soils recover from acid deposition (Lawrence et al., 2020; Sabo et al., 2020; Scanlon et al., 2021), which can result in increased stream TN loads despite increased vegetative uptake (Marinos et al., 2018; Rosi-Marshall et al., 2016).

3.3. RF models predicted TN trend clusters for the entire Bay watershed (Objective 3)

A strength of optimal RF models is the ability to predict clusters of TN trends for the entire Chesapeake Bay watershed at the finer spatial scale of river segments. The spatial distribution of the selected features at the scale of river segments is shown in Figure 6 and Figure S9. In general, AgSurplus_MK and Fertilizer_MK show declines in most river segments over the period of record (Figure 6, a–b). Similarly, Deposition_MK shows declines in all river segments in the watershed (Figure S9). Natural_pct is the highest in the uplands and mountain ridges of the watershed (Figure 6c). Carb_pct is the highest along the valley floor of the Great Appalachian Valley (Figure 6d).

Among the 979 river segments, 292 (30%), 392 (40%), and 295 (30%) were classified as cluster 1, cluster 2, and cluster 3, respectively (Figure 7; see Table S3 for full data). For all the predicted clusters, 85% had high or medium likelihood. In general, cluster 2 segments are located in the lower Susquehanna River, the Eastern Shore, the Maryland Western Shore, and the Shenandoah Valley/Great Appalachian Valley. Cluster 3 segments are in the western Susquehanna River, the lower Rappahannock River, and the James River. Cluster 1 segments are on the northern and western edges of the watershed that are mainly forested and mountainous. These spatial patterns are broadly consistent with the patterns of the NTN stations (Figure 4). Notably, the three clusters are not evenly distributed in terms of proximity to the Bay (Figure 8). In general, cluster 2 dominates the areas closer to the Bay. As distance to the Bay increases, the three clusters have more comparable cumulative areas.

Figure 7.

Figure 7.

Open in a new tab

Map of trend clusters for the Chesapeake Bay watershed at the scale of river segments, as predicted by the random forest ensemble model. For each cluster, dark color indicates high likelihood (≥ 0.667), whereas light color indicates low likelihood (< 0.5). Inset shows the temporal trend of each cluster (Figure 3). State and county boundaries are shown.

Figure 8.

Figure 8.

Open in a new tab

Cumulative area of the river segments for each cluster as a function of distance to the Bay.

These predictions are encouraging in two aspects. First, a large proportion (40%) of the river segments have declining TN trends (cluster 2), and the majority (58%) of these predictions have high likelihoods (≥ 0.667). Second, these declining TN trend watersheds are much closer to the Bay than the V-shaped (cluster 1) or increasing-trend (cluster 3) clusters. The proximity of declining TN trend watersheds (cluster 2) to the Bay can boost restoration efforts because water-quality improvements in areas closer and with a direct connection to the Bay are expected to have a more immediate impact on the estuarine water quality than upland areas (Chesapeake Bay Program, 2017; Clune et al., 2021; Linker et al., 2013a; Shenk and Linker, 2013). Overall, these predictions are useful for understanding likely TN trends across the Bay watershed, including new information on TN trends and responses in the unmonitored areas.

4. Conclusions

We demonstrate that machine learning approaches – i.e., hierarchical clustering and random forest – can be combined to better understand the regional patterns and drivers of TN trends in large river monitoring networks, resulting in information useful for watershed management. To our knowledge, a combined use of these two types of machine learning approaches has not yet been implemented toward understanding regional patterns and drivers of water-quality trends. Although this research is aimed toward the Chesapeake Bay watershed, the combined use of clustering and classification approaches can be applied to other regional monitoring networks to address similar water-quality questions. Our work has several contributions from both methodological and management perspectives, which are summarized below.

  1. Cluster analysis revealed distinct TN trend clusters (Objective 1). Watersheds are complex entities that exhibit strong spatial heterogeneity, but cluster analysis can help reveal similarities. For the NTN stations, TN trends show three distinct clusters, which formed the response variable of the RF models toward identifying trend drivers, leading to insights that are more generalizable from a whole watershed perspective.

  2. RF models identified regional drivers of TN trend clusters (Objective 2). The exhaustive search algorithm developed here can provide an objective way of identifying the optimal combination of explanatory variables for RF or other classification approaches. The optimal RF models selected by this search algorithm can be used to quantify the relative importance and marginal effects of explanatory variables toward cluster assignments, including management-relevant variables that are actionable. In addition, our RF models explicitly incorporated temporal trends in agricultural fertilizer, manure, and agricultural input as well as agricultural surplus. Results provide encouraging evidence that improved nutrient management has resulted in declines in agricultural nonpoint sources, which in turn contributed to water-quality improvement in our period of analysis. Results also show that water-quality improvements are more likely in watersheds underlain by carbonate rocks, reflecting the relatively quick groundwater transport of this terrain. By contrast, water-quality improvements are less likely in watersheds in the Coastal Plain, reflecting the effect of legacy N in groundwater. Lastly, results show degrading trends in forested watersheds, suggesting new and/or remobilized sources of N that may compromise the restoration efforts.

  3. RF models predicted TN trend clusters for the entire Bay watershed (Objective 3). These predictions for river segments provide fine spatial information for the entire watershed (including unmonitored areas), which provide new information on TN trends and responses in the unmonitored areas.

Supplementary Material

s2
s3
s4
s1

Acknowledgments

This work was supported by the U.S. Environmental Protection Agency (to QZ) under Grant No. 07-5-230480) and a National Science Foundation Graduate Research Fellowship (to JTB) under Grant No. 1840380. We thank Dr. Tom Barnum and Dr. Jiajia Lin for reviewing an early version of this manuscript. The views expressed in this article are those of the authors and do not necessarily represent the views or policies of the U.S. Environmental Protection Agency. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government.

Data availability

All data used in this research are described in the Methods section.

  • The annual flow-normalized TN loads at the NTN stations are available from the U.S. Geological Survey (Moyer and Langland, 2020).

  • The land use and N input data for the NTN stations are available from the Chesapeake Assessment Scenario Tool (Chesapeake Bay Program, 2017).

  • The geology data for the NTN stations are available from the U.S. conterminous wall-to-wall anthropogenic land use trends dataset (Falcone, 2015).

  • The physiography data for the NTN stations are available from the U.S. Geological Survey (Wieczorek et al., 2018).

References

  1. Ator SW, Blomquist JD, Webber JS, Chanat JG, 2020. Factors driving nutrient trends in streams of the Chesapeake Bay watershed. Journal of Environmental Quality 49(4), 812–834. [DOI] [PubMed] [Google Scholar]
  2. Ator SW, Brakebill JW, Blomquist JD, 2011. Sources, fate, and transport of nitrogen and phosphorus in the Chesapeake Bay watershed: An empirical model, U.S. Geological Survey, Reston, VA, p. 27. [Google Scholar]
  3. Ator SW, García AM, Schwarz GE, Blomquist JD, Sekellick AJ, 2019. Toward explaining nitrogen and phosphorus trends in Chesapeake Bay tributaries, 1992–2012. Journal of the American Water Resources Association 55(5), 1149–1168. [Google Scholar]
  4. Biau G, Scornet E, 2016. A random forest guided tour. Test 25(2), 197–227. [Google Scholar]
  5. Boesch DF, 2019. Barriers and bridges in abating coastal eutrophication. Frontiers in Marine Science 6, 123. [Google Scholar]
  6. Boyer EW, Goodale CL, Jaworski NA, Howarth RW, 2002. Anthropogenic nitrogen sources and relationships to riverine nitrogen export in the northeastern USA. Biogeochemistry 57(1), 137–169. [Google Scholar]
  7. Boynton WR, Hagy JD, Cornwell JC, Kemp WM, Greene SM, Owens MS, Baker JE, Larsen RK, 2008. Nutrient budgets and management actions in the Patuxent River estuary, Maryland. Estuaries and Coasts 31(4), 623–651. [Google Scholar]
  8. Breiman L, 1984. Classification and regression trees, New York. [Google Scholar]
  9. Breiman L, 2001. Random Forests. Machine Learning 45, 5–32. [Google Scholar]
  10. Burns DA, Bhatt G, Linker LC, Bash JO, Capel PD, Shenk GW, 2021. Atmospheric nitrogen deposition in the Chesapeake Bay watershed: A history of change. Atmospheric Environment 251, 118277. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Casquin A, Dupas R, Gu S, Couic E, Gruau G, Durand P, 2021. The influence of landscape spatial configuration on nitrogen and phosphorus exports in agricultural catchments. Landscape Ecology 36(12), 3383–3399. [Google Scholar]
  12. Chanat JG, Moyer DL, Blomquist JD, Hyer KE, Langland MJ, 2016. Application of a weighted regression model for reporting nutrient and sediment concentrations, fluxes, and trends in concentration and flux for the Chesapeake Bay Nontidal Water-Quality Monitoring Network, results through water year 2012, U.S. Geological Survey, Reston, VA, p. 76. [Google Scholar]
  13. Chanat JG, Yang G, 2018. Exploring drivers of regional water-quality change using differential spatially referenced regression—a pilot study in the Chesapeake Bay watershed. Water Resources Research 54(10), 8120–8145. [Google Scholar]
  14. Chang SY, Zhang Q, Byrnes DK, Basu NB, Van Meter KJ, 2021. Chesapeake legacies: The importance of legacy nitrogen to improving Chesapeake Bay water quality. Environmental Research Letters 16(8), 085002. [Google Scholar]
  15. Chen D, Hu M, Dahlgren RA, 2014. A dynamic watershed model for determining the effects of transient storage on nitrogen export to rivers. Water Resources Research 50(10), 7714–7730. [Google Scholar]
  16. Chesapeake Bay Program, 2017. Chesapeake assessment and scenario tool (CAST) version 2017d.
  17. Cloern JE, 2001. Our evolving conceptual model of the coastal eutrophication problem. Marine Ecology Progress Series 210, 223–253. [Google Scholar]
  18. Clune JW, Capel PD, Miller MP, Burns DA, Sekellick AJ, Claggett PR, Coupe RH, Fanelli RM, Garcia AM, Raffensperger JP, Terziotti S, Bhatt G, Blomquist JD, Hopkins KG, Keisman JL, Linker LC, Shenk GW, Smith RA, Soroka AM, Webber JS, Wolock DM, Zhang Q, 2021. Nitrogen in the Chesapeake Bay watershed—A century of change, 1950–2050, U.S. Geological Survey,. [Google Scholar]
  19. Clune JW, Crawford JK, Chappell WT, Boyer EW, 2020. Differential effects of land use on nutrient concentrations in streams of Pennsylvania. Environmental Research Communications 2(11), 115003. [Google Scholar]
  20. Djodjic F, Bieroza M, Bergstrom L, 2021. Land use, geology and soil properties control nutrient concentrations in headwater streams. Sci Total Environ 772, 145108. [DOI] [PubMed] [Google Scholar]
  21. Dupas R, Ehrhardt S, Musolff A, Fovet O, Durand P, 2020. Long-term nitrogen retention and transit time distribution in agricultural catchments in western France. Environmental Research Letters 15(11), 115011. [Google Scholar]
  22. Dupas R, Jomaa S, Musolff A, Borchardt D, Rode M, 2016. Disentangling the influence of hydroclimatic patterns and agricultural management on river nitrate dynamics from sub-hourly to decadal time scales. Sci Total Environ 571, 791–800. [DOI] [PubMed] [Google Scholar]
  23. Eshleman KN, Sabo RD, 2016. Declining nitrate-N yields in the Upper Potomac River Basin: What is really driving progress under the Chesapeake Bay restoration? Atmospheric Environment 146, 280–289. [Google Scholar]
  24. Eshleman KN, Sabo RD, Kline KM, 2013. Surface water quality is improving due to declining atmospheric N deposition. Environmental Science and Technology 47(21), 12193–12200. [DOI] [PubMed] [Google Scholar]
  25. Falcone JA, 2015. U.S. conterminous wall-to-wall anthropogenic land use trends (NWALT), 1974–2012, U.S. Geological Survey, Reston, VA. [Google Scholar]
  26. Fisher TR, Fox RJ, Gustafson AB, Koontz E, Lepori-Bui M, Lewis J, 2021. Localized water quality improvement in the Choptank estuary, a tributary of Chesapeake Bay. Estuaries and Coasts 44(5), 1274–1293. [Google Scholar]
  27. Golden HE, Rajib A, Lane CR, Christensen JR, Wu Q, Mengistu S, 2019. Non-floodplain wetlands affect watershed nutrient dynamics: A critical review. Environmental Science & Technology 53(13), 7203–7214. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Groffman PM, Driscoll CT, Durán J, Campbell JL, Christenson LM, Fahey TJ, Fisk MC, Fuss C, Likens GE, Lovett G, Rustad L, Templer PH, 2018. Nitrogen oligotrophication in northern hardwood forests. Biogeochemistry 141(3), 523–539. [Google Scholar]
  29. Guyennon N, Salerno F, Rossi D, Rainaldi M, Calizza E, Romano E, 2021. Climate change and water abstraction impacts on the long-term variability of water levels in Lake Bracciano (Central Italy): A random forest approach. Journal of Hydrology: Regional Studies 37, 100880. [Google Scholar]
  30. Hastie T, Tibshirani R, Friedman J, 2009. The elements of statistical learning: Data mining, inference, and prediction, Springer-Verlag New York. [Google Scholar]
  31. Hirsch RM, Moyer DL, Archfield SA, 2010. Weighted Regressions on Time, Discharge, and Season (WRTDS), with an application to Chesapeake Bay river inputs. Journal of the American Water Resources Association 46(5), 857–880. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Hyer KE, Phillips SW, Ator SW, Moyer DL, Webber JS, Felver R, Keisman JL, McDonnell LA, Murphy R, Trentacoste EM, Zhang Q, Dennison WC, Swanson S, Walsh B, Hawkey J, Taillie D, 2021. Nutrient trends and drivers in the Chesapeake Bay Watershed, U.S. Geological Survey, Baltimore, MD, p. 4. [Google Scholar]
  33. Jassby AD, Cloern JE, 2016. wq: Exploring water quality monitoring data. R package version 0.4.8. http://cran.r-project.org/package=wq. [Google Scholar]
  34. Jordan TE, Correll DL, Weller DE, 1997. Relating nutrient discharges from watersheds to land use and streamflow variability. Water Resources Research 33(11), 2579–2590. [Google Scholar]
  35. Kaufman L, Rousseeuw PJ, 1990. Finding groups in data: An introduction to cluster analysis, John Wiley & Sons, Inc. [Google Scholar]
  36. Keisman JLD, Devereux OH, LaMotte AE, Sekellick AJ, Blomquist JD, 2018. Manure and fertilizer inputs to land in the Chesapeake Bay watershed, 1950–2012, U.S. Geological Survey, Reston, VA, p. 37. [Google Scholar]
  37. Kendall MG, 1975. Rank correlation methods, Oxford University Press, London, UK. [Google Scholar]
  38. Kuhn M, Wing J, Weston S, Williams A, Keefer C, Engelhardt A, Cooper T, Mayer Z, Kenkel B, R Core Team, Benesty M, Lescarbeau R, Ziem A, Scrucca L, Tang Y, Candan C, Hunt T, 2020. caret: Classification and Regression Training. [Google Scholar]
  39. Lawrence GB, Scanga SE, Sabo RD, 2020. Recovery of soils from acidic deposition may exacerbate nitrogen export from forested watersheds. Journal of Geophysical Research: Biogeosciences 125(1), e2019JG005036. [Google Scholar]
  40. Lefcheck JS, Orth RJ, Dennison WC, Wilcox DJ, Murphy RR, Keisman J, Gurbisz C, Hannam M, Landry JB, Moore KA, Patrick CJ, Testa J, Weller DE, Batiuk RA, 2018. Long-term nutrient reductions lead to the unprecedented recovery of a temperate coastal region. Proceedings of the National Academy of Sciences of the United States of America 115(14), 3658–3662. [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Liaw A, Wiener M, 2002. Classification and regression by randomForest. R News 2(3), 18–22. [Google Scholar]
  42. Lin J, Compton JE, Hill RA, Herlihy AT, Sabo RD, Brooks JR, Weber M, Pickard B, Paulsen SG, Stoddard JL, 2021. Context is everything: Interacting inputs and landscape characteristics control stream nitrogen. Environmental Science & Technology 55(12), 7890–7899. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Linker LC, Batiuk RA, Shenk GW, Cerco CF, 2013a. Development of the Chesapeake Bay watershed total maximum daily load allocation. Journal of the American Water Resources Association 49(5), 986–1006. [Google Scholar]
  44. Linker LC, Dennis R, Shenk GW, Batiuk RA, Grimm J, Wang P, 2013. Computing atmospheric nutrient loads to the Chesapeake Bay watershed and tidal waters. Journal of the American Water Resources Association 49(5), 1025–1041. [Google Scholar]
  45. Lintern A, McPhillips L, Winfrey B, Duncan J, Grady C, 2020. Best management practices for diffuse nutrient pollution: Wicked problems across urban and agricultural watersheds. Environmental Science & Technology 54(15), 9159–9174. [DOI] [PubMed] [Google Scholar]
  46. Lowrance R, Altier LS, Newbold JD, Schnabel RR, Groffman PM, Denver JM, Correll DL, Gilliam JW, Robinson JL, Brinsfield RB, Staver KW, Lucas W, Todd AH, 1997. Water quality functions of riparian forest buffers in Chesapeake Bay watersheds. Environmental Management 21(5), 687–712. [DOI] [PubMed] [Google Scholar]
  47. Malone TC, Newton A, 2020. The globalization of cultural eutrophication in the coastal ocean: Causes and consequences. Frontiers in Marine Science 7, 670. [Google Scholar]
  48. Marinos RE, Campbell JL, Driscoll CT, Likens GE, McDowell WH, Rosi EJ, Rustad LE, Bernhardt ES, 2018. Give and take: A watershed acid rain mitigation experiment increases baseflow nitrogen retention but increases stormflow nitrogen export. Environmental Science & Technology 52(22), 13155–13165. [DOI] [PubMed] [Google Scholar]
  49. McLauchlan KK, Gerhart LM, Battles JJ, Craine JM, Elmore AJ, Higuera PE, Mack MC, McNeil BE, Nelson DM, Pederson N, Perakis SS, 2017. Centennial-scale reductions in nitrogen availability in temperate forests of the United States. Scientific Reports 7(1), 7856. [DOI] [PMC free article] [PubMed] [Google Scholar]
  50. Moyer DL, Hirsch RM, Hyer KE, 2012. Comparison of two regression-based approaches for determining nutrient and sediment fluxes and trends in the Chesapeake Bay watershed, U.S. Geological Survey, Reston, VA, p. 118. [Google Scholar]
  51. Moyer DL, Langland MJ, 2020. Nitrogen, phosphorus, and suspended-sediment loads and trends measured at the Chesapeake Bay Nontidal Network stations: Water years 1985–2018. [Google Scholar]
  52. Murphy RR, Keisman J, Harcum J, Karrh RR, Lane M, Perry ES, Zhang Q, 2022. Nutrient improvements in Chesapeake Bay: Direct effect of load reductions and implications for coastal management. Environ Sci Technol 56(1), 260–270. [DOI] [PubMed] [Google Scholar]
  53. Murphy J, Sprague L, 2019. Water-quality trends in US rivers: Exploring effects from streamflow trends and changes in watershed management. Sci Total Environ 656, 645–658. [DOI] [PubMed] [Google Scholar]
  54. Murphy RR, Kemp WM, Ball WP, 2011. Long-term trends in Chesapeake Bay seasonal hypoxia, stratification, and nutrient loading. Estuaries and Coasts 34(6), 1293–1309. [Google Scholar]
  55. Newcomer ME, Bouskill NJ, Wainwright H, Maavara T, Arora B, Siirila-Woodburn ER, Dwivedi D, Williams KH, Steefel C, Hubbard SS, 2021. Hysteresis patterns of watershed nitrogen retention and loss over the past 50 years in United States hydrological basins. Global Biogeochemical Cycles 35(4), e2020GB006777. [Google Scholar]
  56. Oelsner GP, Stets EG, 2019. Recent trends in nutrient and sediment loading to coastal areas of the conterminous U.S.: Insights and global context. Sci Total Environ 654, 1225–1240. [DOI] [PubMed] [Google Scholar]
  57. Pennino MJ, Leibowitz SG, Compton JE, Hill RA, Sabo RD, 2020. Patterns and predictions of drinking water nitrate violations across the conterminous United States. Sci Total Environ 722, 137661. [DOI] [PMC free article] [PubMed] [Google Scholar]
  58. Prieto C, Le Vine N, Kavetski D, García E, Medina R, 2019. Flow prediction in ungauged catchments using probabilistic random forests regionalization and new statistical adequacy tests. Water Resources Research 55(5), 4364–4392. [Google Scholar]
  59. R Core Team, 2019. R: A language and environment for statistical computing, R Foundation for Statistical Computing, Vienna, Austria. http://www.R-project.org/. [Google Scholar]
  60. Rosi-Marshall EJ, Bernhardt ES, Buso DC, Driscoll CT, Likens GE, 2016. Acid rain mitigation experiment shifts a forested watershed from a net sink to a net source of nitrogen. Proceedings of the National Academy of Sciences of the United States of America 113(27), 7580–7583. [DOI] [PMC free article] [PubMed] [Google Scholar]
  61. Sabo RD, 2018. Shifting inputs and transformations of nitrogen in forested and mixed land use basins: implications for hydrologic nitrogen loss. Doctoral dissertation, University of Maryland, College Park. 10.13016/nc2a-kw9c. [DOI] [Google Scholar]
  62. Sabo RD, Clark CM, Bash J, Sobota D, Cooter E, Dobrowolski JP, Houlton BZ, Rea A, Schwede D, Morford SL, Compton JE, 2019. Decadal shift in nitrogen inputs and fluxes across the contiguous United States: 2002–2012. Journal of Geophysical Research: Biogeosciences 124(10), 3104–3124. [Google Scholar]
  63. Sabo RD, Clark CM, Compton JE, 2021. Considerations when using nutrient inventories to prioritize water quality improvement efforts across the US. Environmental Research Communications 3(4), 045005. [DOI] [PMC free article] [PubMed] [Google Scholar]
  64. Sabo RD, Elmore AJ, Nelson DM, Clark CM, Fisher T, Eshleman KN, 2020. Positive correlation between wood δ 15N and stream nitrate concentrations in two temperate deciduous forests. Environmental Research Communications 2(2), 025003. [DOI] [PMC free article] [PubMed] [Google Scholar]
  65. Sabo RD, Sullivan B, Wu C, Trentacoste E, Zhang Q, Shenk GW, Linker LC, 2022. Major sources of point and nonpoint source pollution to surface water have declined throughout the Chesapeake Bay watershed. Environmental Research Communications. [DOI] [PMC free article] [PubMed] [Google Scholar]
  66. Sanchez GM, Nejadhashemi AP, Zhang Z, Marquart-Pyatt S, Habron G, Shortridge A, 2015. Linking watershed-scale stream health and socioeconomic indicators with spatial clustering and structural equation modeling. Environmental Modelling & Software 70, 113–127. [Google Scholar]
  67. Sanford WE, Pope JP, 2013. Quantifying groundwater’s role in delaying improvements to Chesapeake Bay water quality. Environmental Science and Technology 47(23), 13330–13338. [DOI] [PubMed] [Google Scholar]
  68. Scanlon TM, Riscassi AL, Galloway JN, 2021. Observed changes in chronic and episodic acidification in Virginia mountain streams in response to the Clean Air Act and amendments. Atmospheric Environment 252, 118279. [Google Scholar]
  69. Sekellick AJ, Devereux OH, Keisman JLD, Sweeney JS, Blomquist JD, 2019. Spatial and temporal patterns of best management practice implementation in the Chesapeake Bay watershed, 1985–2014, U.S. Geological Survey, Reston, VA, p. 25. [Google Scholar]
  70. Sen PK, 1968. Estimates of the regression coefficient based on Kendall’s Tau. Journal of the American Statistical Association 63(324), 1379–1389. [Google Scholar]
  71. Shenk GW, Linker LC, 2013. Development and application of the 2010 Chesapeake Bay watershed total maximum daily load model. Journal of the American Water Resources Association 49(5), 1042–1056. [Google Scholar]
  72. Spearman C, 1904. The proof and measurement of association between two things. American Journal of Psychology 15, 72–101. [PubMed] [Google Scholar]
  73. Stets EG, Sprague LA, Oelsner GP, Johnson HM, Murphy JC, Ryberg K, Vecchia AV, Zuellig RE, Falcone JA, Riskin ML, 2020. Landscape drivers of dynamic change in water quality of U.S. rivers. Environ Sci Technol 54(7), 4336–4343. [DOI] [PubMed] [Google Scholar]
  74. Sun X, Lall U, Merz B, Dung NV, 2015. Hierarchical Bayesian clustering for nonstationary flood frequency analysis: Application to trends of annual maximum flow in Germany. Water Resources Research 51(8), 6586–6601. [Google Scholar]
  75. Tango PJ, Batiuk RA, 2016. Chesapeake Bay recovery and factors affecting trends: Long-term monitoring, indicators, and insights. Regional Studies in Marine Science 4, 12–20. [Google Scholar]
  76. Terziotti S, Capel PD, Tesoriero AJ, Hopple JA, Kronholm SC, 2017. Estimates of nitrate loads and yields from groundwater to streams in the Chesapeake Bay watershed based on land use and geology, U.S. Geological Survey, Reston, VA, p. 20. [Google Scholar]
  77. Tesoriero AJ, Gronberg JA, Juckem PF, Miller MP, Austin BP, 2017. Predicting redox-sensitive contaminant concentrations in groundwater using random forest classification. Water Resources Research 53(8), 7316–7331. [Google Scholar]
  78. U.S. Environmental Protection Agency, 2010. Chesapeake Bay Total Maximum Daily Load for nitrogen, phosphorus and sediment, U.S. Environmental Protection Agency, Annapolis, MD. [Google Scholar]
  79. van Breemen N, Boyer EW, Goodale CL, Jaworski NA, Paustian K, Seitzinger SP, Lajtha K, Mayer B, van Dam D, Howarth RW, Nadelhoffer KJ, Eve M, Billen G, 2002. Where did all the nitrogen go? Fate of nitrogen inputs to large watersheds in the northeastern USA. Biogeochemistry 57(1), 267–293. [Google Scholar]
  80. Van Meter KJ, Van Cappellen P, Basu NB, 2018. Legacy nitrogen may prevent achievement of water quality goals in the Gulf of Mexico. Science 360(6387), 427–430. [DOI] [PubMed] [Google Scholar]
  81. Vidon P, Karwan DL, Andres AS, Inamdar S, Kaushal S, Morrison J, Mullaney J, Ross DS, Schroth AW, Shanley JB, Yoon B, 2018. In the path of the Hurricane: impact of Hurricane Irene and Tropical Storm Lee on watershed hydrology and biogeochemistry from North Carolina to Maine, USA. Biogeochemistry 141(3), 351–364. [Google Scholar]
  82. Ward JH, 1963. Hierarchical grouping to optimize an objective function. Journal of the American Statistical Association 58(301), 236–244. [Google Scholar]
  83. Wieczorek ME, Jackson SE, Schwarz GE, 2018. Select attributes for NHDPlus version 2.1 reach catchments and modified network routed upstream watersheds for the conterminous United States (ver. 3.0, January 2021). U.S. Geological Survey data release, U.S. Geological Survey. [Google Scholar]
  84. Wolfe JD, Shook KR, Spence C, Whitfield CJ, 2019. A watershed classification approach that looks beyond hydrology: Application to a semi-arid, agricultural region in Canada. Hydrology and Earth System Sciences 23(9), 3945–3967. [Google Scholar]
  85. Zhang Q, Ball WP, Moyer DL, 2016. Decadal-scale export of nitrogen, phosphorus, and sediment from the Susquehanna River basin, USA: Analysis and synthesis of temporal and spatial patterns. Science of the Total Environment 563–564, 1016–1029. [DOI] [PubMed] [Google Scholar]
  86. Zhang Q, Blomquist JD, 2018. Watershed export of fine sediment, organic carbon, and chlorophyll-a to Chesapeake Bay: Spatial and temporal patterns in 1984–2016. Science of the Total Environment 619–620, 1066–1078. [DOI] [PubMed] [Google Scholar]
  87. Zhang Q, Brady DC, Boynton WR, Ball WP, 2015. Long-term trends of nutrients and sediment from the nontidal Chesapeake watershed: An assessment of progress by river and season. Journal of the American Water Resources Association 51(6), 1534–1555. [Google Scholar]
  88. Zhang Q, Fisher TR, Trentacoste EM, Buchanan C, Gustafson AB, Karrh R, Murphy RR, Keisman J, Wu C, Tian R, Testa JM, Tango PJ, 2021a. Nutrient limitation of phytoplankton in Chesapeake Bay: Development of an empirical approach for water-quality management. Water Research 188, 116407. [DOI] [PubMed] [Google Scholar]
  89. Zhang Q, Murphy RR, Tian R, Forsyth MK, Trentacoste EM, Keisman J, Tango PJ, 2018. Chesapeake Bay’s water quality condition has been recovering: Insights from a multimetric indicator assessment of thirty years of tidal monitoring data. Science of the Total Environment 637–638, 1617–1625. [DOI] [PMC free article] [PubMed] [Google Scholar]
  90. Zhang Q, Webber JS, Moyer DL, Chanat JG, 2021b. An approach for decomposing river water-quality trends into different flow classes. Science of the Total Environment 755(Pt 2), 143562. [DOI] [PubMed] [Google Scholar]

Associated Data

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

Supplementary Materials

s2
NIHMS1808992-supplement-s2.xlsx (14.6KB, xlsx)
s3
NIHMS1808992-supplement-s3.xlsx (33.3KB, xlsx)
s4
NIHMS1808992-supplement-s4.xlsx (291.2KB, xlsx)
s1
NIHMS1808992-supplement-s1.pdf (1.4MB, pdf)

Data Availability Statement

All data used in this research are described in the Methods section.

  • The annual flow-normalized TN loads at the NTN stations are available from the U.S. Geological Survey (Moyer and Langland, 2020).

  • The land use and N input data for the NTN stations are available from the Chesapeake Assessment Scenario Tool (Chesapeake Bay Program, 2017).

  • The geology data for the NTN stations are available from the U.S. conterminous wall-to-wall anthropogenic land use trends dataset (Falcone, 2015).

  • The physiography data for the NTN stations are available from the U.S. Geological Survey (Wieczorek et al., 2018).

RESOURCES