Optimizing Sampling Strategies for Riverine Nitrate Using High

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


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


Page 2 of 29

Page 3 of 29

70 71


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



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


Page 4 of 29

Page 5 of 29


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



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


Page 6 of 29

Page 7 of 29


= () ()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



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


Page 8 of 29

Page 9 of 29


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



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


Page 10 of 29

Page 11 of 29

268


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



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


Page 14 of 29

Page 15 of 29


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



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


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


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

Optimizing Sampling Strategies for Riverine Nitrate Using High

Recommend Documents