Professional Documents
Culture Documents
Science of The Total Environment
Science of The Total Environment
H I G H L I G H T S G R A P H I C A L A B S T R A C T
a r t i c l e i n f o a b s t r a c t
Article history: Cyanobacteria blooms are a major environmental issue worldwide. Our understanding of the biophysical pro-
Received 2 August 2016 cesses driving cyanobacterial proliferation and the ability to develop predictive models that inform resource
Received in revised form 21 September 2016 managers and policy makers rely upon the accurate characterization of bloom dynamics. Models quantifying re-
Accepted 3 October 2016
lationships between bloom severity and environmental drivers are often calibrated to an individual set of bloom
Available online xxxx
observations, and few studies have assessed whether differences among observing platforms could lead to con-
Editor: D. Barcelo trasting results in terms of relevant bloom predictors and their estimated influence on bloom severity. The aim of
this study was to assess the degree of coherence of different monitoring methods in (1) capturing short- and
Keywords: long-term cyanobacteria bloom dynamics and (2) identifying environmental drivers associated with bloom var-
Cyanobacterial harmful algal blooms iability. Using western Lake Erie as a case study, we applied boosted regression tree (BRT) models to long-term
Water quality time series of cyanobacteria bloom estimates from multiple in-situ and remote sensing approaches to quantify
Long-term time series the relative influence of physico-chemical and meteorological drivers on bloom variability. Results of BRT models
Boosted regression trees showed remarkable consistency with known ecological requirements of cyanobacteria (e.g., nutrient loading,
Biophysical drivers
water temperature, and tributary discharge). However, discrepancies in inter-annual and intra-seasonal bloom
dynamics across monitoring approaches led to some inconsistencies in the relative importance, shape, and
⁎ Corresponding author.
E-mail address: [email protected] (I. Bertani).
1
Joint first authors.
2
Present address: Graduate Degree Program in Ecology, Natural Resource Ecology Lab, Colorado State University, Campus Delivery A245, Fort Collins, CO 80523-1499, USA.
http://dx.doi.org/10.1016/j.scitotenv.2016.10.023
0048-9697/© 2016 Elsevier B.V. All rights reserved.
I. Bertani et al. / Science of the Total Environment 575 (2017) 294–308 295
sign of the modeled relationships between select environmental drivers and bloom severity. This was especially
true for variables characterized by high short-term variability, such as wind forcing. These discrepancies might
have implications for our understanding of the role of different environmental drivers in regulating bloom dy-
namics, and subsequently for the development of models capable of informing management and decision mak-
ing. Our results highlight the need to develop methods to integrate multiple data sources to better characterize
bloom spatio-temporal variability and improve our ability to understand and predict cyanobacteria blooms.
© 2016 Elsevier B.V. All rights reserved.
the CHAB monitoring products included in this study (Section 2.2). We et al., 2013; Obenour et al., 2014; Stumpf et al., 2012), though seasonal
then provide an account of the environmental variables considered as phosphorus load alone may not entirely explain the substantial increase
candidate CHAB predictors (Section 2.3), and we describe the modeling in bloom size observed over the last decade (Obenour et al., 2014). Al-
analyses performed on each of the CHAB monitoring products (Section though several different cyanobacteria taxa can be found in western
2.4). Lake Erie, Microcystis has generally dominated the blooms in recent
years (Bridgeman et al., 2012; Brittain et al., 2000; Chaffin et al., 2013).
2.1. Study site
2.2. CHAB monitoring products in Lake Erie
The western basin of Lake Erie (Fig. 1) has a surface area of approx-
imately 3000 km2. It is relatively shallow (mean depth 7.4 m) and well- We compiled eight in-situ and remote sensing monitoring products,
mixed, with a water residence time of about 51 days (Michalak et al., seven of which provide multiple observations of CHAB size over the
2013). The basin is influenced mainly by the Detroit River to the north course of the CHAB season (Jun–Oct) during 2002–2013. For each sam-
and the Maumee River to the south-west. While the Maumee River pling date in each dataset we expressed bloom size as total metric tons
flow is substantially lower than that of the Detroit River (5% and 90% of chlorophyll-a (MT Chl-a) over the sampled area, following previously
of the total flow discharged annually into the western basin, respective- published protocols to calculate bloom size wherever possible (e.g.,
ly), the two rivers contribute a similar portion (45% each) of the annual Stumpf et al., 2012; Bridgeman et al., 2013) and/or applying appropriate
TP load delivered to the basin from both point and nonpoint sources conversions as described below and in the Supplementary materials.
(IJC, 2014; Maccoux et al., 2016). The even load proportion is due to Comparing absolute bloom sizes derived from different monitoring
the substantially higher TP concentrations typically found in the Mau- products is complicated by considerable differences in the location
mee River, which drains one of the most heavily agricultural watersheds and extent of the sampled areas, efficiencies of sampling methods and
in the Great Lakes region (Richards and Baker, 2002; Richards et al., analytical techniques, and accuracy of the applied conversion factors.
2008; Scavia et al., 2016). Compared to the Maumee, Detroit River nutri- As a result, we limit our analysis to assessing the degree of coherence
ent concentrations are generally too low to significantly contribute to among monitoring products in capturing intra-seasonal and inter-an-
CHAB formation (IJC, 2014; Scavia et al., 2016). Summer cyanobacteria nual variability in relative rather than absolute bloom size.
blooms typically originate and peak in the area directly affected by the
Maumee River plume and tend to move eastward later in the season 2.2.1. UT LEC in-situ sampling (2002–2013)
(Wynne and Stumpf, 2015). Recent studies have shown a positive rela- The University of Toledo Lake Erie Center (UT LEC) collected
tionship between peak summer CHAB size in the western basin and Microcystis samples at six sites from an area of approximately 340 km2
Maumee River spring phosphorus load (Bertani et al., 2016; Michalak in the vicinity of Maumee Bay (Fig. 1) in the years 2002–2013
Fig. 1. Map of the western basin of Lake Erie with in-situ sampling locations shown. EPA GLNPO: US Environmental Protection Agency's Great Lakes National Program Office; UT LEC:
University of Toledo Lake Erie Center; NOAA GLERL: National Oceanic and Atmospheric Administration's Great Lakes Environmental Research Laboratory.
I. Bertani et al. / Science of the Total Environment 575 (2017) 294–308 297
(Bridgeman et al., 2013). Surveys were carried out approximately bi- stations or those located outside of that area but which did not extend
weekly from May through October and samples were collected with beyond the region sampled by Bridgeman et al. (2013), so that the
vertical plankton tows over the entire water column. Microcystis size of the total sampled area remained roughly 300 km2 in all years
biovolume (mL/m2) was determined as reported in Bridgeman et al. (Fig. 1). In 2012–2013, four out of the nine master stations were sam-
(2013). In our analysis, total biovolume was converted to cell dry pled from approximately the same 300 km2 area. Chlorophyll-a concen-
weight using a relationship specifically derived for Microcystis in west- trations were measured from samples collected with a 1-m Niskin
ern Lake Erie (see Fig. 2 in Bridgeman et al., 2013). We then converted bottle lowered below the water surface. We multiplied chlorophyll
cell dry weight to chlorophyll-a using the average chlorophyll content values by the depth of the station (from NOAA bathymetry map;
of Microcystis cells measured in western Lake Erie (0.006125 g/g dry http://www.ngdc.noaa.gov/) and averaged them across stations for
wt; Chaffin et al., 2012). This value is consistent with measurements each sampling date, in accordance with the averaging approach used
of chlorophyll per unit biomass in Lake Erie in 2014 (T. Johengen, pers. for the UT LEC data (Bridgeman et al., 2013). These averages were
comm.) and with average Microcystis cell chlorophyll contents reported then multiplied by an area of 300 km2 to get total MT chlorophyll for
in the literature (Chen et al., 2011; Long et al., 2001; Wang et al., 2007). the sampled area.
To obtain an estimate of total bloom mass for the sampling area in each
sampling date, we averaged the values of MT Chl-a/m2 across the sam- 2.2.3. EPA GLNPO in-situ sampling (2002–2013)
pling sites, following the approach established in Bridgeman et al. Two cruises per year (April and August) have been carried out since
(2013), and multiplied the average by the extent of the sampled area 1983 by the EPA's Great Lakes National Program Office (GLNPO), sam-
(340 km2). pling six offshore stations in the western basin (Fig. 1, http://www3.
epa.gov/greatlakes/monitoring/). Due to the very limited seasonal sam-
2.2.2. NOAA-GLERL in-situ sampling (2008–2013) pling frequency, we did not include this product in our modeling analy-
The National Oceanic and Atmospheric Administration's Great Lakes ses. However, chlorophyll concentrations measured during the August
Environmental Research Laboratory (NOAA-GLERL, hereafter abbreviat- cruise were used to obtain a measure of total MT chlorophyll, which
ed as GLERL) has been monitoring water quality at several sites in the was compared to the inter-annual variability of the bloom estimates de-
western basin of Lake Erie since 2008. While surveys were carried out rived from other data sources. Specifically, we used chlorophyll values
monthly in 2008 at 13 sites across a broad portion of the basin, in measured from samples collected with a Niskin bottle just below the
2009–2011 nine master stations from an area of approximately 300 surface at the two EPA stations falling within the area sampled by the
km2 near the Maumee River inflow (Fig. 1) were sampled weekly to bi- other in-situ surveys (Fig. 1). We multiplied chlorophyll concentrations
weekly from June through September–October of each year. In August– by station depth, averaged across stations, and multiplied that average
September of 2011, when the bloom moved far beyond the Maumee by an area of 300 km2 to obtain estimates of total MT chlorophyll com-
Bay, samples were collected at stations closer to the central portion of parable to those provided by other in-situ monitoring products. EPA sys-
the basin, where the bloom was located. In these cases, we only consid- tematically measured chlorophyll at multiple depths across the water
ered sample sites located within the area covered by the GLERL master column. While we only used surface chlorophyll values to ensure
Fig. 2. Annual bloom severity, calculated as the peak 30-day moving average across each year, for the remote sensing (a) and in-situ (b) monitoring products. UT LEC is given a separate y-
axis for ease of interpretation. SeaWiFS estimates for the year 2009 are omitted due to sensor malfunctions that resulted in only four images available over the whole CHAB season. NCCOS:
NOAA's National Centers for Coastal Ocean Science; MTRI MO SSI: Michigan Tech Research Institute MODIS Surface Scum Index; MTRI MO TOT: Michigan Tech Research Institute MODIS
Total (chlorophyll-a + surface scum); MTRI SW SSI: Michigan Tech Research Institute SeaWiFS Surface Scum Index; MTRI SW TOT: Michigan Tech Research Institute SeaWiFS Total
(chlorophyll-a + surface scum); UT LEC: University of Toledo Lake Erie Center; GLERL: Great Lakes Environmental Research Laboratory; EPA: Environmental Protection Agency.
298 I. Bertani et al. / Science of the Total Environment 575 (2017) 294–308
consistency with GLERL bloom estimates, we compared EPA bloom size 2 m of the water column assuming a concentration equal to the 99th
estimates obtained using chlorophyll measurements taken below the percentile of the distribution of surface chlorophyll concentrations mea-
surface with estimates obtained by calculating the average chlorophyll sured in the western basin by NOAA-GLERL over the period 2008–2013
value across multiple depths throughout the whole water column. The (200 μg/L).
two sets of bloom size measurements align very closely, showing com- Hereafter we refer to the two time series of MTRI surface scum esti-
parable relative inter-annual variability (R2 = 0.95). mates as MO SSI (from MODIS images) and SW SSI (from SeaWiFS im-
ages). While we report surface scum estimates obtained from both
2.2.4. NOAA-NCCOS remote sensing (2002–2013) remote sensing products, only the MO SSI time series was included in
CHAB size estimates have been generated by NOAA's National Cen- our modeling analyses. This is because the SW SSI time series is restrict-
ters for Coastal Ocean Science (NOAA-NCCOS, hereafter abbreviated as ed to a time period when several years had low surface scum occurrence
NCCOS) using images from the MEdium-spectral Resolution Imaging (Fig. 2a), resulting in a relatively low number of data points exhibiting
Spectrometer (MERIS; 2002–2011) and from the Moderate Resolution scum values significantly larger than zero.
Imaging Spectroradiometer (MODIS; 2012–present) (Stumpf et al., In addition to the two surface scum time series, we developed two
2012; Wynne et al., 2013; Wynne et al., 2010; Wynne et al., 2008). separate time series - MO TOT and SW TOT - where total bloom size
Cyanobacteria biomass is quantified and expressed in terms of a was calculated as the sum of the MT chl-a calculated for the surface
“Cyanobacteria Index” (CI) that is positively related to cyanobacteria scum pixels and the MT chl-a calculated for all non-scum pixels in
abundance in the top portion of the water column (Wynne et al., which chl-a N 20 μg/L (Sayers et al., 2016).
2013; Wynne et al., 2010). The CI thus represents the blue-green por- In several MTRI MODIS and SeaWiFS images, cloud cover resulted in
tion of total surface chlorophyll. Individual cloud-free satellite image a varying number of pixels with missing data, possibly leading to CHAB
pixels are compiled into 10-day composites by dividing the CHAB sea- size underestimation. To correct for bias in images with less than one
son (Jun–Oct) into discrete 10-day periods and by taking the highest third of pixels missing, we rescaled the observed chlorophyll mass
CI value observed at each cloud-free pixel during each period (Stumpf using the ratio of the western basin area (~3000 km2) to the total visible
et al., 2012; Wynne and Stumpf, 2015). These composites are developed area in each image. However, if over one third of an image's pixels were
to reflect the total biomass of Microcystis during each 10-day period. missing, the image was considered unreliable, and removed from this
Bloom intensity is calculated by summing CI values across all pixels analysis. In 2009, only four SeaWiFS images are available throughout
within western Lake Erie for each 10-day composite, and annual bloom the CHAB season due to severe sensor malfunctions. We therefore omit-
size is calculated as the maximum 30-day (i.e., three 10-day compos- ted 2009 SeaWiFS data from analyses.
ites) moving average, according to methods established in Stumpf et
al. (2012). One CI corresponds to approximately 1.2 × 1020 2.3. Environmental data and predictor variable development
cyanobacteria cells (Stumpf et al., 2012), and is equivalent to approxi-
mately 4800 metric tons (MT) cyanobacteria dry weight (Obenour et We considered a suite of candidate physical, chemical, and meteoro-
al., 2014). We converted cell dry weight to chlorophyll-a using the aver- logical predictors that are hypothesized to influence CHAB severity and
age chlorophyll content of Microcystis cells measured in western Lake seasonal dynamics (Table 1). Anthropogenic nutrient enrichment is rec-
Erie (0.006125 g/g dry wt; Chaffin et al., 2012). ognized as one of the major drivers of cyanobacteria blooms in aquatic
systems (Brookes and Carey, 2011; Downing et al., 2001; Paerl et al.,
2.2.5. MTRI remote sensing (2002–2013) 2011), and spring TP loading from the Maumee River has been shown
CHABs in western Lake Erie are mapped by the Michigan Tech Re- to explain a large portion of the inter-annual variability in CHAB size
search Institute (MTRI) using MODIS and Sea-Viewing Wide Field-of- in western Lake Erie (Obenour et al., 2014; Stumpf et al., 2012).
View Sensor (SeaWiFS) satellite imagery by using the Color Producing We calculated monthly TP, dissolved reactive phosphorus (DRP),
Agent Algorithm (CPA-A) (Shuchman et al., 2013) to first retrieve and total nitrogen (TN) loads (MT) from the Maumee River using
total chlorophyll-a concentration and then classify pixels as CHABs river nutrient concentration data collected by Heidelberg University's
when chlorophyll exceeds 20 μg/L and water temperature is N 18 °C National Center for Water Quality Research (NCWQR, http://tinyurl.
(Sayers et al., 2016; Shuchman, 2014). The CPA-A is a multi-spectral com/jgq2jxp) and stream flow data from the United States Geological
bio-optical retrieval procedure which simultaneously estimates concen- Survey (USGS, http://www.usgs.gov/water). Missing nutrient concen-
trations of the three primary color producing agents (chlorophyll, col- trations were imputed as described in Obenour et al. (2014). Cumula-
ored dissolved organic matter, and suspended minerals). This tive loads from June, April to June, and February to June were included
approach allows for total chlorophyll estimation in complex waters, in the models as potential explanatory variables.
such as those found in Lake Erie (Shuchman et al., 2013), from SeaWiFS Tributary discharge can also influence CHAB development by affect-
and MODIS data. ing lake turbidity, local circulation patterns, stratification, and residence
Surface scum accumulations (Fig. S3) are mapped separately using a time (Mitrovic et al., 2003; Verspagen et al., 2006). Because blooms in
Surface Scum Index (SSI) algorithm developed by MTRI, which is an ad- the western basin of Lake Erie typically originate and peak in the area in-
aptation of the Normalized Difference Vegetation Index (NDVI) (Rouse fluenced by the Maumee River plume (Wynne and Stumpf, 2015), we
et al., 1973) and which is also applicable to all of the above mentioned included Maumee River flow (m3/s) as a candidate predictor.
sensors. While the CPA-A algorithm maps total chlorophyll-a concen- Lake circulation and mixing regime are affected by wind forcing
trations for pixels where no scum is present, the SSI is a presence/ab- (Beletsky et al., 2013; Michalak et al., 2013; Wynne et al., 2011). Hourly
sence indicator that identifies pixels with surface scum, but does not wind speed (m/s) and wind direction (degrees from true North) were
presently provide a measure of chlorophyll concentration within the acquired from NOAA's National Buoy Data Center (Buoy Station
scum. Chlorophyll concentrations within a Microcystis scum layer typi- 45005, http://www.ndbc.noaa.gov) (Michalak et al., 2013; Zhou et al.,
cally show high spatial variability (G. Fahnenstiel and T. Bridgeman, 2015). Wind speed and direction were missing from only 4.7% of the
pers. comm.), so that given the limited scum sampling data and relative- hours in the study period, and were imputed using a linear regression
ly coarse satellite resolution (~ 1 km), any assumption of the average with data from the nearby Buoy Station THL01. After imputation, only
scum chlorophyll content is a rough approximation. However, to com- 0.6% of the hourly data remained missing. Average daily wind stress
pare the relative variability in surface scum estimates derived from was calculated using the drag coefficient determined by Hsu (1974) fol-
the current version of the SSI algorithm with other available satellite lowing the method reported in Wynne et al. (2010). Wind velocity for
products, we approximated the average chlorophyll content of surface the northerly and westerly components and wind stress were included
scum to be equivalent to the total MT chlorophyll found in the top in the models.
I. Bertani et al. / Science of the Total Environment 575 (2017) 294–308 299
Table 1 preceding each sampling date, to capture potential effects of each vari-
Candidate environmental variables used in the BRT analysis. able at different time scales leading up to each bloom observation
Variable Description Mean (range) (Millie et al., 2014) while minimizing collinearity among time-lagged
WaterTemp_d Average lake surface 20.0 (8.7–27.5)
variables and preventing model over-fitting. For example, wind forcing
(°C) temperature for the western is expected to exert both a short-term (hourly/daily) impact on bloom
basin (from the Great Lakes dynamics by modulating mixing of the water column as well as a lon-
Surface Environmental Analysis ger-term effect by driving lake circulation, currents, and subsequently
(GLSEA) model), where d is the
water/nutrient residence time (Michalak et al., 2013; Zhou et al., 2015).
2, 8, or 30 days previous
Stress_d (Pa) Average wind stress (calculated 0.05 (0.00–0.86) Monthly and yearly temporal trend components were also included
following the method reported in the models to assess whether any consistent pattern in bloom sea-
in Wynne et al., 2010), where d sonality emerged across monitoring products (monthly trend) and
is the 2, 8, or 30 days previous whether our models would confirm previous findings reporting an in-
NortherlyWind_d Average northerly wind 0.36 (−16.57–10.87)
(m/s) component (from the north)
creased susceptibility of western Lake Erie to large blooms (yearly
recorded at National Data Buoy trend; Obenour et al., 2014).
Center's buoy 45005, where d is
the 2, 8, or 30 days previous 2.4. Boosted regression tree analysis
WesterlyWind_d Average westerly wind 0.29 (−13.55–10.26)
(m/s) component (from the west)
recorded at National Data Buoy Boosted regression trees (BRT) are a non-parametric machine learn-
Center's buoy 45005, where d is ing technique that has proven effective in capturing complex biophysi-
the 2, 8, or 30 days previous cal relationships that are difficult to discern when using classical
Irradiance_d Daily average of hourly Climate 224 (7–362) statistical approaches (Bhatt et al., 2013; Elith et al., 2008; Jouffray et
(W/m2) Forecast System (CFSR)
al., 2015; Leathwick et al., 2008). BRT combines regression tree algo-
downward shortwave radiation
flux at water surface in W/m2, rithms - models that recursively split response variables into homoge-
where d is the 2, 8, or 30 days neous groups defined by threshold values of explanatory variables
previous (Breiman et al., 1984) - with boosting, a method for combining models
Discharge_d Average daily Maumee river 112.7 (1.6–2217.2)
3
to enhance predictive performance (Friedman et al., 2000; Schapire,
(m /s) discharge from United States
Geological Survey, where d is the 2003). Classification and regression tree analysis has been widely ap-
2, 8, or 30 days previous plied to identify nonlinear, interactive relationships among variables
TP_June (MT) Sum of Maumee TP daily loads in 120.6 (5.0–496.8) in ecological datasets, and to quantify critical thresholds of environmen-
June tal drivers that trigger ecosystem responses (Cleveland et al., 2011;
TP_JuneApril Sum of Maumee TP daily loads in 673.9 (39.6–1516.5)
Cottenie, 2005; Fernandez et al., 2006; Hambright et al., 2015), includ-
(MT) April–June
TP_JuneFeb (MT) Sum of Maumee TP daily loads in 1343.5 (488.7–2622.3) ing cyanobacteria blooms (Chen and Mynett, 2004; Huber et al., 2012;
February–June Taranu et al., 2015; Wagner and Adrian, 2009). Despite the advantages
DRP_June (MT) Sum of Maumee DRP daily loads 32.3 (0.2–124.5) that regression tree algorithms offer in terms of ability to model discon-
in June
tinuities and interactions, incorporate different types of response and
DRP_JuneApril Sum of Maumee DRP daily loads 139.1 (7.1–265.5)
(MT) in April–June
explanatory variables, and effectively select relevant predictors
DRP_JuneFeb Sum of Maumee DRP daily loads 282.8 (94.6–539.7) (De'ath and Fabricius, 2000; De'ath, 2002), individual trees tend to
(MT) in February–June show relatively low robustness to small variations in the training
TN_June (MT) Sum of Maumee TN daily loads 3090.4 (72.7–10,115.8) dataset, weak predictive performance, and inadequacy in modeling
in June
smooth functions (Hastie et al., 2001; Murtaugh, 2009). Boosting algo-
TN_JuneApril Sum of Maumee TN daily loads 14,069.7 (1092.1–26,281.1)
(MT) in April–June rithms improve these limitations by fitting a large number of trees to
TN_JuneFeb (MT) Sum of Maumee TN daily loads 23,894.1 (6499.0–38,605.0) the data and then combining them to generate more accurate and ro-
in February–June bust predictions (De'ath, 2007; Elith et al., 2008; Friedman and
Year Yearly temporal trend (2002–2013)
Meulman, 2003).
Month Monthly temporal trend (5–10)
Compared with similar techniques that are based on fitting and av-
eraging results from many trees, such as bagged trees and random for-
Water temperature may promote cyanobacteria development both ests (Cutler et al., 2007; Prasad et al., 2006), boosting algorithms differ
directly by affecting algal growth rates, and indirectly by enhancing in that they fit trees sequentially to the data through a stage-wise pro-
water column stability (Jöhnk et al., 2008; Paerl and Huisman, 2008). cedure (Elith et al., 2008). Specifically, trees are fitted iteratively to the
Daily surface water temperatures (°C) were acquired from GLERL's training data, so that at each step a tree is added that minimizes the
Great Lakes Surface Environmental Analysis (GLSEA) mapping effort overall model prediction error. Each newly added tree is fitted to the re-
and averaged over the entire western basin. siduals of the previous collection of trees, thereby progressively focus-
Solar irradiance influences algal photosynthetic and growth rates, ing on unexplained variability in the response (Elith et al., 2008;
and modulates the outcome of competition among phytoplankton Hastie et al., 2001). At each step the fitted values are calculated as a lin-
taxa with different light requirements (Havens et al., 1998; Huisman ear combination of the trees fitted so far, so that the final BRT model can
et al., 1999; Litchman, 1998). Furthermore, light availability influences be viewed as an additive model where each tree represents a model
Microcystis cell buoyancy by controlling carbohydrate production and term (Friedman et al., 2000).
accumulation through photosynthesis (Ibelings et al., 1991; BRTs have been extensively applied as an exploratory tool to com-
Kromkamp and Mur, 1984; Wallace and Hamilton, 1999). Hourly short- pare the nature and relative importance of functional relationships be-
wave radiation data (W/m2) were acquired from NOAA's National Cen- tween variables across different sets of ecological observations
ter for Environmental Prediction's Climate Forecast System Re-Analysis (Buston and Elith, 2011; Descy et al., 2016; Segurado et al., 2016;
(CFSR, http://rda.ucar.edu/), and daily average values were calculated. Tisseuil et al., 2012; Walsh and Webb, 2016). We fit BRT models to
For each meteorological and hydrological predictor (wind velocity, each set of CHAB observations separately to explore relationships be-
wind stress, irradiance, temperature, and river flow), three time-lagged tween environmental drivers and bloom size and to assess the degree
variables were included as candidate predictors in the model. Specifical- of coherence in modeling results across monitoring products. A detailed
ly, we calculated average values for the 2 days, 8 days, and 30 days description of model parameterization and evaluation methods is
300 I. Bertani et al. / Science of the Total Environment 575 (2017) 294–308
provided in the Supplementary materials. Measures of the relative con- Partial dependence plots (Figs. 3 and 4) are used to display fitted
tribution of predictor variables to overall model fit used in machine functions in BRT models and to evaluate relationships between the re-
learning methods can be sensitive to high collinearity among predictors sponse variable and individual predictors. In this analysis, we focus on
and result in biased or unstable predictor ranking (Auret and Aldrich, comparing the shape and sign of functional relationships between envi-
2011; Nicodemus and Malley, 2009). Because we found high collinearity ronmental drivers and bloom size across CHAB monitoring products,
among different nutrient predictors (Pearson's correlation coefficient and on assessing the degree of coherence in identifying critical thresh-
ranging between 0.94–0.97 among TP and DRP variables, 0.75–0.95 olds of environmental drivers that are likely to determine an increase
among TP and TN, and 0.74–0.91 among DRP and TN), we compared or decrease in bloom size.
the influence of different nutrients on model results by fitting three
models to each set of CHAB observations, using TP, DRP, or TN load as
the nutrient predictor, respectively, while leaving all other predictors 3.2.1. Riverine inputs: nutrient loading and river discharge
unchanged. We report results here of models fit using TP, which is a Generally, TP loading variables were identified as positive correlates of
key target of watershed nutrient reduction strategies (United States bloom size in all models. When including TP as the nutrient load predictor,
and Canada, 2012), and discuss differences observed when replacing TP_June was selected as the most important variable in the MO TOT model.
TP predictors with DRP or TN. Results of models fit with DRP or TN are Models displayed a sharp increase in bloom size at TP_June varying be-
included in the Supplementary materials. tween 10 and 95 MT (Table 3; Figs. 3 and 4). Similar positive relationships
occurred with TP_JuneFebruary in the NCCOS and MO SSI models and with
TP_JuneApril in the GLERL and MO SSI models, although the thresholds
3. Results triggering an increase in bloom size differed among models (Table 3).
Replacing TP predictors with DRP or TN did not substantially change
3.1. Inter-annual variability in bloom size the amount of variability explained by the models, nor the functional re-
lationships identified for other predictor variables (Tables S1 and S2).
Time series of annual bloom size expressed as maximum 30-day DRP and TN loads were related positively to bloom size, like TP, al-
moving average observed in each year, similar to the approach used though the relative importance of different loading periods changed de-
by Stumpf et al. (2012), show that most monitoring products agree in pending on nutrient type (Tables S1 and S2).
estimating the largest blooms in 2011 and 2013 (Fig. 2). NCCOS, MO River discharge showed up as a significant predictor in three of the
SSI, and UT LEC show similar overall patterns across the time series, six models. Bloom size was negatively associated with increasing Dis-
with relatively large blooms occurring in 2003 and after 2007 (Fig. 2 charge_30 in the UT LEC and GLERL models, while SW TOT displayed a
and Table 2). While relatively good agreement is observed between u-shaped relationship with Discharge_2 (Table 3; Figs. 3 and 4).
MO TOT and NCCOS annual estimates in years after 2007, MO TOT ex-
hibits relatively larger bloom sizes than NCCOS in the first part of the
time series. Several other discrepancies emerge when comparing rela- 3.2.2. Meteorological variables: water temperature, wind stress, wind direc-
tive inter-annual variability in bloom size across monitoring products tion and irradiance
(Fig. 2). For example, while most products exhibit the largest pre- WaterTemp_30 was selected as the most important predictor in the
2008 bloom in 2003, the largest pre-2008 bloom occurred in 2004 ac- UT LEC and GLERL models and identified as a correlate of bloom size in
cording to MO TOT and SW TOT and in 2006 according to EPA. Similarly, all other models except MO TOT and SW TOT. All models showed a pos-
while MO TOT and NCCOS report blooms of somewhat comparable size itive threshold response of bloom size to WaterTemp_30, with the
in 2008–2010, UT LEC and GLERL show substantial relative inter-annual threshold ranging between 18 and 23 °C, depending on the model
variability in these years, with the 2009 (GLERL) and 2010 (UT LEC) (Table 3; Figs. 3 and 4).
blooms being markedly smaller than those reported in other years. Wind stress appeared in all but the MO SSI model with a relative in-
fluence ≥5% (Table 3). The model fitted to the NCCOS data displayed a
decrease in bloom size for Stress_8 N 0.08 Pa (Fig. 3). The UT LEC
3.2. Environmental drivers of bloom size: comparison across monitoring model exhibited an opposite pattern, with a decrease in bloom size at
products Stress_2 values N 0.02 Pa and an increase at values N0.04 Pa. Other
models showed a positive response of bloom size for stress values
Most BRT models explained between 33% and 50% of the temporal N0.03–0.05 Pa (Figs. 3 and 4).
variability in individual monitoring products, based on cross-validation Wind direction variables exhibited some discrepancies in the shape
performance (Table 3). An exception was the BRT model for scum only of the fitted functions and associated thresholds across models.
(MO SSI), which explained only 18% of response variability. All models NortherlyWind_30 values N−0.2 m/s were associated with an increase
identified seven or more environmental variables as important for in bloom size according to the GLERL model (Fig. 3). Positive values of
bloom prediction (i.e., with a relative influence ≥5%, Table 3). the northerly and westerly wind components were associated with
Table 2
Pearson's correlation coefficients between annual bloom size estimates (peak 30-day moving average) from different monitoring products. Coefficients with p-value b 0.05 are highlighted
in bold, while the number of paired data points used for each calculation is reported in parentheses. Correlation between GLERL and SW SSI/TOT was not calculated as only two paired data
points are available (2008 and 2010). Abbreviations as in Fig. 2.
Table 3
Explanatory variable effects and contributions to each final BRT model. For each variable, the shape of the relationship with the response is shown (+: positive, −: negative, ∩ or ᴜ: hump-
or u-shaped) alongside an approximate threshold(s) that defines that relationship. Relative contributions of each variable are given as a percent and they are scaled so that the sum is 100.
Only variables with a relative influence ≥5% are shown. For each monitoring product, the optimal number of trees used to fit each final BRT model and the associated mean cross-validated
R2 are listed along with their standard deviation. Abbreviations as in Table 1 and Fig. 2.
Fig. 3. Partial dependency plots for UT LEC, GLERL, and NCCOS. For each monitoring product, the six most influential variables are shown in decreasing order of relative importance (see
Table 3). Bloom size has been scaled to 1 at the maximum value observed by each monitoring product to facilitate comparisons of predictor effect size across models. Abbreviations as in
Table 1 and Fig. 2.
302 I. Bertani et al. / Science of the Total Environment 575 (2017) 294–308
Fig. 4. Partial dependency plots for MTRI MO SSI, MTRI MO TOT, and MTRI SW TOT. For each monitoring product, the six most influential variables are shown in decreasing order of relative
importance (see Table 3). Bloom size has been scaled to 1 at the maximum value observed by each monitoring product to facilitate comparisons of predictor effect size across models.
Abbreviations as in Table 1 and Fig. 2.
low bloom sizes in the MO SSI model (Fig. 4), while SW TOT showed an multiple interactions between wind direction variables and nutrient
u-shaped relationship with WesterlyWind_30 (Fig. 4). loading (Tables S3–S5).
The influence of irradiance was ≥5% in all models. According to the
MO TOT and SW TOT models, high irradiance values were mostly associ- 4. Discussion
ated with monotonic decreases in bloom size, with the exception of a
somewhat u-shaped relationship found between Irradiance_30 and SW 4.1. Coherent patterns across CHAB monitoring products
TOT estimates (Fig. 4). The UT LEC, GLERL, and NCCOS models exhibited
a hump-shaped relationship of bloom size with irradiance (Figs. 3 and 4). Most monitoring products exhibit an increase in bloom size in recent
years (Fig. 2), and most models show a positive yearly temporal trend,
independent of other predictors (Table 3). This is consistent with previ-
3.2.3. Temporal trend
ous findings suggesting a gradual increase in the lake's susceptibility to
The yearly temporal trend was selected as the most important pre-
CHAB formation over the past decade (Obenour et al., 2014), similar to
dictor for MO SSI and NCCOS (Table 3). A positive yearly temporal
what has been observed in numerous lakes worldwide (Paerl and Paul,
trend in bloom size was evident in all but the UT LEC model, although
2012; Taranu et al., 2015). Some of the models also exhibit highly
the trend begins in different years according to different monitoring
ranked interactions between the yearly trend and multiple meteorolog-
products (Table 3). The monthly trend had a relative contribution ≥5%
ical factors, such as a positive interaction between WaterTemp_30 and
in the UT LEC and NCCOS models, with a positive threshold in mid-
Year and between Irradiance_2 and Year for the NCCOS model (Fig.
July (Month = 7.5; Table 3).
S4). This further suggests that the system's response to external drivers
might have changed over time, although differences in the interacting
3.2.4. Two-way interactions factors across models make it difficult to formulate hypotheses on the
Models fitted to different monitoring products identified different potential underlying mechanisms.
sets of two-way interactions as most influential, with limited consisten- Most monitoring products also confirm the importance of nutrient
cy across products (Tables S3–S5). A positive interactive effect between loading as a driver of relative variability in CHAB size, although the
WaterTemp_30 and nutrient loading was ranked among the strongest high correlation among TP, DRP, and TN loads did not allow us to iden-
interactions by several models (Tables S3–S5). The NCCOS and MO SSI tify a preferred nutrient predictor. Bioassay studies during western Lake
models showed high relative strength for interactions involving the Erie blooms indicate that phosphorus is often the limiting nutrient for
yearly trend and meteorological predictors, while MO SSI also selected algal growth, although nitrogen limitation can occur in late summer
I. Bertani et al. / Science of the Total Environment 575 (2017) 294–308 303
(Chaffin et al., 2013), and recent work has highlighted the role of nitro- The negative relationship between bloom size and river discharge
gen in promoting the growth of toxic strains of Microcystis in Lake Erie exhibited by the two in-situ monitoring products (Table 3), whose sta-
(Gobler et al., 2016; Harke et al., 2016). The role of phosphorus loads tions are under the direct influence of the Maumee river plume (Fig.
in stimulating CHABs in western Lake Erie has been shown by multiple 1), is consistent with the notion of cyanobacteria growth being favored
studies (Bertani et al., 2016; Michalak et al., 2013; Obenour et al., 2014; by lower summer flushing rates and higher residence time (Elliott,
Stumpf et al., 2016; Stumpf et al., 2012; Verhamme et al., in review), and 2010; Huber et al., 2012; Michalak et al., 2013). The u-shaped relation-
TP and DRP load targets have recently been set for the Maumee River by ship found between river flow in the previous two days and SW TOT
the United States and Canada under the Great Lakes Water Quality might be due to occasional high bloom estimates exhibited by this prod-
Agreement (GLWQA) Amendment of 2012 (United States and Canada, uct during high-discharge events resulting in increased water column
2015). Specifically, a spring (Mar-Jul) TP loading target of 860 MT and turbidity and potential bloom size overestimation (see next section).
a corresponding DRP loading target of 186 MT were recommended to River discharge is used to calculate nutrient loads, so that one might ex-
significantly reduce CHAB formation based on results from a multiple pect a similar effect of these two predictors in the models. However,
modeling approach (Scavia et al., 2016). The thresholds identified for while we included nutrient loading in the model as cumulative values
the April–June TP load by the GLERL model (711 MT, ~ 237 MT/ over different spring periods, river discharge was included as average
month) and for the February–June TP load by the NCCOS model values over the days preceding each bloom estimate throughout the
(1135 MT, ~227 MT/month) are comparable to the TP load thresholds summer. As a result, while the load variables provide an estimate of
determined for the Mar–Jul period by most cyanobacteria models the overall amount of nutrients delivered in spring, the discharge vari-
used in the GLWQA effort (125–308 MT/month; Scavia et al., 2016). ables quantify the effect of intra-seasonal changes in river hydrology
The higher TP_JuneApril and TP_JuneFebruary thresholds extracted by on bloom size during the CHAB season. The effect of river hydrology is
the MO SSI model might be due to the fact that this product only records particularly relevant in the context of potential future climate change-
surface scum, a phenomenon that does not appear to be strictly propor- driven shifts in the timing, frequency, and intensity of extreme weather
tional to overall bloom size. In fact, when comparing time series of total events, such as storms and droughts (Michalak et al., 2013).
bloom size (e.g., MO TOT and NCCOS) with the scum-only MO SSI prod-
uct, a few years show high total bloom size and relatively large spring TP 4.2. Discrepancies across CHAB monitoring products
loads, but comparably low surface scum occurrence (e.g., 2003, 2008,
2010; Figs. 2a and S1). Relationships between bloom size and wind stress were less consis-
Despite the positive relationships generally found between bloom tent across the various monitoring products. NCCOS estimates exhibit a
size and nutrient loads, the fact that different load time windows were negative relationship with wind stress, with a decrease at
selected as influential predictors by different monitoring products sug- Stress_8 N 0.08 Pa (Table 3). In a previous study relating NCCOS bloom
gests that inconsistencies in inter-annual variability in bloom size might estimates and wind stress in western Lake Erie, Wynne et al. (2010) re-
lead to different conclusions on the most critical loading period (Table ported that wind stress N0.1 Pa resulted in a decrease in remotely
3). sensed bloom size due to mixing of cyanobacteria cells through the
Most monitoring products show a positive relationship between water column and subsequent reduction in satellite-detectable near-
bloom size and average water temperature in the preceding 30 days, surface concentrations. The relationship between bloom size and wind
whereas shorter-lagged temperature variables are never selected as in- stress might therefore be influenced by the inherent constraint that re-
fluential predictors (Table 3). The temperature threshold triggering an mote sensing only detects near-surface cyanobacteria accumulations,
increase in bloom size is higher for the scum model (~23 °C) than for especially in turbid, eutrophic freshwater (Kutser et al., 2008; Wynne
other models (~18–21 °C), suggesting that higher critical meteorologi- et al., 2010). The coupling of in-situ surveys with remote sensing ap-
cal thresholds might need to be crossed for surface scum to form. Espe- proaches may overcome this limitation because samples collected
cially high surface water temperatures develop during prolonged low- deeper in the water column may be less strongly affected by wind
mixing, calm summer conditions, which are known to enhance scum stress. In any case, the relationship between CHABs and wind is expect-
formation (Ibelings et al., 2003; Soranno, 1997). All models show max- ed to be complex, because while Microcystis is typically favored by high
imum bloom size when WaterTemp_30 values exceed 23–25 °C, which water column stability, and loses its competitive advantage over non-
is consistent with results of a Lake Erie-specific study based on the buoyant phytoplankton under well-mixed conditions (Ibelings et al.,
GLERL time series (Millie et al., 2014) and more generally with observa- 2003; Jöhnk et al., 2008; Visser et al., 1996), relatively short wind-in-
tions of Microcystis growth rates reaching maxima at temperatures ≥ duced mixing events preceding calm conditions might have a positive
25 °C (Butterwick et al., 2005; Reynolds, 2006). effect on overall cyanobacteria abundance by enhancing re-suspension
The consistent positive relationship between bloom size and of nutrients and overwintering cells (Chaffin et al., 2014; Preston et al.,
WaterTemp_30 indicates a marked seasonality in bloom development 1980; Verspagen et al., 2005).
(Wynne and Stumpf, 2015). Ho and Michalak (2015) compared bloom The positive relationships with wind stress found for other remote
seasonal timing in western Lake Erie using the NCCOS and UT LEC sensing products (i.e., MO TOT and SW TOT) contrast with NCCOS re-
2002–2011 time series, and they found that the average timing of sults. Differences in the estimated effect of wind forcing may be due to
bloom onset was around mid-July for both monitoring products, discrepancies in short-term intra-seasonal bloom dynamics across
which agrees with the positive threshold identified for the monthly monitoring products (Fig. S2). One illustrative example is represented
trend by the UT LEC and NCCOS models (Monthly = 7.5, i.e. mid-July, by the year 2004. In June and September 2004, MO TOT and SW TOT
Table 3) and with a recent in-depth synthesis of the NCCOS time series show short-lived peaks in bloom size that do not appear in the NCCOS
(Wynne and Stumpf, 2015). All other models identified the same sea- product (Fig. S2), resulting in the MO TOT and SW TOT 2004 annual
sonal timing associated with an increase in bloom size, although the bloom size being larger than that recorded in 2003 (Fig. 2a). Inherent
contribution to model fit was below 5% for most models, likely due to differences in the characteristics of the two remote sensing algorithms
the fact that WaterTemp_30 explains a large portion of bloom (CPA-A for MO and SW TOT products versus CI for NCCOS) likely con-
seasonality. tribute to these and other similar short-term discrepancies. While
The positive interaction between WaterTemp_30 and nutrient load- NCCOS data quantify only the cyanobacterial component of the phyto-
ing emerging from multiple models agrees with the known ecology of plankton assemblage, MO and SW TOT estimate total chlorophyll,
Microcystis blooms, which typically peak when high nutrient inputs which includes other algal groups. Although high chlorophyll levels
are combined with favorable physical conditions in late summer during the bloom season have generally been dominated by Microcystis
(Davis et al., 2009; Taranu et al., 2012). spp. in recent years (Bridgeman et al., 2013; Millie et al., 2014), the
304 I. Bertani et al. / Science of the Total Environment 575 (2017) 294–308
contribution of eukaryotic phytoplankton to total chlorophyll can be the characterization of CHAB seasonal dynamics when compared
substantial especially in the beginning and the end of the CHAB season with in-situ measurements.
(Bridgeman et al., 2012; Millie et al., 2009), thereby leading to higher Some inconsistencies in bloom temporal dynamics occur also across
chlorophyll values estimated by MO TOT and SW TOT relative to in-situ monitoring approaches (Fig. S2). Although UT LEC and GLERL
NCCOS. Indeed, when comparing the three satellite products in terms sites cover a similar portion of the lake in terms of areal extent, the
of bloom areal extent only (not shown), without quantifying chloro- somewhat different location of the sampling areas (Fig. 1) might lead
phyll content, some of the observed discrepancies are mitigated (see to differences in the observed CHAB seasonal dynamics due to bloom
Shuchman, 2014). Different sensors and algorithms may also respond horizontal movements and high spatial variability (Ho and Michalak,
differently when large sediment plumes affect water turbidity (Park et 2015). Differences in sampling techniques might also affect in-situ
al., 2010; Wynne et al., 2013), as is often the case in the area influenced bloom estimates. For instance, surface cyanobacteria accumulations
by the Maumee River during high discharge or high wind events. For ex- are better captured through vertical plankton tows (UT LEC) than by
ample, changes in the optical properties of water associated with high sampling below the surface with a Niskin bottle (GLERL). Accordingly,
sediment concentrations may result in occasional overestimation of the GLERL product shows less pronounced relative seasonal peaks com-
chlorophyll-a by the CPA-A algorithm. Accordingly, select high river pared to UT LEC during years with high scum occurrence (e.g., 2011 and
flow events (e.g., 2003, 2004, and 2007) or peaks in wind stress (e.g., 2013; Fig. S2). As mentioned above, the lack of taxonomic specificity of
2005, 2006) are associated with relatively higher bloom sizes detected total chlorophyll measurements might also lead to inconsistencies
by MO TOT and SW TOT compared to NCCOS (Fig. S2). This might also when compared with taxon-specific surveys in cases when Microcystis
explain the positive relationship found between SW TOT estimates is not the dominant taxon. Methodological differences across in-situ sam-
and Discharge_2 N 60 m3/s. MODIS and SeaWiFS bands may also incur pling approaches often depend on the original intended purpose of the
signal saturation under high surface scum (Reinart and Kutser, 2006; monitoring programs and associated research questions. For example,
Wynne et al., 2013), which led to the development of the SSI algorithm. GLERL surveys were initially designed to provide ground-truth informa-
However, chlorophyll concentration estimates for surface scum are pro- tion for satellite algorithm calibration. Similarly, the EPA sampling pro-
visional (see Materials and methods section), increasing the uncertainty gram in western Lake Erie is part of a broader limnological program
in overall bloom intensity estimates when large scums are present. On initiated to monitor water quality in the offshore waters of the Great
the other hand, the NCCOS time series is made of composite estimates Lakes. As a result, EPA stations are spread across the open waters of the
obtained by combining images from multiple days, so that temporal western basin, including regions that are typically not impacted by
variability is reduced, potentially affecting relationships with short- CHABs, and only two stations overlap with the GLERL and UT LEC sam-
lagged environmental drivers. The procedure adopted to generate the pling areas (Fig. 1). Although the original scope of some in-situ monitoring
NCCOS composites, based on summing the maximum biomass values programs might limit their application to characterize absolute bloom
recorded at each pixel over 10-day periods, may also be prone to occa- size, they still provide valuable information on inter-annual variability
sional bloom extent overestimation when the bloom is transported to a in relative bloom intensity to validate, integrate, and augment CHAB esti-
different region of the basin between subsequent satellite images. The mates from other long-term surveys. In this perspective, some of the in-
time of the day when satellites pass over the lake (~ 10 am for MERIS situ and remote sensing datasets considered here have been integrated
and ~2 pm for MODIS AQUA and SeaWiFS) might also affect detected in recent CHAB modeling approaches to support model calibration or val-
near-surface bloom size, due to relatively rapid daily vertical migration idation (Verhamme et al. in review; Obenour et al., 2014).
patterns typically observed for buoyant cyanobacteria (Ibelings et al., In general, method-specific limitations in characterizing bloom var-
1991; Kromkamp and Mur, 1984; Wynne et al., 2013). On a calm day iability need to be taken into careful consideration when using CHAB
in western Lake Erie cyanobacteria tend to form surface scum accumu- monitoring data to infer relationships with environmental drivers. Dis-
lations in mid-day (~11 am) and start to subside in the early afternoon crepancies in short-term bloom temporal dynamics across monitoring
(~2 pm), until no scum is typically present by 6 pm (T. Bridgeman, pers. products may affect the modeled relationships with variables character-
comm.). ized by marked fluctuations on hourly to daily time scales, such as wind,
Discrepancies in intra-seasonal bloom dynamics also occur when leading to contrasting conclusions on their overall effect on bloom size
comparing remote sensing vs. in-situ monitoring products (Fig. S2). In (Table 3). The development of quantitative approaches to integrate
the case of western Lake Erie, inconsistencies in bloom temporal dy- multiple types of bloom observations in CHAB modeling efforts
namics between in-situ and satellite estimates have been partly attrib- may help overcome the specific limitations of individual monitor-
uted to the proximity of in-situ sampling areas to the region of bloom ing methods and improve bloom characterization and modeling
initiation (Ho and Michalak, 2015). More generally, the high spatial capabilities.
patchiness that typically characterizes cyanobacteria blooms is expect- The negative effect of high irradiance levels estimated by some
ed to affect CHAB severity estimates obtained from in-situ vs. satellite models may suggest that photo-inhibition might play a role in influenc-
approaches differently. With respect to spatial coverage, satellite prod- ing bloom seasonal dynamics (Ibelings and Maberly, 1998). However,
ucts represent an obvious enhancement over in-situ sampling, which while photo-inhibition of surface blooms has been occasionally ob-
is typically limited to a few point stations over a restricted portion of served in Lake Erie (Chaffin et al., 2012), Microcystis is known to possess
the lake. For example, extremely high chlorophyll concentrations mea- several mechanisms to adapt to high levels of irradiance (Paerl and
sured at one GLERL site near the mouth of the Maumee in 2008 exert a Otten, 2013b; Paerl et al., 1985; Sommaruga et al., 2009), and the rela-
disproportionate influence on overall bloom size estimates. Although tively high turbidities often observed in western Lake Erie provide fur-
more sophisticated spatial integration methods than that used here ther protection against severe photodamage (Chaffin et al., 2012). The
could improve in-situ estimates, the limitations associated with treating different irradiance thresholds identified by different models (Table 3)
measurements from a limited number of stations as representative of a make it difficult to formulate hypotheses on the irradiance levels that
300 km2 area remain salient. The limited spatial coverage of in-situ sur- might be detrimental for blooms in western Lake Erie, and more re-
veys may also result in CHAB size underestimation when the bloom ex- search is needed on the possible influence of high irradiance on bloom
tends beyond or is transported away from the sampling area, such as in dynamics, especially in view of its potential to promote toxic strain
2011 and 2013 in Lake Erie (Ho and Michalak, 2015). On the other dominance in Microcystis populations (Paerl and Otten, 2013b).
hand, some satellite sensors tend to exhibit signal saturation The scum-only (MO SSI) model explains a lower portion of data var-
when large patches of extremely high cyanobacteria biomass or iability when compared to other remote sensing products (Table 3).
sediment are present (Kutser, 2004; Reinart and Kutser, 2006; Surface scum formation is often a rapid and transient phenomenon,
Wynne et al., 2013), potentially contributing to discrepancies in whose accurate characterization requires an adequately high sampling
I. Bertani et al. / Science of the Total Environment 575 (2017) 294–308 305
frequency. Remote sensing theoretically offers higher temporal resolu- track and forecast CHABs in lakes worldwide, making the implica-
tion than most long-term in-situ monitoring programs. However, limi- tions of our results relevant beyond the study lake.
tations in image availability or quality due to meteorological Our analysis underscores the importance and challenges associated
conditions often cause multi-day gaps that hinder the characterization with comparing multiple types of bloom measurements. Scientists and
of ephemeral surface bloom events. In addition, surface scum occur- resource managers are still faced with significant uncertainties in char-
rence has increased only in the most recent years of the available time acterizing bloom variability, especially at fine temporal scales. Future ef-
period (Fig. 2a), resulting in a limited number of data points on which forts should focus on implementing rigorous methods to generate
to train the models. Interestingly, the scum-only model shows the larg- bloom size estimates that systematically integrate information from
est contribution of wind direction, including several interactive effects multiple data sources. The development of methods that assimilate
between wind direction and various other predictors (Tables 3 and data from multiple CHAB monitoring methods has the potential to
S3–S5). The horizontal distribution and areal coverage of surface scum allow for more accurate bloom detection by leveraging the advantages
might be affected by wind direction in a more direct way than subsur- associated with each monitoring approach while overcoming the re-
face blooms (Wynne et al., 2011). However, given the limitations spective limitations. Enhancing our ability to quantify bloom distribu-
outlined above, and considering that the scum concentration estimates tion and intensity is a key step for both informing resource managers
are provisional and highly uncertain, extra caution should be taken in of bloom spatio-temporal impacts and for improving models aimed at
interpreting results describing scum dynamics. advancing our predictive understanding of harmful algal blooms.
In relation to this, it is important to note that there are uncertainties
associated with some of the assumptions required to estimate total
Acknowledgements
bloom size from the monitoring data (see Materials and methods sec-
tion), and that these uncertainties are not explicitly accounted for in
This work was funded by the University of Michigan Water Center
the modeling analysis. While we followed established workflows to cal-
and by the National Science Foundation (NSF) under grant 1313897.
culate bloom size (e.g., Stumpf et al., 2012; Bridgeman et al., 2013)
We thank Richard Stumpf and Timothy Wynne for providing the
wherever possible and focused our analyses on relative bloom size rath-
NCCOS satellite images. Support for UT LEC data collection was provided
er than absolute bloom size, enhancements in bloom detection technol-
over various periods by: The Ohio Sea Grant College Program, Project R/
ogies under development in western Lake Erie may help quantify and
ER-72, under grant NA16RG2252 from the National Sea Grant College
reduce the uncertainty associated with some of those assumptions in
Program, the Ohio Lake Erie Commission, Lake Erie Protection Fund Pro-
the future. For example, the MTRI SSI algorithm is currently being en-
jects #03-19, #0-08 and #SG406-11, USEPA project #GL2009-1-2, the
hanced to provide quantitative estimates of scum concentrations. A
University of Cincinnati, the USDA Natural Resources Conservation Ser-
monitoring product capable of quantifying surface scum accumulations
vice (MVRC&D), and the NSF under grant 1039043.
is an especially important tool for water resource managers because
surface scums directly impede recreational uses due to distasteful visual
Appendix A. Supplementary data
and odor cues (Fig. S3). A long-term dataset quantifying surface scum
areal coverage and concentration also has the potential to improve
Supplementary data to this article can be found online at http://dx.
our ability to understand and model environmental conditions promot-
doi.org/10.1016/j.scitotenv.2016.10.023.
ing scum development. The increasing deployment of advanced in-situ
and remote sensing bloom monitoring technologies, such as in-situ
water quality sensors and aerial hyperspectral sensors, will also offer References
an unprecedented level of spatio-temporal resolution to augment
Anderson, D.M., Glibert, P.M., Burkholder, J.M., 2002. Harmful algal blooms and eutrophi-
bloom estimates from more traditional monitoring approaches cation: nutrient sources, composition, and consequences. Estuaries 25, 704–726.
(Bullerjahn et al., 2016). Auret, L., Aldrich, C., 2011. Empirical comparison of tree ensemble variable importance
measures. Chemom. Intell. Lab. Syst. 105, 157–170. http://dx.doi.org/10.1016/j.
chemolab.2010.12.004.
Baker, D.B., Confesor, R., Ewing, D.E., Johnson, L.T., Kramer, J.W., Merryfield, B.J., 2014.
5. Conclusions Phosphorus loading to Lake Erie from the Maumee, Sandusky and Cuyahoga rivers:
the importance of bioavailability. J. Great Lakes Res. 40, 502–517. http://dx.doi.org/
10.1016/j.jglr.2014.05.001.
Different CHAB monitoring products exhibit a high degree of coher-
Beletsky, D., Hawley, N., Rao, Y.R., 2013. Modeling summer circulation and thermal struc-
ence in indicating that high nutrient inputs, high water temperature, ture of Lake Erie. J. Geophys. Res. Ocean. 118, 6238–6252. http://dx.doi.org/10.1002/
and low flushing rates are most conducive to bloom development. The 2013JC008854.
associated thresholds extracted by the BRT models are consistent with Bertani, I., Obenour, D., Steger, C., Stow, C., Gronewold, A., Scavia, D., 2016. Probabilistical-
ly assessing the role of nutrient loading in harmful algal bloom formation in western
Microcystis ecology and with recent phosphorus load recommendations Lake Erie. J. Great Lakes Res. http://dx.doi.org/10.1016/j.jglr.2016.04.002.
derived from a multiple models approach. Bhatt, S., Gething, P.W., Brady, O.J., Messina, J.P., Farlow, A.W., Moyes, C.L., Drake, J.M.,
However, the influence of environmental drivers characterized by Brownstein, J.S., Hoen, A.G., Sankoh, O., Myers, M.F., George, D.B., Jaenisch, T., Wint,
G.R.W., Simmons, C.P., Scott, T.W., Farrar, J.J., Hay, S.I., 2013. The global distribution
high short-term variability, such as wind forcing, appears less clear, and burden of dengue. Nature 496, 504–507. http://dx.doi.org/10.1038/nature12060.
and conflicting results emerge when fitting the same model to different Breiman, L., Friedman, J.H., Olshen, R.A., Stone, C.G., 1984. Classification and Regression
monitoring products. This is likely due to the combination of high Trees. Chapman and Hall/CRC.
Bridgeman, T.B., Chaffin, J.D., Kane, D.D., Conroy, J.D., Panek, S.E., Armenio, P.M., 2012.
spatio-temporal variability of cyanobacteria blooms and method- From river to lake: phosphorus partitioning and algal community compositional
specific limitations in capturing such variability at the appropriate changes in Western Lake Erie. J. Great Lakes Res. 38, 90–97. http://dx.doi.org/10.
spatial and temporal scales. Such discrepancies have implications 1016/j.jglr.2011.09.010.
Bridgeman, T.B., Chaffin, J.D., Filbrun, J.E., 2013. A novel method for tracking western Lake
for understanding functional relationships between environmental
Erie Microcystis blooms, 2002–2011. J. Great Lakes Res. 39, 83–89. http://dx.doi.org/
factors and CHAB formation, and ultimately for developing models 10.1016/j.jglr.2012.11.004.
capable of informing resource managers with accurate intra-season- Brittain, S.M., Wang, J., Babcock-Jackson, L., Carmichael, W.W., Rinehart, K.L., Culver, D.A.,
2000. Isolation and characterization of microcystins, cyclic heptapeptide hepatotoxins
al bloom predictions. While previous studies have provided qualita-
from a Lake Erie strain of Microcystis aeruginosa. J. Great Lakes Res. 26, 241–249.
tive assessments of the degree of inter-comparability of different Brookes, J.D., Carey, C.C., 2011. Resilience to blooms. Science 334 (80-.), 46–47. http://dx.
CHAB monitoring approaches in Lake Erie, to our knowledge this is doi.org/10.1126/science.1207349.
the first study applying a quantitative modeling approach to synthe- Bullerjahn, G.S., Mckay, R.M., Davis, T.W., Baker, D.B., Boyer, G.L., Anglada, L.V.D., Doucette,
G.J., Ho, J.C., Irwin, E.G., Kling, C.L., Kudela, R.M., Kurmayer, R., Michalak, A.M., Ortiz,
size and compare multiple time series of CHAB estimates. The mon- J.D., Otten, T.G., Paerl, H.W., Qin, B., Sohngen, B.L., Stumpf, R.P., Visser, P.M.,
itoring approaches considered in this study are routinely used to Wilhelm, S.W., 2016. Global solutions to regional problems: collecting global
306 I. Bertani et al. / Science of the Total Environment 575 (2017) 294–308
expertise to address the problem of harmful cyanobacterial blooms. A Lake Erie case Gobler, C.J., Burkholder, J.M., Davis, T.W., Harke, M.J., Johengen, T., Stow, C.A., Van De
study. Harmful Algae 54, 223–238. http://dx.doi.org/10.1016/j.hal.2016.01.003. Waal, D.B., 2016. The dual role of nitrogen supply in controlling the growth and tox-
Buston, P.M., Elith, J., 2011. Determinants of reproductive success in dominant pairs of icity of cyanobacterial blooms. Harmful Algae 54, 87–97. http://dx.doi.org/10.1016/j.
clownfish: a boosted regression tree analysis. J. Anim. Ecol. 80, 528–538. http://dx. hal.2016.01.010.
doi.org/10.1111/j.1365-2656.2011.01803.x. Hambright, D.K., Beyer, J.E., Easton, J.D., Zamor, R.M., Easton, A.C., Hallidayschult, T.C.,
Butterwick, C., Heaney, S.I., Talling, J.F., 2005. Diversity in the influence of temperature on 2015. The niche of an invasive marine microbe in a subtropical freshwater impound-
the growth rates of freshwater algae, and its ecological relevance. Freshw. Biol. 50, ment. ISME J. 9, 256–264. http://dx.doi.org/10.1038/ismej.2014.103.
291–300. http://dx.doi.org/10.1111/j.1365-2427.2004.01317.x. Harke, M.J., Davis, T.W., Watson, S.B., Gobler, C.J., 2016. Nutrient-controlled niche differen-
Carmichael, W.W., Boyer, G.L., 2016. Health impacts from cyanobacteria harmful algae tiation of western Lake Erie cyanobacterial populations revealed via
blooms: implications for the North American Great Lakes. Harmful Algae 54, metatranscriptomic surveys. Environ. Sci. Technol. 50, 604–615. http://dx.doi.org/
194–212. http://dx.doi.org/10.1016/j.hal.2016.02.002. 10.1021/acs.est.5b03931.
Chaffin, J.D., Bridgeman, T.B., Heckathorn, S.A., Krause, A.E., 2012. Role of suspended sed- Hastie, T., Tibshirani, R., Friedman, J.H., 2001. The Elements of Statistical Learning: Data
iments and mixing in reducing photoinhibition in the bloom-forming cyanobacteri- Mining, Inference, and Prediction. Springer Verlag, New York.
um Microcystis. J. Water Resour. Prot. 4, 1029–1041. http://dx.doi.org/10.4236/ Havens, K.E., 2008. Cyanobacteria blooms: effects on aquatic ecosystems. In: Hudnell, H.K.
jwarp.2012.412119. (Ed.), Cyanobacterial Harmful Algal Blooms - State of the Science and Research
Chaffin, J.D., Bridgeman, T.B., Bade, D.L., 2013. Nitrogen constrains the growth of late sum- Needs. Springer.
mer cyanobacterial blooms in Lake Erie. Adv. Microbiol. 16–26. Havens, K.E., Phlips, E.J., Cichra, M.F., Li, B.L., 1998. Light availability as a possible regulator
Chaffin, J., Sigler, V., Bridgeman, T., 2014. Connecting the blooms: tracking and establish- of cyanobacteria species composition in a shallow subtropical lake. Freshw. Biol. 39,
ing the origin of the record-breaking Lake Erie Microcystis bloom of 2011 using DGGE. 547–556. http://dx.doi.org/10.1046/j.1365-2427.1998.00308.x.
Aquat. Microb. Ecol. 73, 29–39. http://dx.doi.org/10.3354/ame01708. Ho, J.C., Michalak, A.M., 2015. Challenges in tracking harmful algal blooms: a synthesis of
Chen, Q., Mynett, A.E., 2004. Predicting Phaeocystis globosa bloom in Dutch coastal waters evidence from Lake Erie. J. Great Lakes Res. http://dx.doi.org/10.1016/j.jglr.2015.01.
by decision trees and nonlinear piecewise regression. Ecol. Model. 176, 277–290. 001.
http://dx.doi.org/10.1016/j.ecolmodel.2003.10.031. Hsu, S.A., 1974. Experimental results of the drag-coefficient estimation for air–coast inter-
Chen, M., Li, J., Dai, X., Sun, Y., Chen, F., 2011. Effect of phosphorus and temperature on faces. Bound.-Layer Meteorol. 6, 505–507.
chlorophyll a contents and cell sizes of Scenedesmus obliquus and Microcystis Huber, V., Wagner, C., Gerten, D., Adrian, R., 2012. To bloom or not to bloom: contrasting
aeruginosa. Limnology 12, 187–192. http://dx.doi.org/10.1007/s10201-010-0336-y. responses of cyanobacteria to recent heat waves explained by critical thresholds of
Chorus, I., Bartram, J., 1999. Toxic Cyanobacteria in Water: a Guide to Their Public Health abiotic drivers. Oecologia 169, 245–256. http://dx.doi.org/10.1007/s00442-011-
Consequences, Monitoring and Management. E & FN Spon, London. 2186-7.
Cleveland, C.C., Townsend, A.R., Taylor, P., Alvarez-Clare, S., Bustamante, M.M.C., Chuyong, Huisman, J., Jonker, R.R., Zonneveld, C., Weissing, F.J., 1999. Competition for light between
G., Dobrowski, S.Z., Grierson, P., Harms, K.E., Houlton, B.Z., Marklein, A., Parton, W., phytoplankton species: experimental tests of mechanistic theory. Ecology 80,
Porder, S., Reed, S.C., Sierra, C.a., Silver, W.L., Tanner, E.V.J., Wieder, W.R., 2011. Rela- 211–222. http://dx.doi.org/10.1890/0012-9658(1999)080[0211:CFLBPS]2.0.CO;2.
tionships among net primary productivity, nutrients and climate in tropical rain for- Ibelings, B.W., Maberly, S.C., 1998. Photoinhibition and the availability of inorganic carbon
est: a pan-tropical analysis. Ecol. Lett. 14, 939–947. http://dx.doi.org/10.1111/j.1461- restrict photosynthesis by surface blooms of cyanobacteria. Limnol. Oceanogr. 43,
0248.2011.01658.x. 408–419. http://dx.doi.org/10.4319/lo.1998.43.3.0408.
Codd, G.A., Lindsay, J., Young, F.M., Morrison, L.F., Metcalf, J.S., 2005a. Harmful Ibelings, B.W., Mur, L.R., Walsby, a.E., 1991. Diurnal changes in buoyancy and vertical-dis-
cyanobacteria - from mass mortalities to management measures. In: Huisman, J., tribution in populations of Microcystis in 2 Shallow Lakes. J. Plankton Res. 13,
Matthijs, H., Visser, P.M. (Eds.), Harmful Cyanobacteria. Springer, Dordrecht. 419–436.
Codd, G.A., Morrison, L.F., Metcalf, J.S., 2005b. Cyanobacterial toxins: risk management for Ibelings, B.W., Vonk, M., Los, H.F.J., Van Der Molen, D.T., Mooij, W.M., 2003. Fuzzy model-
health protection. Toxicol. Appl. Pharmacol. 203, 264–272. http://dx.doi.org/10.1016/ ing of cyanobacterial surface waterblooms: validation with NOAA-AVHRR satellite
j.taap.2004.02.016. images. Ecol. Appl. 13, 1456–1472. http://dx.doi.org/10.1890/01-5345.
Cottenie, K., 2005. Integrating environmental and spatial processes in ecological commu- IJC, 2014. A Balanced Diet for Lake Erie: Reducing Phosphorus Loadings and Harmful Algal
nity dynamics. Ecol. Lett. 8, 1175–1182. http://dx.doi.org/10.1111/j.1461-0248.2005. Blooms. Report of the Lake Erie Ecosystem Priority.
00820.x. Jöhnk, K.D., Huisman, J., Sharples, J., Sommeijer, B., Visser, P.M., Stroom, J.M., 2008. Sum-
Cutler, D.R., Edwards, T.C., Beard, K.H., Cutler, A., Hess, K.T., Gibson, J., Lawler, J.J., 2007. mer heatwaves promote blooms of harmful cyanobacteria. Glob. Chang. Biol. 14,
Random forests for classification in ecology. Ecology 88, 2783–2792. http://dx.doi. 495–512. http://dx.doi.org/10.1111/j.1365-2486.2007.01510.x.
org/10.1890/07-0539.1. Jouffray, J.-B., Nyström, M., Norström, A., Williams, I., Wedding, L., Kittinger, J., Williams,
Dale, B., Murphy, M., 2014. A retrospective appraisal of the importance of high-resolution G., 2015. Identifying multiple coral reef regimes and their drivers across the Hawaiian
sampling for harmful algal blooms: lessons from long-term phytoplankton monitor- archipelago. Philos. Trans. R. Soc. B 370, 20130268. http://dx.doi.org/10.1098/rstb.
ing at Sherkin Island, S.W. Ireland. Harmful Algae 40, 23–33. http://dx.doi.org/10. 2013.0268.
1016/j.hal.2014.09.007. Kosten, S., Huszar, V.L.M., Bécares, E., Costa, L.S., Donk, E., Hansson, L.-A., Jeppesen, E.,
Davis, T.W., Berry, D.L., Boyer, G.L., Gobler, C.J., 2009. The effects of temperature and nutri- Kruk, C., Lacerot, G., Mazzeo, N., Meester, L., Moss, B., Lürling, M., Nõges, T., Romo,
ents on the growth and dynamics of toxic and non-toxic strains of Microcystis during S., Scheffer, M., 2012. Warmer climates boost cyanobacterial dominance in shallow
cyanobacteria blooms. Harmful Algae 8, 715–725. http://dx.doi.org/10.1016/j.hal. lakes. Glob. Chang. Biol. 18, 118–126. http://dx.doi.org/10.1111/j.1365-2486.2011.
2009.02.004. 02488.x.
De'ath, G., 2002. Multivariate regression trees: a new technique for modeling species-en- Kromkamp, J.C., Mur, L.R., 1984. Buoyant density changes in the cyanobacterium
vironment relationships. Ecology 83, 1105–1117. Microcystis aeruginosa due to changes in the cellular carbohydrate content.
De'ath, G., 2007. Boosted trees for ecological modeling and prediction. Ecology 88, FEMS Microbiol. Lett. 25, 105–109. http://dx.doi.org/10.1016/0378-
243–251. http://dx.doi.org/10.1890/0012-9658(2007)88[243:BTFEMA]2.0.CO;2. 1097(84)90055-7.
De'ath, G., Fabricius, K.E., 2000. Classification and regression trees: a powerful yet simple Kutser, T., 2004. Quantitative detection of chlorophyll in cyanobacterial blooms by satel-
technique for ecological data analysis. Ecology 81, 3178–3192. lite remote sensing. Limnol. Oceanogr. 49, 2179–2189. http://dx.doi.org/10.4319/lo.
DePinto, J.V., Young, T.C., McIlroy, L.M., 1986. Great Lakes water quality improvement. En- 2004.49.6.2179.
viron. Sci. Technol. 20, 752–759. http://dx.doi.org/10.1021/es00150a001. Kutser, T., Metsamaa, L., Vahtmäe, E., Strömbeck, N., 2006. Suitability of MODIS 250 m res-
Descy, J.-P., Leprieur, F., Pirlot, S., Leporcq, B., Van Wichelen, J., Peretyatko, A., Teissier, S., olution band data for quantitative mapping of cyanobacterial blooms. Proc. Est. Acad.
Codd, G.A., Triest, L., Vyverman, W., Wilmotte, A., 2016. Identifying the factors deter- Sci. Biol. Ecol. 55, 318–328.
mining blooms of cyanobacteria in a set of shallow lakes. Eco. Inform. 34, 129–138. Kutser, T., Metsamaa, L., Dekker, A.G., 2008. Influence of the vertical distribution of
http://dx.doi.org/10.1016/j.ecoinf.2016.05.003. cyanobacteria in the water column on the remote sensing signal. Estuar. Coast.
Dodds, W.K., Bouska, W.W., Eitzmann, J.L., Pilger, T.J., Pitts, K.L., Riley, A.J., Schloesser, J.T., Shelf Sci. 78, 649–654. http://dx.doi.org/10.1016/j.ecss.2008.02.024.
Thornbrugh, D.J., 2009. Eutrophication of U.S. freshwaters: analysis of potential eco- Leathwick, J.R., Elith, J., Chadderton, W.L., Rowe, D., Hastie, T., 2008. Dispersal, disturbance
nomic damages. Environ. Sci. Technol. 43, 12–19. and the contrasting biogeographies of New Zealand's diadromous and non-diadro-
Dolan, D.M., 1993. Point source loadings of phosphorus to Lake Erie. J. Great Lakes Res. mous fish species. J. Biogeogr. 35, 1481–1497. http://dx.doi.org/10.1111/j.1365-
1986–1990. 2699.2008.01887.x.
Downing, J.A., Watson, S.B., McCauley, E., 2001. Predicting cyanobacteria dominance in Litchman, E., 1998. Population and community responses of phytoplankton to fluctuating
lakes. Can. J. Fish. Aquat. Sci. 58, 1905–1908. http://dx.doi.org/10.1139/f01-143. light. Oecologia 117, 247–257. http://dx.doi.org/10.1007/s004420050655.
Elith, J., Leathwick, J.R., Hastie, T., 2008. A working guide to boosted regression trees. Long, B.M., Jones, G.J., Orr, P.T., 2001. Cellular microcystin content in N-limited Microcystis
J. Anim. Ecol. 77, 802–813. http://dx.doi.org/10.1111/j.1365-2656.2008.01390.x. aeruginosa can be predicted from growth rate. Appl. Environ. Microbiol. 67, 278–283.
Elliott, J.A., 2010. The seasonal sensitivity of cyanobacteria and other phytoplankton to http://dx.doi.org/10.1128/AEM.67.1.278.
changes in flushing rate and water temperature. Glob. Chang. Biol. 16, 864–876. Maccoux, M., Dove, A., Backus, S., Dolan, D., 2016. Total and soluble reactive phosphorus
http://dx.doi.org/10.1111/j.1365-2486.2009.01998.x. loadings to Lake Erie. J. Great Lakes Res. In press.
Fernandez, D.P., Neff, J.C., Belnap, J., Reynolds, R.L., 2006. Soil respiration in the cold desert Makarewicz, J.C., Bertram, P., 1991. Evidence for the restoration of the Lake Erie ecosys-
environment of the Colorado Plateau (USA): abiotic regulators and thresholds. Bio- tem. Bioscience 41, 216–223.
geochemistry 78, 247–265. http://dx.doi.org/10.1007/s10533-005-4278-0. Michalak, A.M., Anderson, E.J., Beletsky, D., Boland, S., Bosch, N.S., Bridgeman, T.B., Chaffin,
Friedman, J.H., Meulman, J.J., 2003. Multiple additive regression trees with application in J.D., Cho, K., Confesor, R., Daloglu, I., Depinto, J.V., Evans, M.A., Fahnenstiel, G.L., He, L.,
epidemiology. Stat. Med. 22, 1365–1381. http://dx.doi.org/10.1002/sim.1501. Ho, J.C., Jenkins, L., Johengen, T.H., Kuo, K.C., Laporte, E., Liu, X., McWilliams, M.R.,
Friedman, J., Hastie, T., Tibshirani, R., 2000. Additive logistic regression: a statistical Moore, M.R., Posselt, D.J., Richards, R.P., Scavia, D., Steiner, A.L., Verhamme, E.,
view of boosting (with discussion and a rejoinder by the authors). Ann. Stat. 28, Wright, D.M., Zagorski, M.A., 2013. Record-setting algal bloom in Lake Erie caused
337–407. by agricultural and meteorological trends consistent with expected future conditions.
I. Bertani et al. / Science of the Total Environment 575 (2017) 294–308 307
Proc. Natl. Acad. Sci. U. S. A. 110, 6448–6452. http://dx.doi.org/10.1073/pnas. Shen, L., Xu, H., Guo, X., 2012. Satellite remote sensing of harmful algal blooms (HABs)
1216006110. and a potential synthesized framework. Sensors 12, 7778–7803. http://dx.doi.org/
Millie, D.F., Fahnenstiel, G.L., Dyble Bressie, J., Pigg, R.J., Rediske, R.R., Klarer, D.M., Tester, 10.3390/s120607778.
P.A., Litaker, R.W., 2009. Late-summer phytoplankton in western Lake Erie (Lauren- Shuchman, R.A., 2014. Harmful algal bloom mapping in the Great Lakes. Michigan Tech
tian Great Lakes): bloom distributions, toxicity, and environmental influences. Research Institute Final Report. July 1, 2011–June 30, 2014. Chicago, IL..
Aquat. Ecol. 43, 915–934. http://dx.doi.org/10.1007/s10452-009-9238-7. Shuchman, R.A., Leshkevich, G., Sayers, M.J., Johengen, T., Brooks, C.N., Pozdnyakov, D.,
Millie, D.F., Weckman, G.R., Fahnenstiel, G.L., Carrick, H.J., Ardjmand, E., Young Ii, W.A., 2013. An algorithm to retrieve chlorophyll, dissolved organic carbon, and suspended
Sayers, M.J., Shuchman, R.A., 2014. Using artificial intelligence for CyanoHAB niche minerals from great lakes satellite data. J. Great Lakes Res. 39, 78–89.
modeling: discovery and visualization of Microcystis–environmental associations Smith, V.H., 2003. Eutrophication of freshwater and coastal marine ecosystems: a global
within western Lake Erie. Can. J. Fish. Aquat. Sci. 71, 1642–1654. http://dx.doi.org/ problem. Environ. Sci. Pollut. Res. 10, 126–139.
10.1139/cjfas-2013-0654. Sommaruga, R., Chen, Y., Liu, Z., 2009. Multiple strategies of bloom-forming microcystis to
Mitrovic, S.M., Oliver, R.L., Rees, C., Bowling, L.C., Buckney, R.T., 2003. Critical flow veloci- minimize damage by solar ultraviolet radiation in surface waters. Microb. Ecol. 57,
ties for the growth and dominance of Anabaena circinalis in some turbid freshwater 667–674. http://dx.doi.org/10.1007/s00248-008-9425-4.
rivers. Freshw. Biol. 48, 164–174. http://dx.doi.org/10.1046/j.1365-2427.2003. Soranno, P.A., 1997. Factors affecting the timing of surface scum and epilimnetic blooms
00957.x. of blue-green algae in a eutrophic lake. Can. J. Fish. Aquat. Sci. 54, 1965–1975.
Murtaugh, P.A., 2009. Performance of several variable-selection methods applied to real http://dx.doi.org/10.1139/cjfas-54-9-1965.
ecological data. Ecol. Lett. 12, 1061–1068. http://dx.doi.org/10.1111/j.1461-0248. Srivastava, A., Singh, S., Ahn, C.Y., Oh, H.M., Asthana, R.K., 2013. Monitoring approaches for
2009.01361.x. a toxic cyanobacterial bloom. Environ. Sci. Technol. 47, 8999–9013. http://dx.doi.org/
Nicodemus, K.K., Malley, J.D., 2009. Predictor correlation impacts machine learning algo- 10.1021/es401245k.
rithms: implications for genomic studies. Bioinformatics 25, 1884–1890. http://dx. Stumpf, R.P., Wynne, T.T., Baker, D.B., Fahnenstiel, G.L., 2012. Interannual variability of
doi.org/10.1093/bioinformatics/btp331. cyanobacterial blooms in Lake Erie. PLoS One 7, e42444. http://dx.doi.org/10.1371/
Obenour, D., Gronewold, A., Stow, C.A., Scavia, D., 2014. Using a Bayesian hierarchical journal.pone.0042444.
model to improve Lake Erie cyanobacteria bloom forecasts. Water Resour. Res. 50, Stumpf, R., Johnson, L., Wynne, T., Baker, D., 2016. Forecasting annual cyanobacterial
7847–7860. http://dx.doi.org/10.1002/2014WR015616.Received. bloom biomass to inform management decisions in Lake Erie. J. Great Lakes Res.
Paerl, H.W., Huisman, J., 2008. Blooms like it hot. Science 320 (80-.), 57–58. http://dx.doi. http://dx.doi.org/10.1016/j.jglr.2016.08.006.
org/10.1126/science.1155398. Taranu, Z.E., Zurawell, R.W., Pick, F., Gregory-Eaves, I., 2012. Predicting
Paerl, H.W., Huisman, J., 2009. Climate change: a catalyst for global expansion of harmful cyanobacterial dynamics in the face of global change: the importance of scale
cyanobacterial blooms. Environ. Microbiol. Rep. and environmental context. Glob. Chang. Biol. 18, 3477–3490. http://dx.doi.
Paerl, H.W., Otten, T.G., 2013a. Harmful cyanobacterial blooms: causes, consequences, and org/10.1111/gcb.12015.
controls. Microb. Ecol. 65, 995–1010. http://dx.doi.org/10.1007/s00248-012-0159-y. Taranu, Z.E., Gregory-Eaves, I., Leavitt, P.R., Bunting, L., Buchaca, T., Catalan, J., Domaizon, I.,
Paerl, H.W., Otten, T.G., 2013b. Blooms bite the hand that feeds them. Science 342 (80-.), Guilizzoni, P., Lami, A., McGowan, S., Moorhouse, H., Morabito, G., Pick, F.R.,
433–434. http://dx.doi.org/10.1126/science.1245276. Stevenson, M.A., Thompson, P.L., Vinebrooke, R.D., 2015. Acceleration of
Paerl, H.W., Paul, V.J., 2012. Climate change: links to global expansion of harmful cyanobacterial dominance in north temperate-subarctic lakes during the
cyanobacteria. Water Res. 46, 1349–1363. http://dx.doi.org/10.1016/j.watres.2011. Anthropocene. Ecol. Lett. http://dx.doi.org/10.1111/ele.12420.
08.002. Tisseuil, C., Vrac, M., Grenouillet, G., Wade, A.J., Gevrey, M., Oberdorff, T., Grodwohl, J.,
Paerl, H.W., Bland, P.T., Bowles, N.D., Haibach, M.E., 1985. Adaptation to high-intensity, Lek, S., 2012. Strengthening the link between climate, hydrological and species
low-wavelength light among surface blooms of the cyanobacterium Microcystis distribution modeling to assess the impacts of climate change on freshwater bio-
aeruginosa. Appl. Environ. Microbiol. 49, 1046–1052. diversity. Sci. Total Environ. 424, 193–201. http://dx.doi.org/10.1016/j.scitotenv.
Paerl, H.W., Hall, N.S., Calandrino, E.S., 2011. Controlling harmful cyanobacterial blooms in 2012.02.035.
a world experiencing anthropogenic and climatic-induced change. Sci. Total Environ. United States, Canada, 2012. Great Lakes Water Quality Agreement - Annex 4.
409, 1739–1745. http://dx.doi.org/10.1016/j.scitotenv.2011.02.001. United States, Canada, 2015. Recommended Phosphorus Loading Targets for Lake Erie -
Park, Y.-J., Ruddick, K., Lacroix, G., 2010. Detection of algal blooms in European waters Annex 4 Objectives and Targets Task Team Final Report to the Nutrients Annex
based on satellite chlorophyll data from MERIS and MODIS. Int. J. Remote Sens. 31, Subcommittee.
6567–6583. Vanderploeg, H.A., Liebig, J.R., Carmichael, W.W., Agy, M.A., Johengen, T.H., Fahnenstiel,
Perovich, G., Dortch, Q., Goodrich, J., Berger, P.S., Brooks, J., Evens, T.J., Gobler, C.J., Graham, G.L., Nalepa, T.F., 2001. Zebra mussel (Dreissena polymorpha) selective filtration pro-
J., Hyde, J., Karner, D., O'Shea, D.K., Paul, V., Paerl, H., Piehler, M., Rosen, B.H., moted toxic Microcystis blooms in Saginaw Bay (Lake Huron) and Lake Erie. Can.
Santelmann, M., Pat Tester, J.W., 2008. Causes, prevention, and mitigation. In: J. Fish. Aquat. Sci. 58, 1208–1221.
Hudnell, H.K. (Ed.), Cyanobacterial Harmful Algal Blooms - State of the Science and Verhamme, E., Redder, T., Schlea, D., Grush, J., Bratton, J., DePinto, J., n.d. Development of
Research Needs. Springer. the Western Lake Erie Ecosystem Model (WLEEM): application to connect phospho-
Prasad, A.M., Iverson, L.R., Liaw, A., 2006. Newer classification and regression tree tech- rus loads to cyanobacteria biomass. J. Great Lakes Res. (in review).
niques: bagging and random forests for ecological prediction. Ecosystems 9, Verspagen, J.M.H., Snelder, E.O.F.M., Visser, P.M., Jöhnk, K.D., Ibelings, B.W., Mur, L.R.,
181–199. http://dx.doi.org/10.1007/s10021-005-0054-1. Huisman, J., 2005. Benthic-pelagic coupling in the population dynamics of the harm-
Preston, T., Stewart, W.D.P., Reynolds, C.S., 1980. Bloom-forming cyanobacterium ful cyanobacterium Microcystis. Freshw. Biol. 50, 854–867. http://dx.doi.org/10.1111/
Microcystis aeruginosa overwinters on sediment surface. Nature 288, 365–367. j.1365-2427.2005.01368.x.
http://dx.doi.org/10.1038/288365a0. Verspagen, J.M.H., Passarge, J., Jöhnk, K.D., Visser, P.M., Boers, P., Laanbroek, H.J., Huisman,
Reinart, A., Kutser, T., 2006. Comparison of different satellite sensors in detecting J., 2006. Water management strategies against toxic microcystis blooms in the Dutch
cyanobacterial bloom events in the Baltic Sea. Remote Sens. Environ. 102, 74–85. Delta. Ecol. Appl. 16, 313–327.
Reynolds, C., 2006. Ecology of Phytoplankton. Cambridge University Press, New York. Visser, P.M., Ibelings, B.W., Van der Veer, B., Koedood, J., Mur, L.R., 1996. Artificial mixing
Richards, R.P., Baker, D.B., 2002. Trends in water quality in LEASEQ rivers and streams prevents nuisance blooms of the cyanobacterium Microcystis in Lake Nieuwe Meer,
(northwestern Ohio), 1975–1995. Lake Erie agricultural systems for environmental the Netherlands. Freshw. Biol. 36, 435–450. http://dx.doi.org/10.1046/j.1365-2427.
quality. J. Environ. Qual. 31, 90–96. http://dx.doi.org/10.2134/jeq2002.0090. 1996.00093.x.
Richards, R.P., Baker, D.B., Crumrine, J.P., Kramer, J.W., Ewing, D.E., Merryfield, B.J., 2008. Wagner, C., Adrian, R., 2009. Cyanobacteria dominance: quantifying the effects of climate
Thirty-year trends in suspended sediment in seven Lake Erie tributaries. J. Environ. change. Limnol. Oceanogr. 54, 2460–2468. http://dx.doi.org/10.4319/lo.2009.54.6_
Qual. 37, 1894–1908. http://dx.doi.org/10.2134/jeq2007.0590. part_2.2460.
Rouse, J.W., Haas, R.H., Schell, J.A., Deering, D.W., 1973. Monitoring vegetation systems in Wallace, B.B., Hamilton, D.P., 1999. The effect of variations in irradiance on buoyancy reg-
the Great Plains with ERTS. Third ERTS Symposium, NASA SP-351 I, pp. 309–317. ulation in Microcystis aeruginosa. Limnol. Oceanogr. 44, 273–281. http://dx.doi.org/
Sayers, M., Fahnenstiel, G.L., Shuchman, R.A., Whitley, M., 2016. Cyanobacteria blooms in 10.4319/lo.1999.44.2.0273.
three eutrophic basins of the Great Lakes: a comparative analysis using satellite re- Walsh, C.J., Webb, J.A., 2016. Interactive effects of urban stormwater drainage, land
mote sensing. Int. J. Remote Sens. 37, 4148–4171. clearance, and flow regime on stream macroinvertebrate assemblages across a
Scavia, D., David Allan, J., Arend, K.K., Bartell, S., Beletsky, D., Bosch, N.S., Brandt, S.B., large metropolitan region. Freshw. Sci. 35, 324–339. http://dx.doi.org/10.1086/
Briland, R.D., Daloǧlu, I., DePinto, J.V., Dolan, D.M., Evans, M.A., Farmer, T.M., Goto, 685105.
D., Han, H., Höök, T.O., Knight, R., Ludsin, S.A., Mason, D., Michalak, A.M., Peter Wang, J., Xie, P., Guo, N., 2007. Effects of nonylphenol on the growth and microcystin pro-
Richards, R., Roberts, J.J., Rucinski, D.K., Rutherford, E., Schwab, D.J., Sesterhenn, duction of Microcystis strains. Environ. Res. 103, 70–78. http://dx.doi.org/10.1016/j.
T.M., Zhang, H., Zhou, Y., 2014. Assessing and addressing the re-eutrophication of envres.2006.05.013.
Lake Erie: central basin hypoxia. J. Great Lakes Res. 40, 226–246. http://dx.doi.org/ Watson, S.B., Miller, C., Arhonditsis, G., Boyer, G.L., Carmichael, W., Charlton, M.N.,
10.1016/j.jglr.2014.02.004. Confesor, R., Depew, D.C., Ludsin, S.A., Matisoff, G., Mcelmurry, S.P., Ho, T.O.,
Scavia, D., DePinto, J.V., Bertani, I., 2016. A multi-model approach to evaluating target Murray, M.W., Richards, R.P., Rao, Y.R., Steffen, M.M., Wilhelm, S.W., 2016. The re-eu-
phosphorus loads for Lake Erie. J. Great Lakes Res. http://dx.doi.org/10.1016/j.jglr. trophication of Lake Erie: harmful algal blooms and hypoxia. Harmful Algae 56,
2016.09.007. 44–66. http://dx.doi.org/10.1016/j.hal.2016.04.010.
Schapire, R.E., 2003. The boosting approach to machine learning - an overview. In: Wynne, T., Stumpf, R., 2015. Spatial and temporal patterns in the seasonal distribution of
Denison, D.D., Hansen, M.H., Holmes, C.C., Mallick, B., Yu, B. (Eds.), Nonlinear Estima- toxic cyanobacteria in Western Lake Erie from 2002–2014. Toxins (Basel). 7,
tion and Classification, Lecture Notes in Statistics. Springer New York, New York, NY 1649–1663. http://dx.doi.org/10.3390/toxins7051649.
http://dx.doi.org/10.1007/978-0-387-21579-2. Wynne, T.T., Stumpf, R.P., Tomlinson, M.C., Warner, R.A., Tester, P.A., Dyble, J., Fahnenstiel,
Segurado, P., Branco, P., Jauch, E., Neves, R., Ferreira, M.T., 2016. Sensitivity of river fishes G.L., 2008. Relating spectral shape to cyanobacterial blooms in the Laurentian
to climate change: the role of hydrological stressors on habitat range shifts. Sci. Total Great Lakes. Int. J. Remote Sens. 29, 3665–3672. http://dx.doi.org/10.1080/
Environ. 562, 435–445. http://dx.doi.org/10.1016/j.scitotenv.2016.03.188. 01431160802007640.
308 I. Bertani et al. / Science of the Total Environment 575 (2017) 294–308
Wynne, T.T., Stumpf, R.P., Tomlinson, M.C., Dyble, J., 2010. Characterizing a cyanobacterial Wynne, T.T., Stumpf, R., Briggs, T., 2013. Comparing MODIS and MERIS spectral shapes for
bloom in Western Lake Erie using satellite imagery and meteorological data. Limnol. cyanobacterial bloom detection. Int. J. Remote Sens. 34, 6668–6678. http://dx.doi.org/
Oceanogr. 55, 2025–2036. http://dx.doi.org/10.4319/lo.2010.55.5.2025. 10.1080/01431161.2013.804228.
Wynne, T., Stumpf, R., Tomlinson, M.I., Schwab, D., Watabayashi, G., Christensen, J., 2011. Zhou, Y., Michalak, A.M., Beletsky, D., Rao, Y.R., Richards, R.P., 2015. Record-breaking Lake
Estimating cyanobacterial bloom transport by coupling remotely sensed imagery and Erie hypoxia during 2012 drought. Environ. Sci. Technol. 49, 800–807. http://dx.doi.
a hydrodynamic model. Ecol. Appl. 21, 2709–2721. org/10.1021/es503981n.