Revealing Biotic and Abiotic Controls of Harmful Algal Blooms in a

Subscriber access provided by the Henry Madden Library | California State University, Fresno

Article

Revealing biotic and abiotic controls of harmful algal blooms in a shallow subtropical lake through statistical machine learning Natalie Genevieve Nelson, Rafael Muñoz-Carpena, Edward J. Phlips, David Kaplan, Peter Sucsy, and J. Hendrickson Environ. Sci. Technol., Just Accepted Manuscript • DOI: 10.1021/acs.est.7b05884 • Publication Date (Web): 24 Feb 2018 Downloaded from http://pubs.acs.org on February 25, 2018

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 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 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 35

Environmental Science & Technology

1

TITLE

2

Revealing biotic and abiotic controls of harmful algal blooms in a shallow subtropical

3

lake through statistical machine learning

4

AUTHORS

5

Natalie G. Nelson1,2, Rafael Muñoz-Carpena1*, Edward J. Phlips3, David Kaplan4, Peter

6

Sucsy5, John Hendrickson5

7

AUTHOR AFFILIATIONS

8

1

9

Florida, Gainesville, FL, USA

Hydrology & Water Quality, Agricultural & Biological Engineering, University of

10

2

11

Raleigh, NC, USA

12

3

13

of Florida, Gainesville, FL, USA

14

4

15

Engineering Sciences Department, University of Florida, Gainesville, FL, USA

16

5

17

* corresponding author: [email protected]

At present: Biological & Agricultural Engineering, North Carolina State University,

Fisheries & Aquatic Sciences, School of Forest Resources & Conservation, University

Engineering School of Sustainable Infrastructure and Environment, Environmental

St. Johns River Water Management District, Palatka, FL, USA

ACS Paragon Plus Environment

1


Page 2 of 35

18

ABSTRACT

19

Harmful algal blooms are a growing human and environmental health hazard globally.

20

Eco-physiological diversity of the cyanobacteria genera that make up these blooms

21

creates challenges for water managers tasked with controlling the intensity and frequency

22

of blooms, particularly of harmful taxa (e.g., toxin producers, N2 fixers). Compounding

23

these challenges is the ongoing debate over the efficacy of nutrient management

24

strategies (phosphorus-only versus nitrogen and phosphorus), which increases decision-

25

making uncertainty. To help improve our understanding of how different cyanobacteria

26

respond to nutrient levels and other biophysical factors, we analyzed a unique 17-year

27

dataset comprising monthly observations of cyanobacteria genera and zooplankton

28

abundances, water quality, and flow a bloom-impacted, subtropical, flow-through lake in

29

Florida (USA). Using the Random Forests machine learning algorithm, an assumption-

30

free ensemble modeling approach, we characterized and quantified relationships among

31

environmental conditions and five dominant cyanobacteria genera. Results highlighted

32

nonlinear relationships and critical thresholds between cyanobacteria genera and

33

environmental covariates, the potential for hydrology and temperature to limit the

34

efficacy of cyanobacteria bloom management actions, and the importance of a dual

35

nutrient management strategy for reducing bloom risk in the long-term.

36


2

Page 3 of 35


37 38

INTRODUCTION Increases in the frequency, magnitude, duration, and geographic range of harmful

39

cyanobacterial blooms related to cultural eutrophication and climate change (1–3) have

40

raised concerns about undesirable future changes in the structure and function of

41

freshwater ecosystems (4–8). Of particular concern are cyanobacteria species that

42

threaten ecological and public health through the production of toxins, formation of

43

scums that shade out benthic primary producers, and excess accumulation of biomass that

44

can result in hypoxia (9).

45

Although cyanobacteria are found in most aquatic environments, their blooms are

46

particularly widespread and intense in eutrophic freshwater systems (9). The ubiquity of

47

cyanobacteria is attributable to a diversity of physiological adaptations that allow them to

48

successfully compete for limited resources in a wide range of habitat types (1,2,10). Such

49

adaptations include N2 fixation (11,12), luxury nitrogen (N) and phosphorus (P) uptake

50

(10,13), buoyancy regulation using gas vesicles (14), sustained growth at elevated

51

temperatures (7), and grazer avoidance (15,16), among others.

52

Considering the eco-physiological diversity exhibited by cyanobacteria taxa offers

53

insight into how cyanobacteria outcompete other phytoplankton, but it does not

54

necessarily help clarify the factors that drive the dynamics of cyanobacteria populations

55

and formation of bloom events in individual ecosystems (17,18). This issue is central to

56

challenges faced by water managers who are tasked with developing appropriate actions

57

for not only controlling the intensity and frequency of blooms, but also reducing the

58

potential for blooms of harmful taxa (e.g., toxin producers). These challenges are further

59

compounded by an ongoing debate regarding the efficacy of P-only (“single nutrient”)


3


60

versus N-and-P (“dual nutrient”) management strategies for cyanobacteria bloom

61

abatement in aquatic ecosystems (19–23).

62

Page 4 of 35

The primary goal of this study was to address the growing need for increased

63

understanding of the environmental conditions that control blooms of important

64

cyanobacteria genera and to inform water managers’ adoption of nutrient management

65

strategies for bloom mitigation. Our specific objectives were to: (1) quantify and

66

characterize taxa-specific relationships between cyanobacteria biomass and

67

environmental conditions; and (2) evaluate the efficacy of nutrient management strategies

68

relative to several important cyanobacteria taxa. To do so, we applied the Random

69

Forests method, an assumption-free, ensemble-modeling machine learning approach, to a

70

long-term (17-year) dataset composed of monthly observations of physical and chemical

71

parameters, flow, and biomass of cyanobacteria and zooplankton taxa in Lake George, a

72

subtropical flow-through lake of the iconic St. Johns River (Florida). To our knowledge,

73

this is the first investigation to concurrently characterize relationships between bloom-

74

forming cyanobacteria genera and biological, chemical, and physical covariates, as well

75

as to apply the Random Forests methodology to the assessment of cyanobacteria

76

dynamics. Results offer new insights into how diverse cyanobacterial assemblages

77

respond to top-down and bottom-up controls over a long-term record, offering insights

78

that are broadly applicable to other freshwater systems.

79 80 81 82

METHODS Site description – The St. Johns River is a subtropical blackwater system that drains 24,424 km2 of Florida’s northeastern region (USA) (24). The St. Johns flows from


4

Page 5 of 35


83

south to north through approximately 500 km of meandering river reaches and flow-

84

through lakes before discharging into the Atlantic Ocean (Figure S1). The largest of these

85

flow-through lakes is Lake George (18,934 ha), a shallow system (mean depth of 3 m)

86

located approximately 210 km upstream from the mouth of the river (25). Hydrologic

87

inputs to Lake George include the St. Johns River (78.9% of the lake’s volume), rainfall

88

(8.7%), flow from artesian springs (8.3%), and runoff (4.1%) (25). Turnover times in the

89

lake are estimated to range from 24 to 180 days depending on flow conditions (26). Lake

90

George’s waters are amber in color due to the presence of high levels of colored

91

dissolved organic matter, concentrations of which peak from fall to early spring (24).

92

Lake George is eutrophic (TP concentrations range from approximately 30-160

93

µg L-1) and affected by recurrent and severe cyanobacteria blooms (24,27). In terms of

94

carbon biomass, cyanobacteria account for over 90% of the total phytoplankton observed

95

in this system. Cyanobacteria in Lake George display distinct seasonal patterns that

96

diverge from traditional spring-summer-fall-winter variation. Instead, cyanobacteria

97

periodicity in this subtropical, flow-through setting is better explained by “cold” (January

98

– April), “warm” (May – August), and “flushing” (September – December) seasons

99

(27,28). On average, cyanobacteria carbon biomass in Lake George is greatest in the

100 101

warm season, followed by the flushing season (27). Dataset – Data evaluated in this study included monthly water quality

102

observations, phytoplankton and zooplankton species counts, and flow rates from

103

October 1993 to December 2010. Data were collected at a long-term monitoring site

104

(LG1.2, see Figure S1a) located at the outlet of Lake George. Phytoplankton and

105

zooplankton species identifications and enumerations were done using inverted light


5


106

microscopes according to methods described by Srifa et al. (27). Water quality

107

parameters associated with the collection of water for plankton counts were obtained

108

from the St. Johns River Water Management District. Regularly-measured constituents

109

included alkalinity, chlorophyll, chloride, color, conductivity, dissolved oxygen, pH,

110

pheophytin, Secchi depth, sulfate, total dissolved and suspended solids, total organic

111

carbon, turbidity, water temperature, dissolved ammonium, dissolved phosphate, total

112

phosphorus (TP), and total Kjeldahl nitrogen (TKN). Light extinction due to tripton

113

(Ktripton, m-1) was predicted from the chlorophyll-a, color, and Secchi depth data

114

(Equation S1).

115

Page 6 of 35

Zooplankton and cyanobacteria biovolumes were transformed into carbon

116

biomass (mg C L-1) using a carbon conversion factors of 0.075 and 0.22 pg C µm-3,

117

respectively (29,30). Daily flow data were collected in the St. Johns River by the U.S.

118

Geological Survey (USGS) at site 02236000 in Deland, FL (“flow monitoring” site in

119

Figure S1a). The USGS maintains a flow monitoring site nearer the inlet to Lake George

120

(site 02236125 in Astor, FL), but these data included several gaps over the study period.

121

Flow data collected at the Deland and Astor sites were compared to ensure that the

122

Deland data were representative of Lake George’s inflow dynamics (Figure S2); the data

123

were related by a Pearson’s correlation coefficient of 0.96 and a Nash-Sutcliffe

124

Efficiency (Equation S1) of 0.915, leading us to conclude that the use of Deland data was

125

acceptable for this analysis. Data processing details are provided in the Supporting

126

Information.

127

Random Forests modeling – We utilized Random Forest models to quantify and

128

characterize taxa-specific cyanobacteria-environmental relationships. Random Forests is


6

Page 7 of 35


129

a machine learning algorithm used to fit a large ensemble (or, “forest”) of randomly-

130

assembled de-correlated classification (discrete data) or regression (continuous data) trees

131

to bootstrapped samples of a response variable, and the outputs of these trees are

132

averaged to produce a simulated response (31–33). The Random Forests modeling

133

framework was selected for this analysis because of its abilities to (1) handle data

134

produced from complex interactions (33) and (2) uncover nonlinear and linear

135

relationship structures (32,34), making this approach ideal for uncovering the functions

136

relating cyanobacteria genera to environmental and trophic covariates.

137

While Random Forests models are often viewed as “black boxes,” methods exist

138

for quantifying how these statistically derived models relate input explanatory variables

139

to produce simulated responses. Such measures include permutation importance, partial

140

dependence, and relative sensitivity. Permutation importance describes the change in

141

Mean Squared Error (MSE) that occurs when a fitted model is run with a randomly

142

permuted explanatory variable (34,35). Partial dependence is a measure of an explanatory

143

variable’s influence on the response variable given the effects of all other explanatory

144

variables in the model (33,34), and is calculated across the range of each explanatory

145

variable’s observations. Relative sensitivity is calculated from the partial dependence

146

curves to summarize the ranges over which partial dependencies varied relative to

147

changes in the explanatory variables; if partial dependence greatly varies over a narrow

148

range of an explanatory variable (i.e., the partial dependence curve has a steep slope),

149

then the response variable is said to have a high relative sensitivity to that explanatory

150

variable.


7


151

Models were fit and tested through 5-fold cross validation (33), which was

152

performed using the following steps: (1) partition dataset into training and testing folds,

153

(2) “grow” Random Forests, (3) quantify model performance using the Nash-Sutcliffe

154

Efficiency, (4) tabulate permutation importance, partial dependence, and relative

155

sensitivity of all the explanatory variables, and (5) repeat steps (2)-(4) 4 times, with each

156

new fold representing the “testing” set on each iteration. This modeling framework was

157

repeated 5 times for each response variable to evaluate variability in model outputs; thus

158

a total of 25 random forest models were produced with each application (5 repetitions of

159

each 5-fold cross validation). Details associated with each of these steps are provided in

160

the Supporting Information.

161

Response and explanatory variables – The dominant cyanobacteria genera –

162

defined here as genera that accounted for 5% or greater of the total cyanobacteria carbon

163

biomass (mg C L-1) over the study period – were designated as the response variables in

164

the random forests modeling analysis.

165

Page 8 of 35

TKN (mg L-1), TP (mg L-1), dissolved NH4 (mg L-1), dissolved PO4 (mg L-1),

166

water temperature (°C), average flow (m3 s-1), partial light extinction due to tripton

167

(Ktripton, m-1), color (CPU), copepods (mg C L-1), and rotifers (mg C L-1) were included as

168

explanatory variables in the Random Forest models. Ktripton was calculated from Secchi

169

depth, chlorophyll-a, and color (see Supporting Information). This reduced set of

170

factors was selected based on reported relationships between these covariates and

171

cyanobacteria (2,25,27,36–38). Excluded variables are outlined in the Supporting

172

Information, along with justification for their exclusion. All explanatory variables were


8

Page 9 of 35


173

lagged by one time step (1 month) to reflect the division rates of cyanobacteria in natural

174

environments (13,39).

175 176 177

RESULTS Summary of observations – The Anabaena (N2 fixers), Cylindrospermopsis (N2

178

fixers), Planktolyngbya (non-fixers), Microcystis (non-fixers), and Oscillatoria (non-

179

fixers) groups dominated the Lake George cyanobacteria community over the study

180

period (Table S1). It is important to note that the Oscillatoria group includes several

181

genera (e.g., Planktothrix and Limnothrix) that were further divided from Oscillatoria

182

and renamed during the data collection period (40); similarly, the Anabaena group

183

includes members of the genus recently renamed to Dolichospernum (41). Given that

184

these changes occurred well into the study period, we have reported these groups per the

185

prior nomenclature since observations were recorded in context of these broader

186

classifications.

187

Cyanobacteria biomass displayed strong inter-annual (Figure S3) and seasonal

188

variability (Figure S4) across the dominant genera. Of the dominant genera, Oscillatoria

189

produced the greatest total biomass over the study period, and Anabaena the least (Table

190

S1). Interestingly, Microcystis regularly bloomed prior to 2001, and minimally varied at

191

low biovolumes thereafter (Figure S3). Cyanobacteria predominantly flourished in the

192

warm season, followed by the flushing season (Figure S4); Anabaena typically bloomed

193

prior to the other genera on a seasonal basis (Figure S4). Additional description of the

194

observed data is included in the Supporting Information.


9


195

Page 10 of 35

Model performance – Models fit the training data well, as measured by Nash

196

Sutcliffe Efficiency (NSE) across the 25 model applications per response variable (Figure

197

1a); the minimum NSE value was 0.77 (an Anabaena model) and the max was 0.91 (a

198

Planktolyngbya model). Previous studies have proposed that models with NSE > 0.65 are

199

“acceptable” (42), indicating that all the training Random Forest models presented here

200

satisfactorily simulated cyanobacteria genera dynamics. However, models did not

201

perform as well when predicting the testing data (Figure 1b). Specifically, the model

202

suites included efficiencies below 0 when evaluated against the testing data, which

203

suggests that findings reported here were not generalizable across the entire study period;

204

rather, marked effects of particular relationships were important for subsets of the period

205

of record.

206

Permutation importance – Color and water temperature were associated with

207

the greatest permutation importance (quantified through percent change in MSE) in the

208

Cylindrospermopsis (Figure 2b), Planktolyngbya (Figure 2c), and Oscillatoria (Figure

209

2e) models; these explanatory variables were also among the most important in the

210

Anabaena (Figure 2a) and Microcystis (Figure 2d) models. Overall, permutation of

211

explanatory variables did not dramatically change the percent MSE in any of the models.

212

However, water temperature appeared to be a particularly influential variable in the

213

Planktolyngbya model (Figure 2c). As the Planktolyngbya model suite had the highest

214

testing NSE results, the variable importance results suggest that permuted importance

215

also served as a predictor of model testing success.

216 217

Partial dependence – Plots of partial dependence revealed predominantly nonlinear model relationships between the lagged explanatory variables and


10

Page 11 of 35


218

cyanobacteria genera (Figure 3). Partial dependence curves correspond to the

219

relationships linking response and explanatory variables, and the curves are comprised of

220

the average modeled values (in units of the response variable) across the range of

221

explanatory variable observations. The steepest curves were associated with TKN, TP,

222

average flow, color, water temperature, and copepod abundance, whereas the curves for

223

NH4, PO4, Ktripton, and rotifer abundance curves were largely invariant.

224

Non-fixing cyanobacteria genera (Planktolyngbya, Microcystis, and Oscillatoria)

225

displayed a relatively strong partial dependence on TKN (Figure 3a). Specifically, these

226

curves depict a threshold in the relationship between TKN and non-fixers where partial

227

dependence rose sharply TKN concentrations between 1.5-2.5 mg L-1, but plateaued in

228

the range of 2.5-3 mg L-1. Note, however, that there were few observations at high TKN

229

concentrations, limiting interpretation in this range. Anabaena and Cylindrospermopsis

230

showed little partial dependence on TKN; however, these partial dependence curves

231

revealed a slight sensitivity of N2 fixers to TKN at low TKN concentrations.

232

The cyanobacteria genera largely lacked partial dependence on TP, with

233

Microcystis being an exception (Figure 3b). Microcystis’ partial dependence on TP

234

indicated a sensitivity to TP at low concentrations (decreasing carbon biomass with

235

increasing TP) up until a concentration approximately 0.1 mg L-1, above which there was

236

no additional dependence (i.e., similar to the “threshold” response described above).

237

Average flow, color, and water temperature emerged as explanatory variables on

238

which all of the cyanobacteria genera displayed notable partial dependence (Figure 3e).

239

Specifically, cyanobacteria were more abundant at low flows and colors, and higher

240

temperatures. Although the color partial dependence curves were essentially identical


11


241

across all cyanobacteria genera except Anabaena (Figure 3f), which peaked at

242

approximately 120 CPU.

Page 12 of 35

243

Relative sensitivity – Relative sensitivities were calculated based on the median

244

scaled partial dependence curves (Figure 4). The cyanobacteria genera were, on average,

245

most sensitive to the physical covariates, followed by nutrient concentrations, then

246

zooplankton abundance (average relative sensitivities to physical factors, nutrients, and

247

zooplankton across all genera equaled 1.29, 0.45, and 0.41, respectively). Planktolyngbya

248

(Figure 4c) and Oscillatoria (Figure 4e) were most sensitive to changes in water

249

temperature, Anabaena (Figure 4a) and Microcystis (Figure 4d) were most sensitive to

250

changes in average flow, and Cylindrospermopsis (Figure 4b) was most sensitive to water

251

color. With regards to nutrients, the genera were generally more sensitive to TKN than

252

TP, though Microcystis was similarly sensitive to TKN and TP (Figure 4d).

253 254 255

DISCUSSION Historically, a number of researchers have argued that nutrient management is the

256

most effective strategy for preventing cyanobacteria blooms regardless of the dominant

257

taxa involved (17,43,44). While the results of this study provide some support for the

258

potential efficacy of nutrient reduction strategies, certain aspects of the results highlight

259

how nutrients alone do not always adequately explain trends in biomass of certain

260

cyanobacteria taxa of particular concern from a management perspective (e.g. harmful

261

algal bloom species). The importance of the latter caveat is illustrated by the relative

262

sensitivity of different cyanobacteria genera to the range of explanatory variables

263

included in this modeling effort. Overall, the genera were relatively more sensitive to


12

Page 13 of 35


264

TKN, TP, average flow, color, and copepod abundance. Because the genera were found

265

to be largely insensitive to changes in NH4, PO4, light extinction due to tripton, and

266

rotifer abundance, these covariates were excluded from the ensuing discussion.

267

Sensitivity of Cyanobacterial Biomass to Nitrogen and Phosphorus Levels

268

N2 fixing cyanobacteria were relatively insensitive to TKN (Figure 2a) due to

269

their ability to utilize N2 as a nitrogen source for growth (13). In contrast, non-fixing

270

genera showed increasing partial dependence on TKN with increasing concentrations

271

until a threshold of 1.5-2.5 mg L-1 (Figure 3a, 2a). This finding indicates that

272

management actions seeking to reduce TKN may produce observable decreases in non-

273

fixer cyanobacteria biomass when this concentration threshold is crossed. The sharpest

274

thresholding behavior was associated with Microcystis, which reflects the observation

275

that after 2001 no major Microcystis blooms were observed in Lake George (Figure S3).

276

The post-2001 period was also characterized by lower baseline and peak levels of TKN.

277

It is possible that the relatively strong reliance of Microcystis on external nitrogen made it

278

less competitive during the post-2001 period, as observed in some other ecosystems (45–

279

47).

280

Most cyanobacteria genera were less sensitive to changes in TP concentrations

281

than to TKN, as demonstrated by the flatter TP partial dependence curves (Figure 3b).

282

The somewhat elevated dependence of Microcystis biomass on the lower end of the TP

283

concentration scale may reflect how elevated biomass levels generally appeared in mid-

284

late summer when TP levels in the water column were often lower than in spring or early

285

summer (Figure 3d). Microcystis aeruginosa is well known to have strong buoyancy

286

regulation capacity (13,48). It is possible that during mid-summer, when average wind


13


Page 14 of 35

287

mixing energy is lower than in the spring, M. aeruginosa takes advantage of benthic

288

fluxes of bioavailable phosphorus, which are enhanced by lower oxygen levels near the

289

bottom of the water column (49,50), particularly under high summer water temperatures.

290

The ability of buoyancy-regulating cyanobacteria to take advantage of elevated nutrient

291

levels in different layers of the water column has been observed in other ecosystems (51).

292

In more general terms, many of the major bloom-forming cyanobacteria species in Lake

293

George are known to have differing degrees of buoyancy regulation (52), which may

294

contribute to the apparent low sensitivity to TP levels, since sediment phosphorus

295

reserves can serve as sources of nutrients somewhat independent of water column

296

concentrations (51).

297

Implications for Nutrient-Based Cyanobacteria Bloom Mitigation

298

The cyanobacteria-nutrient relationships identified here provide some insight into

299

the ongoing discussions in the literature of two nutrient management approaches:

300

phosphorus only (19,20) versus dual control of both nitrogen and phosphorus (21–23).

301

Although our results are from a single, shallow, subtropical lake, they highlight nuances

302

that are broadly applicable when weighing the potential efficacy of these two approaches.

303

The insensitivity of the cyanobacteria groups’ biomass to the range of TP values

304

observed in Lake George raises questions about whether phosphorus reduction alone will

305

significantly reduce intensities of certain types of cyanobacteria blooms. While the results

306

of recent nutrient enrichment bioassay experiments in Lake George identified

307

phosphorus- and phosphorus/nitrogen co-limited conditions for phytoplankton growth,

308

these conclusions were produced from data collected during an extreme drought in 2000,

309

when TP concentrations were very low (e.g., near 30 µg L-1) (25). Additionally, the


14

Page 15 of 35


310

modest sensitivity of N2 fixing species biomass to TP values below 0.06 µg L-1 observed

311

in the Random Forests results, provides some limited evidence for the potential control of

312

this functional group, possibly related to the relatively high demand some nitrogen fixers

313

have for phosphorus (53,54). Recent mechanistic modeling efforts by the St. Johns River

314

Water Management District identified an annual mean threshold of 0.05-0.063 mg TP L-1

315

as a long-term target that would be effective for mitigating blooms (50). The differences

316

between the findings of these prior studies and those presented here are challenging to

317

reconcile, but could be explained by the aforementioned influences of internal loads from

318

sediments and water column cycling of nutrients, as observed in some ecosystems (51).

319

The range of results among this and other studies on cyanobacteria-phosphorus

320

relationships in Lake George warrant further examination, focusing on the feasibility of

321

reducing phosphorus levels within the context of both internal and external phosphorus

322

sources.

323

In contrast, the sensitivity of non-N2 fixing cyanobacteria to TKN (Figure 2a)

324

suggests that significant reductions in nitrogen load do have the potential to reduce bloom

325

intensities of those genera, though N2-fixing cyanobacteria are unlikely to be affected.

326

Combined with those of TP, these TKN results suggest that the potential exists for

327

control of N2 fixers through phosphorus reductions, and non-fixers through nitrogen

328

reductions. However, other studies have shown that N2 fixation in many lakes may

329

represent less than 50% of the N2 fixer nitrogen demand (55–57); therefore, these

330

modeling results should be interpreted in the context of other experimental findings.

331 332

The value of considering nitrogen load reduction in part depends on the specific management priorities for Lake George. For example, toxic blooms of the non-N2 fixing


15


Page 16 of 35

333

cyanobacterium Microcystis aeruginosa are a recurring problem for Lake George and the

334

downstream portions of the lower St. Johns River (58). The negative response of

335

Microcystis and other non-N2 fixing cyanobacterial biomass to TKN concentrations

336

below 1.5 mg L-1 (Figure 3a) suggests that significant reductions in external nitrogen

337

loads may contribute to less intense blooms of these groups. However, this conclusion

338

may be in part dependent on the response of the N2 fixing cyanobacteria to reduced

339

nitrogen loads. Currently, the N2 fixers Anabaena (most prominently the renamed genus

340

Dolichospermum) and Cylindrospermopsis are major contributors to the nitrogen budget

341

of Lake George (59).

342

The latter observation raises two important considerations related to possible

343

consequences of nitrogen reduction strategies: (1) the relative importance of Anabaena

344

and Cylindrospermopsis biomass may increase relative to non-fixers and result in

345

alternative challenges to the ecosystem, and (2) increases in N2 fixer biomass may lead to

346

increases in nitrogen loads via N2 fixation to the system, which could ultimately fuel non-

347

fixer blooms, particularly downstream of the lake. These considerations, and those

348

associated with phosphorus reduction scenarios, argue for exploration of a dual

349

phosphorus and nitrogen management strategy. The potential feasibility or efficacy of

350

such an approach in other lakes more generally depends on the specific structure of

351

individual ecosystems and the character of their external and internal sources of nutrient

352

load.

353

Sensitivity of Cyanobacterial Biomass to Other Abiotic and Biotic Variables

354 355

Beyond nutrient-based management strategies, our study also identified important connections among cyanobacteria and other abiotic and biotic factors, which help to put


16

Page 17 of 35


356

the nutrient response observations into a more holistic management context. One of the

357

challenges facing water managers is that many of these factors (e.g., water temperature,

358

flow, color) are largely uncontrollable in the short-term, even though human activities

359

will likely have strong impacts on them in future decades and centuries, such as changes

360

in temperature, rainfall (e.g. as it relates to watershed inputs) and hydrologic conditions

361

(e.g. sea level rise, lake stage). For example, the strong positive relationship between

362

cyanobacteria biomass and temperature also corroborates the hypotheses of a number of

363

researchers that anticipated future global temperature increases will selectively favor

364

cyanobacteria (4,7,60).

365

It is also important to examine potential interactive effects among abiotic

366

explanatory variables to further define the extent to which nutrients influence

367

cyanobacteria dynamics, particularly when certain nutrients are not present in limiting

368

concentrations (17,61–63), as appears to be the case for phosphorus in Lake George.

369

High flow conditions in the fall and winter not only lead to reduced water residence times

370

and reduced light transmission due to elevated color levels, but occur during the period of

371

lowest incident light flux and water temperatures, all of which help to explain the low

372

potential for blooms of phytoplankton in general, despite the fact that it is also the period

373

of highest nutrient concentrations (24,27,36). The importance of these relationships is

374

reflected in the strong negative correlations between cyanobacterial biomass and both

375

flow and color.

376

Previous studies of Lake George have highlighted how hydrologic conditions can

377

override or exaggerate the effects of other environmental factors, such as nutrient levels,

378

on cyanobacteria dynamics (27), corroborating the results presented here. Understanding


17


Page 18 of 35

379

the degree to which the hydrologic regime decouples cyanobacteria from nutrient

380

dynamics is critical to understanding the potential efficacy of water quality management

381

practices. We caution, however, that there are substantial physical and/or fiscal barriers

382

associated with changing a system’s hydrology. Unless managers have the capacity to store

383

and release large amounts of freshwater, their ability to control bloom potentials is likely limited

384

to nutrient management. Therefore, it is ultimately the “nutrient knob” (64) that provides the most

385

reliable management strategy for long-term bloom mitigation (43,44,64).

386

Another consideration for defining the dynamics of cyanobacteria populations is

387

the role of top-down controls, both in terms of direct grazing pressure and indirect trophic

388

cascade effects. The data available for the two dominant groups of zooplankton in Lake

389

George, rotifers and copepods, provide some insight into top-down effects. Non-N2 fixing

390

cyanobacteria had a positive near linear relationship with rotifer biomass, suggesting that

391

rotifers respond positively to increases in cyanobacteria, but do not appear to be able to

392

suppress bloom formation, as indicated by the lack of depression of cyanobacteria

393

biomass even at the highest abundances of rotifers (Figure 3j). The relative resistance of

394

many cyanobacteria to strong top-down control by zooplankton grazers has been widely

395

discussed in the literature, as it relates to the presence of grazing inhibitors (2,15,16), or

396

secondary trophic effects (60,65).

397

For copepods, the relationships to cyanobacteria were positive but non-linear,

398

with a narrow threshold range of response (Figure 3i). Interpreting the implications of

399

this relationship is complicated by the diverse trophic character of copepods, which

400

include many carnivorous, as well as, omnivorous taxa (66). The sharp response

401

threshold for the cyanobacteria-copepod relationship may in part reflect the seasonal

402

coincidence of peaks in copepod and cyanobacteria biomass. In addition, top-down


18

Page 19 of 35


403

pressure by copepods on smaller zooplankton grazers of phytoplankton (e.g. protozoans,

404

rotifers) may lower grazing pressure on cyanobacteria.

405

Methodological Implications

406

Ideally, mechanistic water quality modeling would effectively simulate time-

407

varying dynamics of cyanobacteria remains, however the capacity of such models

408

remains limited (67,68). At present, data-intensive, assumption-free methods (e.g.,

409

Random Forests models) are among the most rigorous tools available for the study of

410

complex organism community dynamics using long-term monitoring data. The Random

411

Forests machine learning algorithm utilized here offers a straightforward and assumption-

412

free approach to developing an initial appraisal of how important potentially manageable

413

(e.g., nutrient loads) versus non-manageable (e.g., water temperature) system features are

414

as predictors of cyanobacteria dynamics.

415

Although the Random Forests models were employed for explanatory modeling

416

(i.e., describing observed dynamics) in the case presented here, the use of 5-repeated 5-

417

fold cross validation provided insight into the degree to which the models could be used

418

for general prediction purposes. All of the model suites performed reasonably well when

419

applied to testing data (Figure 1b), which suggests that the explanatory variables

420

considered in this analysis are important descriptors of cyanobacteria dynamics.

421

However, some efficiencies associated with individual testing model runs were low (i.e.,

422

NSE < 0), which suggests that additional descriptors should be incorporated into the

423

models to increase their predictive capacities. Future research efforts should consider

424

evaluating whether cyanobacteria genera, as opposed to environmental covariates, better

425

predict the dynamics of other genera. Moreover, as this study was limited by the temporal


19


Page 20 of 35

426

resolution of the long-term monitoring data program, opportunities exist to expand this

427

analysis by repeating it in a setting with higher resolution data; given higher resolution

428

data, the robustness of the sensitivities and nonlinearities reported here could be assessed

429

across a variety of temporal scales.

430

Applying Random Forests models to a broad variety of aquatic systems impacted

431

by cyanobacteria blooms would help to advance our understanding of the factors that can

432

be managed to mitigate blooms, and such applications should be the focus of future

433

studies. This analysis serves as a case study to guide other researchers and practitioners in

434

their efforts to leverage machine learning tools for the study of water quality phenomena.

435 436 437

ACKNOWLEDGEMENTS This material is based upon work supported by the National Science Foundation

438

Graduate Research Fellowship under Grant No. DGE-0802270, USDA NIFA Hatch

439

Project 1011481, and St. John River Water Management District Contract No. 28650.

440

RMC acknowledgers support for the UF Water Institute Fellowship. The authors declare

441

that there are no conflicts of interest.

442 443

SUPPORTING INFORMATION

444

Document containing additional details on data processing, modeling workflow, and

445

summary of observed data.

446


20

Page 21 of 35


447

REFERENCES

448

(1)

449 450

(6157), 433–434. (2)

451 452

Paerl, H. W.; Otten, T. G. Blooms bite the hand that feeds them. Science 2013, 342

Paerl, H. W.; Fulton, R. S.; Moisander, P. H.; Dyble, J. Harmful freshwater algal blooms, with an emphasis on cyanobacteria. Sci. World 2001, 1, 76–113.

(3)

Glibert, P.; Anderson, D.; Gentien, P.; Granéli, E.; Sellner, K. The Global,

453

Complex Phenomena of Harmful Algal Blooms. Oceanography 2005, 18 (2), 136–

454

147.

455

(4)

456 457

Paerl, H. W.; Huisman, J. Climate change: A catalyst for global expansion of harmful cyanobacterial blooms. Environ. Microbiol. Rep. 2009, 1 (1), 27–37.

(5)

O’Neil, J. M.; Davis, T. W.; Burford, M. A.; Gobler, C. J. The rise of harmful

458

cyanobacteria blooms: The potential roles of eutrophication and climate change.

459

Harmful Algae 2012, 14, 313–334.

460

(6)

Carey, C. C.; Ibelings, B. W.; Hoffmann, E. P.; Hamilton, D. P.; Brookes, J. D.

461

Eco-physiological adaptations that favour freshwater cyanobacteria in a changing

462

climate. Water Res. 2012, 46 (5), 1394–1407.

463

(7)

Paerl, H. W.; Huisman, J. Blooms Like It Hot. Science (80-. ). 2008, 320, 57–58.

464

(8)

Havens, K.; Paerl, H.; Phlips, E.; Zhu, M.; Beaver, J.; Srifa, A. Extreme Weather

465

Events and Climate Variability Provide a Lens to How Shallow Lakes May

466

Respond to Climate Change. Water 2016, 8 (6), 229.

467

(9)

468 469

Paerl, H. W. Nuisance phytoplankton blooms in coastal, estuarine, and inland waters. Limnol. Oceanogr. 1988, 33 (4_part_2), 823–843.

(10)

Oliver, R. L.; Hamilton, D. P.; Brookes, J. D.; Ganf, G. G. Physiology, Blooms


21


Page 22 of 35

470

and Prediction of Planktonic Cyanobacteria. In Ecology of Cyanobacteria II: Their

471

Diversity in Space and Time; 2012; pp 1–13.

472

(11)

Howarth, R. W.; Marino, R.; Cole, J. J. Nitrogen fixation in freshwater, estuarine,

473

and marine ecosystems. 1. Rates and importance. Limnol. Oceanogr. 1988, 33 (4,

474

part 2), 688–701.

475

(12)

Howarth, R. W.; Marino, R.; Cole, J. J. Nitrogen fixation in freshwater, estuarine,

476

and marine ecosystems. 2. Biogeochemical control. Limnol. Oceanogr. 1988, 33

477

(4, part 2), 688–701.

478

(13)

479 480

Reynolds, C. Ecology of Phytoplankton; Cambridge University Press: Cambridge, UK, 2006.

(14)

Oliver, R. L.; Walsby, A. E. Direct evidence for the role of light-mediated gas

481

vesicle collapse in the buoyancy regulation of Anabaena flos-aquae

482

(cyanobacteria). Limnol. Ocean. 1984, 29 (4), 879–886.

483

(15)

484 485

Lampert, W. Laboratory studies on zooplankton-cyanobacteria interactions. New Zeal. J. Mar. Freshw. Res. 1987, 21, 483–490.

(16)

Wilson, A. E.; Sarnelle, O.; Tillmanns, A. R. Effects of cyanobacterial toxicity and

486

morphology on the population growth of freshwater zooplankton: Meta-analyses

487

of laboratory experiments. Limnol. Ocean. 2006, 51 (4), 1915–1924.

488

(17)

Paerl, H. W.; Otten, T. G. Duelling “CyanoHABs”: Unravelling the environmental

489

drivers controlling dominance and succession among diazotrophic and non-N2-

490

fixing harmful cyanobacteria. Environ. Microbiol. 2016, 18 (2), 316–324.

491 492

(18)

Reynolds, C. S.; Huszar, V.; Kruk, C.; Naselli-Flores, L.; Melo, S. Towards a functional classification of the freshwater phytoplankton. J. Plankton Res. 2002,


22

Page 23 of 35


493 494

24 (5), 417–428. (19)

Schindler, D. W.; Hecky, R. E.; Findlay, D. L.; Stainton, M. P.; Parker, B. R.;

495

Paterson, M. J.; Beaty, K. G.; Lyng, M.; Kasian, S. E. M. Eutrophication of lakes

496

cannot be controlled by reducing nitrogen input: Results of a 37-year whole-

497

ecosystem experiment. Proc. Natl. Acad. Sci. 2008, 105 (32), 11254–11258.

498

(20)

Schindler, D. W.; Carpenter, S. R.; Chapra, S. C.; Hecky, R. E.; Orihel, D. M.

499

Reducing phosphorus to curb lake eutrophication is a success. Environ. Sci.

500

Technol. 2016, 50 (17), 8923–8929.

501

(21)

Conley, D. J.; Paerl, H. W.; Howarth, R. W.; Boesch, D. F.; Seitzinger, S. P.;

502

Havens, K. E.; Lancelot, C.; Likens, G. E. Controlling eutrophication: nitrogen and

503

phosphorus. Science (80-. ). 2009, 323 (5917), 1014–1015.

504

(22)

Paerl, H. W. Controlling Eutrophication along the Freshwater–Marine Continuum:

505

Dual Nutrient (N and P) Reductions are Essential. Estuaries and Coasts 2009, 32

506

(4), 593–601.

507

(23)

Paerl, H. W.; Scott, J. T.; McCarthy, M. J.; Newell, S. E.; Gardner, W. S.; Havens,

508

K. E.; Hoffman, D. K.; Wilhelm, S. W.; Wurtsbaugh, W. A. It Takes Two to

509

Tango: When and Where Dual Nutrient (N & P) Reductions Are Needed to Protect

510

Lakes and Downstream Ecosystems. Environ. Sci. Technol. 2016, 50 (20), 10805–

511

10813.

512

(24)

Phlips, E. J.; Cichra, M.; Aldridge, F. J.; Jembeck, J.; Hendrickson, J.; Brody, R.

513

Light availability and variations in phytoplankton standing crops in a nutrient-rich

514

blackwater river. Limnol. Oceanogr. 2000, 45 (4), 916–929.

515

(25)

Piehler, M. F.; Dyble, J.; Moisander, P. H.; Chapman, A. D.; Hendrickson, J.;


23


Page 24 of 35

516

Paerl, H. W. Interactions between nitrogen dynamics and the phytoplankton

517

community in Lake George, Florida, USA. Lake Reserv. Manag. 2009, 25 (1), 1–

518

14.

519

(26)

Stewart, J.; Sucsy, P.; Hendrickson, J. Meteorological and subsurface factors

520

affecting estuarine conditions within Lake George in teh St Johns River, Florida.

521

In 7th International Conference on HydroScience and Engineering; Philadelphia,

522

USA, 2006.

523

(27)

Srifa, A.; Phlips, E. J.; Cichra, M. F.; Hendrickson, J. C. Phytoplankton dynamics

524

in a subtropical lake dominated by cyanobacteria: cyanobacteria “Like it Hot” and

525

sometimes dry. Aquat. Ecol. 2016, 1–12.

526

(28)

Srifa, A.; Phlips, E. J.; Hendrickson, J. How many seasons are there in a sub-

527

tropical lake? A multivariate statistical approach to determine seasonality and its

528

application to phytoplankton dynamics. Limnologica 2016, 60, 39–50.

529

(29)

530 531

Latja, R.; Salonen, K. Carbon analysis for the determination of individual biomass of planktonic animals. Verh. Internat. Verein Limnol. 1978, 20, 2556–2560.

(30)

Work, K.; Havens, K.; Sharfstein, B.; East, T. How important is bacterial carbon to

532

planktonic grazers in a turbid, subtropical lake? J. Plankton Res. 2005, 27 (4),

533

357–372.

534

(31)

Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32.

535

(32)

Cutler, D. R.; Edwards, T. C.; Beard, K. H.; Cutler, A.; Hess, K. T.; Gibson, J.;

536

Lawler, J. J. Random forests for classification in ecology. Ecology 2007, 88 (11),

537

2783–2792.

538

(33)

Hastie, T.; Tibshirani, R.; Friedman, J. The Elements of Statistical Learning: Data


24

Page 25 of 35


539

Mining, Inference, and Prediction, Second.; Springer Science & Business Media:

540

New York, NY, 2009.

541

(34)

542 543

Annual MPSA Conference; 2015; pp 1–31. (35)

544 545

Jones, Z.; Linder, F. Exploratory Data Analysis using Random Forests. In 73rd

Jones, Z. M.; Linder, F. J. Edarf: Exploratory Data Analysis using Random Foresets. J. Open Source Softw. 2016, 1 (6).

(36)

Phlips, E. J.; Hendrickson, J.; Quinlan, E. L.; Cichra, M. Meteorological influences

546

on algal bloom potential in a nutrient-rich blackwater river. Freshw. Biol. 2007, 52

547

(11), 2141–2155.

548

(37)

Leonard, J. A.; Paerl, H. W. Zooplankton community structure, micro-zooplankton

549

grazing impact, and seston energy content in the St. Johns river system, Florida as

550

influenced by the toxic cyanobacterium Cylindrospermopsis raciborskii.

551

Hydrobiologia 2005, 537 (1–3), 89–97.

552

(38)

553 554

Paerl, H. W.; Otten, T. G. Harmful Cyanobacterial Blooms: Causes, Consequences, and Controls. Microb. Ecol. 2013, 65, 995–1010.

(39)

Stolte, W.; Garcés, E. Ecological aspects of harmful algal in situ population

555

growth rates. In Ecology of Harmful Algae; Springer-Verlag: Berlin, 2006; pp

556

139–152.

557

(40)

Suda, S.; Watanabe, M. M.; Otsuka, S.; Mahakahant, A.; Yongmanitchai, W.;

558

Nopartnaraporn, N.; Liu, Y.; Day, J. G. Taxonomic revision of water-bloom-

559

forming species of oscillatorioid cyanobacteria. Int. J. Syst. Evol. Microbiol. 2002,

560

52 (2002), 1577–1595.

561

(41)

Wacklin, P.; Hoffmann, L.; Komárek, J. Nomenclatural validation of the


25


562

genetically revised cyanobacterial genus Dolichospermum (Ralfs ex Bornet et

563

Flahault) comb. nova. Fottea 2009, 9 (1), 59–64.

564

(42)

Page 26 of 35

Ritter, A.; Muñoz-Carpena, R. Performance evaluation of hydrological models:

565

Statistical significance for reducing subjectivity in goodness-of-fit assessments. J.

566

Hydrol. 2013, 480, 33–45.

567

(43)

Mantzouki, E.; Visser, P. M.; Bormans, M.; Ibelings, B. W. Understanding the key

568

ecological traits of cyanobacteria as a basis for their management and control in

569

changing lakes. Aquat. Ecol. 2016, 50 (3), 333–350.

570

(44)

Rigosi, A.; Carey, C. C.; Ibelings, B. W.; Brookes, J. D. The interaction between

571

climate warming and eutrophication to promote cyanobacteria is dependent on

572

trophic state and varies among taxa. Limnol. Ocean. 2014, 59 (1), 99–114.

573

(45)

Gerloff, G. C.; Skoog, F. Nitrogen as a Limiting Factor for the Growth of

574

Microcystis Aeruginosa in Southern Wisconsin Lakes. Ecology 1957, 38 (4), 556–

575

561.

576

(46)

Baldia, S. F.; Evangelista, A. D.; Aralar, E. V.; Santiago, A. E. Nitrogen and

577

phosphorus utilization in the cyanobacterium Microcystis aeruginosa isolated from

578

Laguna de Bay, Philippines. J. Appl. Phycol. 2007, 19, 607–613.

579

(47)

Peng, G.; Fan, Z.; Wang, X.; Chen, C. Photosynthetic response to nitrogen source

580

and different ratios of nitrogen and phosphorus in toxic cyanobacteria, Microcystis

581

aeruginosa FACHB-905. J. Limnol. 2016, 75 (3), 560–570.

582

(48)

Visser, P. M.; Ibelings, B. W.; Mur, L. R.; Walsby, A. E. The Ecophysiology of

583

the Harmful Cyanobacterium Microcystis. In Harmful Cyanobacteria; 2005; pp

584

109–142.


26

Page 27 of 35

585


(49)

Malecki, L. M.; White, J. R.; Reddy, K. R. Nitrogen and Phosphorus Flux Rates

586

from Sediment in the Lower St. Johns River Estuary. J. Environ. Qual. 2004, 33,

587

1545–1555.

588

(50)

589 590

Hendrickson, J.; Mattson, R.; Sucsy, P. Recommended Ecosystem Performance Targets to Achieve Designated Use in Lake George, Florida; Palatka, FL, 2017.

(51)

Cottingham, K. L.; Ewing, H. A.; Greer, M. L.; Carey, C. C.; Weathers, K. C.

591

Cyanobacteria as biological drivers of lake nitrogen and phosphorus cycling.

592

Ecosphere 2015, 6 (1), 1–19.

593

(52)

Reynolds, C. S.; Oliver, R. L.; Walsby, A. E. Cyanobacterial dominance: The role

594

of buoyancy regulation in dynamic lake environments. New Zeal. J. Mar. Freshw.

595

Res. 1987, 21, 379–390.

596

(53)

Andersson, A.; Höglander, H.; Karlsson, C.; Huseby, S. Key role of phosphorus

597

and nitrogen in regulating cyanobacterial community composition in the northern

598

Baltic Sea. Estuar. Coast. Shelf Sci. 2015, 164, 161–171.

599

(54)

Vahtera, E.; Conley, D. J.; Gustafsson, B. G.; Kuosa, H.; Pitkänen, H.; Savchuk,

600

O. P.; Tamminen, T.; Viitasalo, M.; Voss, M.; Wasmund, N.; Wulff, F. Internal

601

Ecosystem Feedbacks Enhance Nitrogen-fixing Cyanobacteria Blooms and

602

Complicate Management in the Baltic Sea. Ambio 2007, 36 (2–3), 186–194.

603

(55)

Scott, J. T.; Doyle, R. D.; Prochnow, S. J.; White, J. D. Are watershed and

604

lacustrine controls on planktonic N2 fixation hierarchically structured? Ecol. Appl.

605

2008, 18 (3), 805–819.

606 607

(56)

Paerl, H. W.; Scott, J. T. Throwing fuel on the fire: Synergistic effects of excessive nitrogen inputs and global warming on harmful algal blooms. Environ. Sci.


27


608 609

Technol. 2010, 44 (20), 7756–7758. (57)

Scott, J. T.; McCarthy, M. J. Nitrogen fixation may not balance the nitrogen pool

610

in lakes over timescales relevant to eutrophication management. Limnol.

611

Oceanogr. 2010, 55 (3), 1265–1270.

612

(58)

Hendrickson, J.; Lowe, E. F.; Ph, D.; Dobberfuhl, D.; Ph, D.; Campbell, D.

613

Characteristics of Accelerated Eutrophication in the Lower St . Johns River

614

Estuary and Recommended Targets to Achieve Water Quality Goals for the

615

Fulfillment of TMDL and PLRG Objectives; Palatka, Florirda, 2003.

616

(59)

617 618

Doron, M. Aquatic nitrogen fixation: patterns, rates and controls in a shallow, subtropical lake, University of Florida, 2010.

(60)

619 620

Page 28 of 35

Paerl, H. W.; Paul, V. J. Climate change: Links to global expansion of harmful cyanobacteria. Water Res. 2012, 46 (5), 1349–1363.

(61)

Soares, M. C. S.; Rocha, M. I. D. A.; Marinho, M. M.; Azevedo, S. M. F. O.;

621

Branco, C. W. C.; Huszar, V. L. M. Changes in species composition during annual

622

cyanobacterial dominance in a tropical reservoir : physical factors , nutrients and

623

grazing effects. Aquat. Microb. Ecol. 2009, 57 (November), 137–149.

624

(62)

Arhonditsis, G. B.; Stow, C. a.; Paerl, H. W.; Valdes-Weaver, L. M.; Steinberg, L.

625

J.; Reckhow, K. H. Delineation of the role of nutrient dynamics and hydrologic

626

forcing on phytoplankton patterns along a freshwater-marine continuum. Ecol.

627

Modell. 2007, 208 (2–4), 230–246.

628

(63)

Bormans, M.; Ford, P. W.; Fabbro, L. Spatial and temporal variability in

629

cyanobacterial populations controlled by physical processes. J. Plankton Res.

630

2005, 27 (1), 61–70.


28

Page 29 of 35

631


(64)

Paerl, H. W.; Hall, N. S.; Peierls, B. L.; Rossignol, K. L. Evolving Paradigms and

632

Challenges in Estuarine and Coastal Eutrophication Dynamics in a Culturally and

633

Climatically Stressed World. Estuaries and Coasts 2014, 37 (2), 243–258.

634

(65)

635 636

Ecol. Monogr. 1993, 63 (2), 129–149. (66)

637 638

Sarnelle, O. Herbivore Effects on Phytoplankton Succession in a Eutrophic Lake.

Kleppel, G. S. On the diets of calanoid copepods. Mar. Ecol. Prog. Ser. 1993, 99, 183–195.

(67)

Shimoda, Y.; Arhonditsis, G. B. Phytoplankton functional type modelling:

639

Running before we can walk? A critical evaluation of the current state of

640

knowledge. Ecol. Modell. 2016, 320, 29–43.

641 642

(68)

Rigosi, A.; Fleenor, W.; Rueda, F. State-of-the-art and recent progress in phytoplankton succession modelling. Environ. Rev. 2010, 18 (NA), 423–440.

643


29


Page 30 of 35

644

FIGURE CAPTIONS

645

Figure 1. Nash-Sutcliffe Efficiencies (NSE) of the genera models when tested against (a)

646

training, and (b) testing data subsets. Each boxplot contains n=25 values corresponding to

647

the 5-repeated 5-fold cross-validation performed for each Random Forests model; each

648

value is shown as a green point.

649 650

Figure 2. Permuted importance of explanatory variables. X-axes are in units of % change

651

in Mean Squared Error. Each boxplot includes n=25 values corresponding to the 5-

652

repeated 5-fold cross-validation performed for each model.

653 654

Figure 3. Scaled partial dependence plots of the cyanobacteria genera; solid lines equal

655

the median partial dependences of the n=25 models per genus, and corresponding shaded

656

areas are bound by the min and max partial dependences. Y-axes correspond to carbon

657

biomass, but the curves have been min-max normalized by genus to enable comparisons

658

among the genera, making the y-axes unitless. Partial dependence plots for each genus

659

(non-scaled) are included in the Supporting Information. The point clouds included at the

660

top of each panel reflect the densities of the explanatory covariates’ observed values.

661 662

Figure 4. Relative sensitivities of the genera to the explanatory variables; y-axes

663

correspond to relative sensitivity values (unitless).


30

TOC Page 31Environmental of 35 Science & Technology (a) TKN (mg L−1)

1.00

(b) TP (mg L−1)

Random Forests modeling of cyanobacteria genera

(c)

0.75 0.50 0.25

ACS Paragon Plus Environment 0. 16 0. 00

2 0. 1

8 0. 0

4 0. 0

3

2

1

0.00

sc

to ria

illa

is


O

st

ro cy

ic

M

gb ya

yn

s

op si

en a

ab a

ol

kt

an

Pl

is

to ria

illa

rm

pe

in dr os

yl

An

sc


O

st

ro cy

ic

M

s

by a

ng

ly

to

an k

(a) Training

C

Pl

op si

en a

ab a

rm

pe

in dr os

yl

1.0

C

An

Model efficiency (NSE)

Figure 1 Page 32 of 35

(b) Testing

0.5

0.0

−0.5

−1.0

Figure 2 35 Page 33 of


(a) Anabaena

(b) Cylindrospermopsis

Ktripton

Color

Color

Water temp.

Copepods

Ktripton

Avg. flow

TP

Water temp.

Avg. flow

Dissolved PO4

Dissolved NH4

TP

TKN

Dissolved NH4

Dissolved PO4

Rotifers

Rotifers

TKN

Copepods 0

10

20

30

0

(c) Planktolyngbya TKN

Color

Color

Avg. flow

Avg. flow

Ktripton

TP

TP

Water temp.

Dissolved PO4

Copepods

TKN

Dissolved NH4

Copepods

Rotifers

Rotifers

Dissolved PO4

Dissolved NH4

Ktripton 10

20

30

(e) Oscillatoria Color Water temp. Dissolved PO4 Avg. flow Ktripton TP Rotifers TKN Dissolved NH4


Copepods 0

10

20

20

30

20

30

(d) Microcystis

Water temp.

0

10

30

0

10

9

3

8

6

6

tp rotifers po4d

copepods tkn

rotifers tp tp

(a) TKN (mg L−1)

4

22 copepods tkn 3 Figure tkn

nh4d Environmental Sciencenh4d & Technology

(b) TP (mg L−1)

Page 34 of 35

n

(c) Dissolved NH4 (mg L−1) (d) Dissolved PO4 (mg L−1)

0. 16

0. 06 6 0 0. 12 0. 09 9 0

nh4d tripton

color

0. 3 03 0 0. 08

80 0. 00 0 0. 04

0. 10 63 0

tp po4d

tripton tripton

40

20 0. 05 2

90

0. 16 0. 00 0 1

nh4d

po4d

0. 12

00. .008 8

tp po4d

60

30

0

00. .004 4

80

3 63 0

40

20

2 2

0.75

0 1 1

90 90

1.00

c

Qavg

0.50

−1 Qavg −1 wtemp 3 −1 ) TP (mg ) (c) (Dissolved (mg(Ld(−1 )CPU Dissolved ) ) PO4 (mg ) (wtemp f)4Color (g)L Ktripton (m−1) (e)LAverage flow m s ) NH

4

0. 3 10

0

30 .0 0 9

0

10 .0 1 0 30 .0 5 2 0 20 .0 0 6

0. 16 0. 0 00 000

color wtemp

0.

50 0

0. 03 20 0 0 0 5 .08 0. 06 30 0 0. 0.1 1 2 40 0 0 .0 0 9

10 0

0.

.0 4

04. 16 00. 0. 00 00 0

312 0.1 0

tripton Qavg

0.

2

08

0. 05

11

color

0.

000. 04

0. 16 0. 00

0. 12 0. 3 0. 09 09

tripton Qavg

0. 0. 03 03 2 0.08 0. 0. 06 06

0.00

0. 0. 00 00 1 0. 04

0.25

(h) Water temperature (oC)

30 50 0

25 40 0

20 30 0

15 20 0

10 10 0

0

0 4

3

30

2

0 20

30

25

20

115 5

110 0

0 30

0 20

0 10

0

0 50

10 1 0

50

wtemp

wtemp

0.50

00

0

0

30

40

0

0 20

0 10

0

0 4

30

3

2 20 0

101 0

50

0.75

00

0

1.00

0.25

) (( ) ) Color((iCPU ) Copepods

( )

o −1 temperature C

0.75

tp

nh4d

po4d po4d

0.50

33

Anabaena Cylindrospermopsis Planktolyngbya Microcystis Oscillatoria 22

1

90

60

30

0

80

60

40

0

20

1.00

Qavg Qavg

00..0 099

00..0 066

0. 0 33

0. 00

30

9205

0. 10

ACS Paragon Plus Environment 6105 20

0. 05

30 10

0. 16

0. 00 color

2 80 3 0 4

0. 08

0. 12

601

tripton

20 00 30 0 420 0 50 400 0

10

0

0.00

0. 04

0.25

30

25

20

15

430

325

0

22

15

1

tkn tkn

−1

10

rotifers

(h) Water ) (m ()j) Rotifers (mgC L )

−1 g Ktripton mgC L

10

10 0.0 0 20 0 0 0. 0 30 3 0 0. 0 40 6 0 0. 0 50 9 0 0

0. 010

copepods

30

2

5 300.0 5 0

0. 0 08 1 0. 0 12 10 015 0. 16 200.020 00

.0

4

0.00

Page 35 of 35



Revealing Biotic and Abiotic Controls of Harmful Algal Blooms in a

Recommend Documents