Subscriber access provided by UNIVERSITY OF LEEDS
Article
Optimizing Sampling Strategies for Riverine Nitrate using High-Frequency Data in Agricultural Watersheds Kaycee N Reynolds, Terrance D. Loecke, Amy J. Burgin, Caroline A. Davis, Diego Riveros-Iregui, Steven A Thomas, Martin A. St.Clair, and Adam S. Ward Environ. Sci. Technol., Just Accepted Manuscript • DOI: 10.1021/acs.est.5b05423 • Publication Date (Web): 18 May 2016 Downloaded from http://pubs.acs.org on May 18, 2016
Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a free service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are accessible to all readers and citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.
Environmental Science & Technology is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.
Page 1 of 29
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
Environmental Science & Technology
Optimizing Sampling Strategies for Riverine Nitrate using High-Frequency Data in Agricultural Watersheds
Kaycee N. Reynolds1, Terrance D. Loecke*1,2, Amy J. Burgin1,2, Caroline A. Davis3, Diego Riveros-Iregui4, Steven A. Thomas1, Martin A. St. Clair5, and Adam S. Ward6 1
School of Natural Resources, University of Nebraska-Lincoln, Lincoln, NE (
[email protected],
[email protected]) 2
Present address: Kansas Biological Survey and University of Kansas, Lawrence, KS (
[email protected],
[email protected]) 3
IIHR Hydroscience & Engineering, University of Iowa, Iowa City, IA (
[email protected]) 4
Department of Geography, University of North Carolina, Chapel Hill, NC (
[email protected])
5
Department of Chemistry, Coe College, Cedar Rapids, IA (
[email protected])
6
School of Public and Environmental Affairs, Indiana University, Bloomington, IN (
[email protected]) Running Title: Sampling Frequency *Corresponding Author: Terrance D. Loecke Kansas Biological Survey University of Kansas 2101 Constant Ave. Lawrence KS, 66047 Date of Preparation: 21 March 2016, revisions post review for ES&T Revision/Version: 18 Word count: 4,600 (excludes references) # Figures: 4 # Tables: 2 # References: 47 In Review: Environmental Science & Technology Type of Contribution: Research Article Key Findings: 1) Fixed time-interval sampling has the greatest return on sampling investment by reducing uncertainty faster in estimates of mean nitrate concentration, proportion of yearly samples exceeding the nitrate drinking water standard, peak nitrate concentration, and nitrate flux compared to event-based strategies. 2) Storm sampling using a historical discharge threshold optimally characterized nitrate flux only at coarse frequencies. 3) Traditional sampling frequencies (i.e., bi-weekly and monthly) sufficiently characterize the four nitrate parameters considered in this study.
1
ACS Paragon Plus Environment
Environmental Science & Technology
51
Abstract: Understanding linked hydrologic and biogeochemical processes such as nitrate loading
52
to agricultural streams requires that the sampling bias and precision of monitoring strategies be
53
known. An existing spatially distributed, high-frequency nitrate monitoring network covering
54
~40% of Iowa (USA) provided direct observations of in situ nitrate concentrations at a temporal
55
resolution of 15 minutes. Systematic subsampling of nitrate records allowed for quantification of
56
uncertainties (bias and precision) associated with estimates of various nitrate parameters,
57
including: mean nitrate concentration, proportion of samples exceeding the nitrate drinking water
58
standard (DWS), peak (>90th quantile) nitrate concentration, and nitrate flux. We subsampled
59
continuous records for 47 site-year combinations mimicking common, but labor intensive, water
60
sampling regimes (e.g. time-interval, stage-triggered and dynamic-discharge storm sampling).
61
Our results suggest that time-interval sampling most efficiently characterized all nitrate
62
parameters, except at coarse frequencies for nitrate flux. Stage-triggered storm sampling most
63
precisely captured nitrate flux when using less than 0.19% of possible 15-minute observations for
64
a site-year. The time-interval strategy had the greatest return on sampling investment by most
65
precisely and accurately quantifying nitrate parameters per sampling effort. These uncertainty
66
estimates can aid in designing sampling strategies focused on nitrate monitoring in the tile-
67
drained Midwest (USA), or similar agricultural regions.
68 69
Keywords: High-frequency monitoring, Nitrate, water quality, error estimation, agriculture
2
ACS Paragon Plus Environment
Page 2 of 29
Page 3 of 29
70 71
Environmental Science & Technology
Introduction Agricultural practices accelerate nitrate leaching in cultivated landscapes,1 particularly in
72
the Midwestern US where intensive agriculture and artificially-drained fields predominate.2–5
73
Stream nitrate concentration variability can occur at timescales spanning minutes (episodic
74
response), weeks (seasonal cycles) and decades (long-term trends) in response to dynamic
75
hydrological processes.6,7 Variation in nitrate concentration associated with specific hydrologic
76
events (e.g., storm events, climatic variability) can be a major source of sampling error when
77
estimating nitrate flux.8 Inter-annual variation in nitrate flux from agricultural basins is primarily
78
controlled by discharge (Q)9, with most nutrient losses occurring when discharge ≥ median Q,
79
and more than 50% of export occurring during episodic precipitation events accompanied by
80
extreme flow (>90th percentile Q).10,11 Generally, maximum storm-related nitrogen fluxes are one
81
to two orders of magnitude higher than pre-storm fluxes.8 Thus, accurately capturing nitrate flux
82
during extreme flow conditions is critical for reliable flux estimates.
83
Despite the fact that sampling uncertainty is an inverse function of sampling frequency,
84
logistical and budget constraints typically limit sampling efforts, which contributes to amplified
85
errors in water quality parameter estimates.8,12,13 Water quality monitoring is often initiated to: 1)
86
collect data for model development or validation, 2) verify effectiveness of best management
87
practices, 3) measure water quality status for regulatory compliance, and 4) document existing
88
background conditions to evaluate trends.12,14 Historically, water quality monitoring has relied on
89
grab sampling at discrete time-intervals (e.g., bi-weekly or monthly) throughout the year, from
90
which annual metrics are estimated.15 Although infrequent concentration measurements combined
91
with mean daily stream Q may be acceptable for evaluating nitrate load on a seasonal or annual
92
scale, such infrequent measurements may not sufficiently capture individual or sequenced event-
93
related nitrate fluxes.6 More recently, monitoring agencies have invested in in situ sensors that
94
measure nitrate concentrations at high-frequency intervals (e.g., seconds to minutes). These
95
sensors provide nearly 3000 times more nitrate concentration data than monthly grab sampling. In
3
ACS Paragon Plus Environment
Environmental Science & Technology
96
Iowa, continuous nitrate sensing is used to manage drinking-water supplies and wastewater
97
discharges as well as to track the effectiveness of agricultural best management practices.16
98 99
Traditional labor-intensive sampling methods for estimating nutrient loads have cumulative uncertainties introduced by four procedural categories: stream flow measurement,
100
sample collection, sample preservation/storage, and laboratory analysis.17 Of these, the strategy
101
for collecting samples has the potential to introduce the most uncertainty, accounting for 4-48%
102
of the error under typical collection scenarios.17 With limited monitoring resources, infrequent
103
and cost-efficient sampling regimes will likely be the norm for the foreseeable future,18 thus
104
making an assessment of the sampling uncertainty of simplified strategies critical.
105
The use of spatially distributed, continuous real-time data presents an unparalleled
106
opportunity to quantify uncertainties of conventional sampling strategies across a range of
107
watersheds. We use such a database to ask: How is nitrate concentration and flux estimation
108
uncertainty affected by sampling strategy? We address this question using a high-frequency
109
nitrate monitoring dataset (n≈830,000) collected in watersheds covering 40% of the state of Iowa,
110
the second largest agricultural producer in the United States.19 With more frequent sampling, we
111
predicted that uncertainties associated with nitrate concentration and flux estimates would
112
decrease. We expected the most frequent time-interval sampling strategy (i.e., daily) to provide
113
the most precise estimates of mean nitrate concentration, proportion of yearly samples exceeding
114
the nitrate DWS, peak nitrate concentration, and nitrate flux. Alternatively, we hypothesized that
115
while the storm-based sampling regimes would likely increase uncertainty in mean nitrate
116
concentration estimates, by targeting extreme Q, this regime is more likely to capture maximum
117
concentrations and may be superior for characterizing the instances of stream water exceeding the
118
nitrate DWS, peak nitrate concentrations and nitrate flux. Understanding how both timing and
119
frequency of monitoring affects accuracy of nitrate parameter estimates will provide valuable
120
insight for informed management and optimization of monitoring strategies in agricultural
121
watersheds.
4
ACS Paragon Plus Environment
Page 4 of 29
Page 5 of 29
Environmental Science & Technology
122
Materials & Methods
123
Catchment characteristics
124
Nitrate (NO3--N) concentration and discharge (Q) were measured using in situ continuous
125
monitoring sensors at 17 locations established by the USGS. Collectively, these 17 sites cover 11
126
HUC-8 watersheds and drain 58,831 km2 (~40%) of Iowa (Figure SI-S1). Land used in these
127
watersheds is predominantly agricultural, defined here as corn and soybean row crop (mean =
128
72%; Table SI-S1), with growing urbanization in Des Moines, Cedar Rapids, and Iowa City. The
129
watersheds used in our analyses were on average 45.5% corn, 26.2% soy and 15.8% other
130
production.20 Iowa has a distinct four-season climate, with warm and humid summers, cool
131
autumns, cold winters, and wet springs, reflective of the state’s mid-latitude and interior
132
continental position.21 The annual average temperature ranges from 7.2-11.1°C and annual
133
average precipitation ranges from 66-96.5 cm (from the NW to SE corners), with more than 75%
134
of the annual precipitation falling during the growing season (i.e. April-September).22
135
Due to poor soil drainage that causes fields to otherwise be submerged following storm
136
events, agricultural fields are often outfitted with underground tiles that drain water quickly from
137
the soil-rooting zone.23 The extent of artificial drainage is difficult to estimate, but conservatively
138
covers 32% of Iowa cropland.24 The presence of drainage tiles rapidly connects the land surface,
139
vadose zone, and shallow groundwater to the stream network. Tile drainage allows for the rapid
140
transport of nitrate, originating from agricultural practices, into nearby surface waters; a system
141
that results in high nitrate concentrations in Iowa streams.25,26 Nitrate yields from Iowa rivers are
142
some of the highest in the nation,2 often ranking the state near the highest in terms of estimated
143
total flux of nitrogen contributed to the Gulf of Mexico Hypoxic “Dead Zone”.27
144
High-frequency monitoring sensors
145
As of 2013, the United States Geological Survey (USGS) and partner organizations
146
monitored nitrate continuously at 44 locations in the continental USA;28 17 of these locations
147
were used in this study (Table SI-S1). The USGS measures Q and in situ stream nitrate
5
ACS Paragon Plus Environment
Environmental Science & Technology
148
concentration at a 15-minute interval. Stream nitrate is recorded as nitrate+nitrite as N (mg NO3--
149
N L-1) using a Hach Nitratax plus sc Sensor (2 mm path length; Loveland, CO, USA), which has
150
a measurement range of 0.1-50 mg NO3--N L-1 and a 0.1 mg NO3--N L-1 resolution (measuring
151
error: ±3% of the mean MW± 0.5). Gaps in the datasets reflect operational issues with the
152
equipment or periods when the sensor was retrieved for calibration and servicing. Most of the in
153
situ sensors operate from early spring (mid-March to mid-April) through early winter (mid-
154
November to mid-December) when data collection is suspended because the sensors cannot
155
function below 2°C.29 Stream Q was estimated using site-specific rating curves based on
156
continuously recorded gage height and discretely measured streamflow.30 Long-term Q data were
157
obtained from the USGS National Water Information System.31
158
Nitrate Parameters
159
We analyzed results for four nitrate parameters: mean nitrate concentration (CM),
160
proportion of observations exceeding the 10 mg NO3--N L-1 EPA Maximum Contaminant Level
161
(MCL) for nitrate, (MCLP),32 peak (>90th quantile) nitrate concentrations (CP), and nitrate flux
162
(FT). Each parameter was evaluated over the range of dates included in each site-year. Mean
163
nitrate concentration (CM) was calculated for each of the 47 site-years as =
Σ()
164
where C(t) represents the 15-minute nitrate concentration readings and n is the number of
165
concentration observations for a site-year based on calendar years. The MCLP, often referred to as
166
the nitrate DWS, was determined for each site year using the empirical cumulative distribution
167
function in R.33 The 90th percentile of the nitrate concentration for each site year (CP) was
168
determined using R. Instantaneous nitrate mass flux is the product of Q and concentration at each
169
time step. Nitrate mass flux (FT), often referred to as load, over a period of interest from 0 to T,
170
can be calculated by integration as
6
ACS Paragon Plus Environment
Page 6 of 29
Page 7 of 29
Environmental Science & Technology
= () ()d
171
where, 0 and T are the earliest and latest sample date/times, respectively, for a site-year.
172
Estimation of nitrate flux is typically based on available data, which often consists of frequent Q
173
(e.g. often mean daily Q) and infrequent concentration.34 The advent of high-temporal resolution
174
sensors for measuring in-stream nitrate concentrations enables us to calculate these quantities at
175
much higher temporal resolution than heretofore possible and reduce error introduced by
176
averaging.
177
Sampling regimes
178
Three separate sampling regimes were tested here: time-interval, stage-triggered storm
179
sampling, and dynamic-Q storm sampling (Figure 1). Time-interval sampling mimics manual
180
grab sampling at four time-intervals: daily (Figure 1A), weekly, bi-weekly and monthly.
181
Sampling was limited to weekdays (i.e. Monday-Friday) and typical working hours (09:00-
182
17:00). Storm sampling mimics the use of a programmed auto-sampler triggered based on Q, [e.g.
183
Teledyne ISCO (Lincoln, NE, USA)]. Both stage-triggered and dynamic-Q storm sampling
184
randomly drew from a storm-hydrograph at four frequencies: once during an event, twice during
185
an event, daily during an event (Figure 1B-C) and hourly during an event. In addition, two
186
stratified sampling frequencies: monthly + hourly during an event, and weekly + hourly during an
187
event were considered, combining time-interval and storm sampling strategies. All sampling
188
strategies were applied to each of the 47 site-year combinations. For each strategy we drew
189
random samples from the continuous dataset, which included the timestamp and associated nitrate
190
concentration, Q, and instantaneous flux. Sampling effort, defined as the proportion of samples
191
used for each sampling frequency, was calculated as the number of sample draws from the total
192
number of 15-minute samples available for a site-year. Using systematic and Monte Carlo
193
sampling, we generated sample subsets from the continuously collected nitrate data using
194
predetermined, common sampling strategies. As the number of sample iterations increases, the
7
ACS Paragon Plus Environment
Environmental Science & Technology
195
variability will decrease until it reaches a point where we can assume it is representative of the
196
population.35 Based on previous simulations in the literature,35,36 500-interations was deemed
197
sufficient for this study. Subsampling was conducted with replacement.
198
Unique storm events for each site-year dataset were labeled based on stage-triggered and
199
dynamic-Q thresholds. The stage-triggered approach set the storm threshold as the 90th percentile
200
of the historical (all possible years 1907-2013) Q data for each site (Figure 1B). Extreme
201
discharges (>90th percentile) contribute >50% of the nitrate export in the Midwest.11 The start of a
202
unique event was signaled when this Q threshold was surpassed for two consecutive readings
203
(i.e., ~30 min) and ended when Q fell below this threshold for two consecutive readings (Figure
204
1B). For the dynamic-Q threshold approach, we calculated the distribution of percent change in Q
205
over two hour intervals for each site-year. The 99th percentile of this distribution was used as the
206
dynamic-Q threshold for that site-year. The 99th percentile was chosen to capture a similar
207
number of events as the stage-triggered threshold. This more flexible threshold allowed us to
208
sample both small and large storm events within each site-year. A unique event began when the
209
percent change in Q exceeded this threshold (Figure 1C) and ended when Q returned to its start
210
value for that event or, in instances where Q does not return to its start value, one time point
211
before the next event was initiated (Figure 1C). Only events longer than 48 sampling periods (i.e.,
212
½ a day) were included; events were also limited to a 5-day duration maximum. Once unique
213
events were labeled for each site-year based on the two separate thresholds, storm sampling at
214
each of the aforementioned frequencies was carried out to simulate sampling either by manual
215
storm-chasing (i.e., sampling once or twice during an event) or by an automatic Q-induced
216
instrument sampling daily or hourly during an event.
217
Using the subsampled time series, nitrate concentration was linearly interpolated between
218
the sampled points (i.e., daily, weekly, monthly) for each 15-minute time step from the first draw
219
(minimum Date/Time) for that site-year to the last draw (maximum Date/Time). This was paired
220
with the corresponding 15-minute Q value collected at the same date/time for that same site-year.
8
ACS Paragon Plus Environment
Page 8 of 29
Page 9 of 29
Environmental Science & Technology
221
By integrating the instantaneous flux over the time period from initial to final sample for that site-
222
year, we were able to calculate nitrate flux as:
= () ()d
223
where C0 is instantaneous nitrate concentration, Q0 is the paired instantaneous discharge, and 0
224
and T are the earliest and latest date/time for that sample iteration. These calculated subsample
225
nitrate fluxes were compared to the nitrate flux for that site-year, calculated from the continuous
226
nitrate concentration and Q data.
227
Quantifying Sampling Uncertainty
228 229
For each sampling strategy-frequency combination (i.e., sampling regime) and nitrate parameter, the sampling bias was quantified relative to the parameter mean as follows: ! =
"# × 100 $̅
230
where ME is the mean error (i.e., deviation) between the calculated nitrate parameter from each
231
subsample (n=500) and that calculated from the 15-minute complete sensor data set, $̅ is the
232
calculated nitrate parameter from the complete data set of that particular site-year combination.
233
Relative bias is expressed as a percentage.
234 235
For each sampling strategy-frequency combination (i.e., sampling regime) and nitrate parameter, the sampling precision was quantified as the coefficient of variation (CV) as follows: ) =
"*# × 100 $̅
236
where RMSE is the root mean squared error (i.e., deviation) between the calculated water
237
parameter from each subsample (n=500) and that calculated from the 15-minute complete sensor
238
data set, $̅ is the calculated nitrate parameter from the complete data set of that particular site-
239
year combination, and CV is expressed as a percentage.
240 241
Statistical Analysis and Model Development
9
ACS Paragon Plus Environment
Environmental Science & Technology
242
Kruskal-Wallis multiple comparison tests were used to determine time-interval sampling
243
effects on sampling precision for each of the four nitrate parameters.37 A linear mixed effects
244
analysis of the relationship between natural log of CV and sampling effort was conducted using
245
the lme4 package in R.38 We included sampling effort (i.e. proportion of possible samples) and
246
sampling type (with an interaction term) as fixed effects. The random effects in the model were
247
an intercept for site-year, as well as by-site-year random slopes to control for inherent differences
248
among sites and years. The response, CV(%), was natural log-transformed to reduce
249
heteroskedasticity and to meet the assumption that residuals in the model have similar deviations
250
from predicted values.
251 252 253
Results Near continuous monitoring of 17 Iowa streams yielded 829,968 instantaneous in situ
254
nitrate concentration measurements and 1,158,100 Q estimates. Mean nitrate concentrations
255
ranged from 1.5 (±0.4) to 32.5 (±6.7) mg NO3--N L-1 (±1 standard deviation) and mean Q ranged
256
from 0.5 (±0.9) to 376.9 (±361) m3 s-1 (Table SI-S2). Across the 47 site-years, the highest
257
monthly mean nitrate concentrations were observed in May (11.8±6.5 mg NO3--N L-1) and June
258
(10.7±5.7 mg NO3--N L-1), while the highest mean Q was observed in June (156±296 m3 s-1) and
259
March (140±275 m3 s-1) (Table SI-S3). The mean nitrate concentration across all sites was highest
260
in 2008 (9.3 mg NO3--N L-1) and had the largest range in 2013 (±7.9 mg NO3--N L-1), while Q
261
was highest on average and most variable in 2010 (195±282 m3 s-1; Table SI-S4 and SI-S5). The
262
nitrate drinking water standard (DWS; 10 mg NO3--N L-1) was exceeded in 37 of the 47 site-
263
years. Six of the 10 site-years that did not exceed the DWS occurred during 2012, the driest year.
264
Across all sites, instantaneous nitrate concentration was most variable from May-July, while
265
instantaneous Q was most variable during March, May and June (Table SI-S3). Peaks in
266
discharge occur most frequently May and June, which corresponds to the highest average rainfall
267
of all months (12.1 and 13.0 cm respectively).39
10
ACS Paragon Plus Environment
Page 10 of 29
Page 11 of 29
268
Environmental Science & Technology
Sampling bias and precision vary with monitoring strategy and parameter of interest
269
(Figures 2 and 3). Time-interval sampling decreased bias and increased precision for
270
characterizing nitrate concentration more than either stage triggered or dynamic-Q threshold
271
strategies (Figure 2A-B and Figure 3A-B). This lack of systematic bias indicates sampling during
272
“normal work hours” sufficiently characterizes diurnal nitrate concentration patterns across these
273
watersheds. Within the time-interval strategy, precision improves on average 11 to 14-fold
274
among nitrate concentration parameters as sampling frequency increases (i.e., from 9.7%-0.9%
275
for monthly to daily frequencies when estimating mean nitrate concentration; Figure 3A). The
276
precision of daily sampling averages 0.9% of the parameter mean for mean nitrate concentration,
277
0.9% for the proportion of samples exceeding the nitrate DWS, and 0.9% for the peak nitrate
278
concentration estimates (Table 1). For all frequencies in this strategy, even traditional monthly
279
sampling, median CV for nitrate concentration parameters were 1.3% of total possible
329
samples, in which case, it actually amplified uncertainty making dynamic-Q the more
330
advantageous storm sampling approach (Figure 4C). Additionally, when using > 0.2% of possible
331
samples the time-interval strategy was more precise for flux estimates than stage-trigged storm
332
sampling (first crossover point; Figure 4D).
333 334
Discussion
335
Optimizing monitoring strategies by nitrate parameter
336
Optimal sampling strategy is a function of the parameter of interest, the costs associated
337
with equipment and person-power, and the ultimate objective of the monitoring. The time-interval
338
strategy was the best monitoring strategy to improve precision without sacrificing accuracy (i.e.,
339
systematic bias) in mean nitrate concentration, proportion of samples exceeding the nitrate DWS,
340
and peak nitrate concentration (Figures 4A-C). However, optimization for nitrate flux was
341
dependent on the sampling effort (Figure 4D). When sampling < 0.2% of the total possible 15-
342
minute samples at a site, stage-triggered storm sampling was optimal; when sampling more
343
frequently, the time-interval strategy quickly became optimal. For our 47 site-years, this meant
344
stage-triggered storm sampling was optimal at coarse frequencies (1-2 samples per event), while
345
the time-interval strategy optimally characterized nitrate flux at weekly to daily frequencies
13
ACS Paragon Plus Environment
Environmental Science & Technology
346
(Figures 4D and SI-S2). Dynamic-Q storm sampling was the least optimal strategy for any of the
347
nitrate parameters. Thus, if characterizing nitrate concentration parameters is the primary interest,
348
as is often the case for understanding water quality entering a municipal water supply, a time-
349
interval based strategy is the best investment of monitoring resources. However, if the monitoring
350
goal is to understand fluxes of nitrate heading towards connected downstream waters, as is the
351
case for efforts to reduce nutrient loading to coastal areas including the Gulf of Mexico, stage-
352
triggered sampling is best given infrequent sampling, but can quickly be outpaced by time-
353
interval strategies.
354
Estimating sampling bias and precision allows for optimization of monitoring
355
methodologies, improvement in decision-making and regulatory development, and contributes to
356
model calibration, validation and ultimately application.40 Quality assurance protocols are
357
implemented by water monitoring agencies to reduce overall uncertainty; however, sampling bias
358
and precision estimations were previously under-quantified or overlooked due to lack of high-
359
frequency data with which to compare sample estimates. Many previous studies have noted the
360
substantial errors in water quality parameter estimates that result from infrequent sampling;
361
however, most lacked continuous data sets with which to compare parameters estimated from
362
infrequent sampling.34,41–44 For example, Alewell et al.41 aimed to understand if the gain of
363
information with ‘high resolution’ daily measurements outweighed the additional efforts and
364
associated costs compared to weekly or bi-weekly sampling. Their results indicated that bi-
365
weekly sampling accurately characterized ion fluxes in a semi-natural ecosystem and that any
366
additional information gained by daily sampling was not worth the additional effort. However,
367
they noted the dynamic nature of agricultural systems and conceded that these conclusions may
368
not hold in agricultural landscapes. Our results indicate sampling precision is significantly
369
improved when going from bi-weekly or weekly to daily sampling for all four nitrate parameters;
370
however, there is no significant difference in sampling precision or bias between bi-weekly and
371
weekly sampling. Therefore bi-weekly sampling is recommended for mean and peak nitrate
14
ACS Paragon Plus Environment
Page 14 of 29
Page 15 of 29
Environmental Science & Technology
372
concentrations in this system, as it sufficiently characterized these nitrate parameters, contributing
373
on average 3.6% and 4.9% margin of precision to estimates across all 47 site-years (Table 1).
374
There was no significant difference in sampling precision induced by bi-weekly vs. monthly
375
sampling for the proportion of samples exceeding the nitrate DWS or nitrate flux, so the coarser
376
frequency, monthly, could be a reasonable approach to reduce costs.
377
Combining data-driven monitoring optimization with non-data sampling limitations
378
Our results indicate that daily time-interval sampling resulted in the best sampling
379
precision for all nitrate concentration related parameter estimates, while the most frequent stage-
380
triggered storm sampling frequency resulted in the best sampling precision in nitrate flux
381
estimates. Although these strategies provided the most precise and accurate estimates, they may
382
not be economically feasible long-term or across a large spatial extent. Understanding the
383
marginal return on sampling effort allows us to obtain more useful information and adjust
384
sampling strategy or frequency to balance effort and accuracy. Based on our results, traditional
385
sampling frequencies (i.e., bi-weekly and monthly) would characterize mean nitrate
386
concentration, the proportion of samples exceeding the DWS, peak nitrate concentrations and
387
nitrate flux, without bias with a precision of 90th quantile), and (D) flux as a function of sample proportion. The proportion of possible samples used represents sampling intensity; as you go from left to right across the x-axis, sampling gets more frequent.
670
27
ACS Paragon Plus Environment
Environmental Science & Technology
Page 28 of 29
671
List of Tables
672 673 674
Table 1. Median precision for all sampling regimes across 47 site-years. Each row represents a sampling strategy-frequency combination. The columns show the median precision induced by a particular regime as the coefficient of variation with Q1 and Q3 in parentheses to indicate spread. Sampling Regime
Sampling Frequency
Time-Interval
Daily
Time-Interval
Weekly
Time-Interval
Bi-Weekly
Time-Interval
Monthly
Stage-Triggered Storm Stage-Triggered Storm Stage-Triggered Storm Stage-Triggered Storm Stage-Triggered Storm Stage-Triggered Storm Dynamic-Q Storm Dynamic-Q Storm Dynamic-Q Storm Dynamic-Q Storm Dynamic-Q Storm Dynamic-Q Storm
Weekly + Hourly/event Monthly + Hourly/event Hourly/event Daily/event 2x/event 1x/event Weekly + Hourly/event Monthly + Hourly/event Hourly/event Daily/event 2x/event 1x/event
Median CV for CM (Q1-Q3) 0.9 (0.4-1.9) 2.6 (1.8-5.9) 3.6 (2.7-6.5) 9.7 (7.4-14.2) 51.6 (19.1-90.5) 53.1 (19.3-96.1) 53.6 (20.4-96.8) 51.23 (20.3-96.4) 49.7 (18.2-84.2) 49.6 (18.7-84.7) 20.2 (8.2-39.6) 21.4 (9.4-41.5) 21.8 (9.6-42.5) 20.1 (8.5-35.6) 17.1 (8.4-33.2) 18.0 (9.4-33.8)
Median CV for MCLP (Q1-Q3) 0.9 (0.8-1.7) 3.9 (2.3-6.2) 5.7 (4.0-9.7) 13.3 (8.2-17.9) 58.6 (21.4-79.8) 66.7 (21.7-88.5) 69.4 (21.8-96.1) 60.7 (21.1-94.4) 65.7 (21.3-96.7) 66.8 (23.6-96.6) 16.1 (5.9-33.7) 16.7 (6.1-35.3) 17.1 (6.2-36.0) 13.0 (5.1-31.1) 12.7 (5.7-30.6) 14.0 (7.7-31.6)
675 676 677
28
ACS Paragon Plus Environment
Median CV for CP (Q1-Q3) 0.9 (0.5-1.9) 3.6 (2.2-6.0) 4.9 (3.0-6.7) 12.4 (7.7-15.9) 10.0 (5.5-18.8) 10.1 (5.6-19.0) 10.1 (5.6-19.1) 9.6 (4.8-17.2) 8.8 (4.7-16.8) 7.1 (5.4-16.5) 12.5 (4.5-4.7) 12.8 (4.9-26.5) 12.9 (5.0-27.7) 10.8 (5.5-26.5) 12.3 (6.7-25.7) 13.1 (7.4-22.9)
Median CV for FT (Q1-Q3) 1.1 (0.5-2.0) 3.2 (2.5-5.6) 5.4 (3.7-9.5) 9.9 (7.7-15.5) 0.7 (0.4-1.5) 1.7 (1.0-3.5) 1.5 (0.3-3.6) 2.2 (1.0-4.8) 8.1 (5.2-10.8) 10.3 (6.5-13.7) 1.8 (1.3-3.1) 3.9 (2.6-7.1) 5.8 (2.1-13.2) 9.2 (4.3-16.3) 11.1 (7.3-16.6) 12.3 (8.3-19.7)
Page 29 of 29
678 679 680 681 682 683
Environmental Science & Technology
Table 2. Linear mixed-model intercepts and slopes for each sampling strategy. Each row represents a sampling strategy: Time-Interval (TI), Stage-Triggered Storm Sampling (ST), and Dynamic-Q Storm Sampling (DQ. The columns show the intercepts and slopes with standard error for each of the mixed-models in Figure 3. Note: *Slope not significantly different from 0 (t value0.05), aNot significantly different from one another, +Not highly significant (p-value