Remote Sensing of Environment 155 (2014) 161–177

Contents lists available at ScienceDirect

Remote Sensing of Environment

j ourna l homepage: www.e lsev ie r .com/ locate / rse

Eutrophication and cyanobacterial blooms in South Africaninland waters: 10 years of MERIS observations

Mark William MatthewsDepartment of Oceanography, University of Cape Town, Rondebosch, 7701 Cape Town, South Africa

http://dx.doi.org/10.1016/j.rse.2014.08.0100034-4257/© 2014 Elsevier Inc. All rights reserved.

a b s t r a c t

a r t i c l e i n f o

Article history:Received 21 May 2014Received in revised form 25 July 2014Accepted 5 August 2014Available online 26 September 2014

Keywords:Remote sensingInland watersSouth AfricaEutrophicationCyanobacteriaChlorophyll-aMERISMPH algorithm

The medium resolution imaging spectrometer full resolution (MERIS FR) archive (2002 to 2012) over SouthAfrica has been processed with the maximum peak height (MPH) algorithm for the 50 largest standing waterbodies in South Africa. Time series of chlorophyll a (chl-a), cyanobacteria and surface scum area coverage wereused to establish the status, seasonality and trends for each of the water bodies. The majority (62%) of the 50water bodies were hypertrophic (mean chl-a N 30 mg m−3), while 23 water bodies had intermediate to exten-sive cyanobacteria coverage. Cyanobacterial surface scum events posing a high health risk occurred in 26 of thewater bodies in varying extents. Analysis of significant trends showed that water bodies both worsened and im-provedwith regard to eutrophication, cyanobacteria and surface scum coverage between 2005 and 2012. Valida-tion of the MPH algorithm using an independent in situ dataset demonstrated that gross trophic status throughchl-a can be accurately determined in both eukaryote and cyanobacteria-dominant waters. Chl-a estimation inoligo/mesotrophic waters remains challenging due to a wide range of potential sources of error. This study isthe first of its kindproviding quantitative chl-a and phytoplankton species information in lakes from a time seriesof satellite remotely senseddata on a sub-continental scale, demonstrating howglobal analyses of changes in lakebiogeochemistry might be performed in the future. It demonstrates the pivotal role that satellite remote sensingcan play in supplementing in situ monitoring efforts, particularly in the developing world. The study concludesthat eutrophication and cyanobacterial blooms are widespread in South African water bodies and remain issuesof critical concern for water quality.

© 2014 Elsevier Inc. All rights reserved.

1. Introduction

Globally, lakes play an important role in regulating earth's climate,acting as sentinels and regulators of climate change (Williamson,Saros, Vincent, & Smol, 2009), as well as being a crucial life-giving re-source for humanity. Global observations of lakes using satellite remotesensing has rapidly progressed in the last two decades with global timeseries observations of surfacewater level (Crétaux et al., 2011) andwet-land inundation having been obtained (Prigent, Papa, Aires, Rossow, &Matthews, 2007). However, there exists a paucity of information re-garding biogeochemical water quality information on a global scale forlakes, which constitutes a significant gap in understanding the role oflakes in earth's biogeochemical cycle (nutrients and carbon) and theirresponse to global change. Phytoplankton are an ideal indicator ofresponses to changes in biogeochemistry in lakes, with chl-a andcyanobacterial biomass estimates being robust measures of ecologicalintegrity (Carvalho et al., 2012). Anthropogenic changes in nutrientcycles, hydrology and climate have led to an increased incidence ofcyanobacterial algal blooms globally (Michalak et al., 2013; Paerl &Huisman, 2009). However currently there exists no analysis concerningthe possible extent and magnitude of these changes on phytoplankton

biomass (chl-a) and cyanobacterial blooms on global or continentalscales.

Satellite remote sensing of biogeochemical parameters in water hasbeen extensively applied to inland waters (for detailed reviews of stud-ies and methods see Matthews, 2011; Odermatt, Gitelson, Brando, &Schaepman, 2012). Time series observations of parameters related tophytoplankton and water clarity derived from various satellite instru-ments exist for several lakes (e.g. Hu et al., 2010; Olmanson, Bauer, &Brezonik, 2008; Stumpf, Wynne, Baker, & Fahnenstiel, 2012). However,these and other existing studies do not provide quantitative chl-a esti-mates, nor do they cover a large number of lakes on a sub-continentalscale. Therefore their usefulness in contributing to global biogeochemi-cal models is severely limited. Global, continuous estimates of biogeo-chemical parameters in lakes from satellite have been hindered by theabsence of suitable products (i.e. algorithms) and by the lack of contin-uous satellite missions with the required spatial, spectral, temporal andradiometric resolutions. Recent South African studies demonstrate themedium resolution imaging spectrometer (MERIS) as the optimal pastsensor for providing detailed water quality information productsowing to its spectral, temporal, radiometric and spatial resolution(Matthews, Bernard, & Robertson, 2012; Matthews, Bernard, & Winter,

162 M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

2010). Briefly,MERIS has a spatial resolution of 260 by290mdependingon the altitude, an acquisition frequency of 2 to 3 days, and 15 spectralbands ideally positioned for water related applications. These sensorspecifications allow for sufficient observational frequency and sensitiv-ity to enable viable change detection on an approximately weekly time-scale, not offered by themany currently available high spatial resolutionvisible radiometers (e.g. Landsat or SPOT). Therefore, MERIS provided aunique opportunity for observing lakes on a global scale and high fre-quency, unparalleled by current in situ monitoring capabilities whichmay also be hindered by high costs or lack of capacity.

The maximum peak height (MPH) algorithm designed for opera-tional use with MERIS is unique in that it provides quantitative chl-aestimates for a wide range of trophic states, and robust identificationof cyanobacteria (seeMatthews et al., 2012). TheMPH algorithm avoidserror-prone atmospheric correction schemes required bymost productsthrough correcting only for gaseous absorption and Rayleigh scattering(not aerosols). The MPH algorithm can currently only be applied withthe MERIS sensor or hyperspectral instruments, since other current op-erational satellite-based instruments lack sufficient spectral resolutionin the red. However, its applicability to the forthcoming ESA SentinelOcean and Land Colour Instrument (OLCI) should provide good productcontinuation for at least a further decade. This study presents resultsfrom the MPH algorithm applied to the full time series of MERIS ob-servations (2002 to 2012) for the 50 largest South African inlandwater bodies. The study is unique with respect to both the quantitativetime series products and its application to water bodies on a national/sub-continental scale (South Africa). This study is the first of its kindin aiming to provide quantitative biogeochemical products and cyano-bacteria identification for a large number of lakes on a sub-continentalscale from the MERIS full archive.

A great deal of information has been published on the occurrenceand distribution of eutrophication and cyanobacteria in South Africanwater bodies (e.g. Harding & Paxton, 2001; Oberholster & Ashton,2008; Oberholster, Botha, & Cloete, 2005; van Ginkel, 2011, 2012; vanGinkel, Silberbauer, & Vermaak, 2000). This includes recent reviews oneutrophication research (van Ginkel, 2011, 2012). In particular, moni-toring efforts have been driven by the national eutrophication monitor-ing programme (NEMP) by the Department of Water Affairs (DWA)(van Ginkel et al., 2000). These studies indicate that eutrophicationand cyanobacterial blooms are widespread and extensive in SouthAfrican water bodies, and poisonings of domestic and wild animals bycyanobacterial toxins are a frequent occurrence (Oberholster & Botha,2010; Oberholster et al., 2005). However, they are based on in situdiscrete-point samples which are limited both spatially and temporally.Therefore, despite the great monitoring efforts, there remains muchuncertainty regarding the overall status of eutrophication and cyano-bacterial blooms in South African water bodies both from a temporaland spatial perspective. This is exacerbated by the very large numberof manmade impoundments (497 reservoirs with capacity larger thanonemillion cubic metres) as a result of the aridity of the region and rel-ative scarcity of water in South Africa, and almost complete absence oflarge natural lakes (Oberholster & Ashton, 2008).

Furthermore, for most South African water bodies the phenology,seasonal variability and trends of phytoplanktonproduction and speciescomposition remain largely uncharacterised. The rate at which the oc-currence of problem eukaryote species, for example the dinoflagellateCeratium hirundinella, is increasing is also largely unknown (Hart &Wragg, 2009; Van Der Walt, 2011; van Ginkel, Hohls, & Vermaak,2001). Information on the severity and occurrence of cyanobacterialscums (or mats) is also absent for South African water bodies. Surfacescumshave significant negative ecological consequences on the diversityand functioning of the plankton community and higher order organisms,and associated toxin production is a health threat for potable and recre-ational water use (Chorus, Falconer, Salas, & Bartram, 2000; Oberholster& Botha, 2010). Thus the lack of information on their occurrence and ex-tent in South African water bodies constitutes a risk to public health.

Cyanobacterial surface scums also serve as an important ecological indi-cator of over-enrichment and meteorological warming and senescence(Michalak et al., 2013; Paerl & Huisman, 2009). Therefore, surfacescum occurrence is likely to be a good indicator of lakes or regions sub-ject to severe eutrophication or potential climate warming.

This paper aims to characterise the status, seasonality, and trends ofeutrophication (chl-a) and cyanobacterial blooms in 50 South Africanwater bodies using the full MERIS archive. In addition, the study seeksto validate the chl-a estimates from the MPH algorithm using an inde-pendently acquired in situ dataset. The provision of a satellite baseddataset supplementing in situ monitoring data seeks to fill an informa-tion gap in the limnology of South African inlandwaters, as well as pro-vide input into biogeochemical ecosystemmodels and regional climatechange studies incorporating lakes, and to models predicting phyto-plankton biomass and surface scum formation from meteorologicaland hydrological variables (e.g. Soranno, 1997).

2. Methods

2.1. Reservoir selection and data processing

The 50 largest water bodies in South Africa by surface area wereselected for the analysis (see Table 1 for surface area and coordinates).The surface area of the water bodies was obtained from high resolutionshapefiles (RQS, 2004). Seasonal pans, estuaries and water bodies sub-ject to tidal influence were excluded. Weekly water level data availablefor most of the water bodies was used to identify low-water periodsbetween 2002 and 2012 (DWA, 2013a). Water bodies with extendedlow-water (drought) periods or which were too narrow to be viewedwithout overwhelming adjacency effects (less than 600 m wide) wereexcluded.

The MPH algorithm was used to compute the concentration ofsurface chl-a as well as identify cyanobacterial blooms and surfacescum in the water bodies (for detailed description see Matthews et al.,2012). Its usage here is intended to: 1) identify the trophic statusof the water bodies through chl-a estimates; 2) identify cyano-bacterial blooms, and; 3) identify cyanobacterial surface scum condi-tions (defined here by chl-a N 350 mg m−3). Chl-a values as low as350 mg m−3 have been associated with scum conditions (Jupp,Kirk, & Harris, 1994). The threshold value is chosen to encompasssurface accumulations of cyanobacteria which, although not strictlypossessing a 753 nm peak position, constitute a high risk for scumformation, which exists for chl-a values as low as 50 mg m−3

(Chorus et al., 2000). The threshold value necessarily dictates thequantity of scum detected, with lower values being more conserva-tive from a health-risk perspective, and vice versa.

ArchivedMERIS full resolution (FR) level 1P datawere obtained overSouth Africa between the years 2002 and 2012. The files contain 15bands with uncorrected top-of-atmosphere (TOA) radiances. MERISdata were processed using open source processing tools in a Linux envi-ronment using Python programming language (V. 2.7.1). A schematic ofthe data processing chain is shown in Fig. 1. The accurate MERIS ortho-rectified geo-location operational software (AMORGOS, ACRI-ST) wasused to ortho-rectify the data. AMORGOS provides accurate geo-location for MERIS imagery to within one pixel accuracy (b300 m)which is essential for extracting small targets such as the water bodiesin this study. The inputs are the MERIS FR files and the auxiliary orbitfiles corresponding to the specific scene (downloaded from ESA).AMORGOSuses a digital elevationmodel (DEM) to compute the latitudeand longitude of a pixel's first line of sight and intersection with theearth's surface. The outputs are geo-corrected latitude and longitudebands for the scene written into a new product file. Importantly, thedata are not altered by the geo-correction process. A few scenes whichhad persistent geo-location errorswere removed after visual inspection.

The files were processed after ortho-correction using the BasicENVISAT Toolbox for (A)ATSR and MERIS (BEAM) V. 4.9.0.1. Within

Table 1Reservoir characteristics and statistics derived from 10 years of MERIS observations. Bold font indicates trends significant at the 95% confidence interval of the Student t-distribution. The

mean values for chl-a chl� �

, cyanobacteria Acy

� �and surface scum Asc

� �coverage estimated by the MPH algorithm are indicated along with the amplitude (A) and phase (ϕ) (timing)

indicated by abbreviated month of the year. The trends for both the yearly averages (achl) and the anomalies (achl′) for chl-a are indicated as well as for cyanobacteria aAcy

� �and surface

scum aAsc

� �area coverage (when applicable).

Reservoir name Area Lat. Lon. Alt. N chl Achl ϕchl achl′ achl Acy ϕAcyaAcy Asc ϕAsc

aAsc

km2 m asl mg m−3 mg m−3 mg m−3 y−1 mg m−3 y−1 % % y−1 % % y−1

Gariep 346.6 −30.69 25.71 1260 650 137.2 253.1 Feb −4.2 1.2 11.6 Mar −1 0.1 Apr −0.01Vaal 251.4 −26.90 28.14 1490 756 99.6 119.5 Feb −0.9 0.7 27.1 Apr 0.1 0.2 Mar 0.04Bloemhof 186.5 −27.67 25.65 1235 521 142.8 180.5 Feb 1.7 5.5 22.1 Mar 1.7 0.8 Apr −0.05Pongolapoort 113.4 −27.37 31.95 140 625 3.8 7.9 Jan −0.3 −0.4 0.3 Feb −0.01 0.0 Feb –

Vanderkloof 113.1 −30.15 24.88 1170 628 55.3 140.6 Feb 12.7 17.3 5.1 Mar −0.7 0.0 Mar −0.01Sterkfontein 60.8 −28.43 29.02 1702 566 3.3 6.3 Jan 0.1 0.0 0.0 Jul 0 – – –

Lake Sibhayi 54.7 −27.37 32.69 20 647 10.8 11.0 Nov 0.2 0.2 2.9 Jun 0.07 – – –

Darlington 52.2 −33.15 25.15 239 412 120.5 31.8 Mar 5.3 6.1 19.8 May −2.1 1.3 Apr −0.06Theewaterskloof 50.8 −34.03 19.20 307 176 43.1 52.8 Nov −7 −9.2 10.4 Sep −0.8 – – –

Heyshope 48.8 −27.03 30.50 1308 727 23.7 25.8 Jan 2.1 2.4 1.8 Jun 0.1 – – –

Kalkfontein 42.0 −29.52 25.26 1241 548 117.3 151.4 Mar 5.1 8.9 26.8 Jan −1.2 0.3 Jan 0.00Grootdraai 32.7 −26.93 29.33 1549 731 81.8 124.3 Jan −0.6 0.6 18.9 Mar 0.3 0.0 Mar 0.00Spitskop 31.7 −28.09 24.55 1045 525 163.4 237.4 Aug 6.9 6.4 31.7 Apr −1.5 1.3 Dec −0.10Erfenis 29.9 −28.55 26.84 1331 534 351.4 451.7 Feb −42.1 −38.5 21.4 Apr 1.4 0.1 Apr 0.01Kuhlange (Kosi Lake) 28.3 −26.98 32.84 2 629 8.3 10.1 Nov −1.9 −2.2 13.1 Jul −3.2 – – –

Allemanskraal 26.4 −28.30 27.21 1371 573 248.0 483.2 Feb 25.8 32.2 22.7 Apr −0.5 0.1 Apr 0.10Woodstock 26.1 −28.72 29.21 1180 584 14.3 26.0 Mar 0.4 0.9 0.5 Jun −0.03 – – –

Loskop 23.0 −25.43 29.32 1006 863 58.1 145.2 Feb 0.6 1.2 0.4 Feb 0.3 0.0 Feb 0.03AlbertFalls 21.8 −29.44 30.41 658 514 8.4 17.1 Jan −1.7 −1.9 0.0 Jun 0 – – –

Brandvlei 18.7 −33.71 19.43 203 214 6.4 10.4 Oct −1.3 −2.8 1.6 May −2.3 – – –

Ntshingwayo 18.7 −27.99 29.91 1248 566 314.2 717.1 Mar 1.4 1.9 15.3 Dec 2.1 0.0 Feb 0.00Tzaneen 16.8 −23.79 30.15 728 592 39.9 47.0 Jan −5.8 −5.5 0.0 Sep – – – –

Hartbeespoort 15.6 −25.75 27.86 1165 775 92.6 186.0 Jan −1.0 −1.1 48.6 Feb 0.5 10.2 Jan 0.08Krugersdrift 15.3 −28.87 26.00 1252 538 317.9 455.7 Apr 3.4 11.2 17.1 Dec 1.3 0.1 Feb −0.01Umtata (Mthatha) 15.2 −31.54 28.73 695 598 478.3 572.3 Jan −53.9 −56.3 9.8 Apr −0.3 – – –

Voëlvlei 14.5 −33.36 19.04 75 242 114.7 130.1 Nov −28.8 −42.7 10.0 Mar −1 – – –

Midmar 14.3 −29.51 30.19 1046 533 4.4 8.8 Jan 0.0 0.0 0.0 Sep – – – –

Xonxa 14.0 −31.82 27.14 939 625 65.6 147.7 Mar 12 16.6 2.5 Jan −0.3 0.0 Jan –

Spioenkop 13.9 −28.69 29.49 1091 583 67.5 193.7 Jan 1.4 5.6 1.3 Jun 0.1 – – –

Ncora 12.8 −31.78 27.65 1038 627 380.5 592.6 Mar −20.1 −15.6 10.5 May 0.01 – – –

Barberspan 12.7 −26.58 25.59 1354 637 42.0 31.9 Feb −3.6 −3.5 52.8 May −6.3 0.0 May –

Klipvoor 12.6 −25.15 27.82 994 777 145.7 79.0 Jan −6.7 −7.1 6.0 Dec 1.2 0.5 Dec 0.10Grassridge 12.5 −31.77 25.47 1055 648 176.7 256.7 Feb −1.1 2.7 14.7 Oct −0.5 – – –

Koppies 11.4 −27.23 27.69 1424 670 133.0 287.3 Feb −0.7 2.8 46.9 Oct −0.3 0.0 Jan –

Zaaihoek 10.9 −27.43 30.09 1737 707 14.3 4.0 Apr −2.6 −2.7 0.8 Jun −0.3 – – –

Lubisi 10.7 −31.79 27.41 1020 669 82.8 231.1 Mar 11.2 15.3 4.0 Feb 0.09 0.0 Apr –

Lake Chrissiesmeer 10.4 −26.33 30.22 1676 723 116.8 134.5 Jan 1.5 1.8 37.1 Jun −0.6 0.1 Mar 0.02FlagBoshielo 10.4 −24.83 29.44 821 862 42.6 68.7 Jan −3.5 −3.1 18.6 Aug 3.3 – – –

Goedertrou 10.1 −28.77 31.44 228 541 20.1 20.3 Jun −1.3 −1.1 0.5 Nov −0.3 – – –

Rustfontein 10.1 −29.30 26.63 1374 557 115.6 174.3 Jan 16.7 21.6 21.4 Apr 1.3 0.0 May −0.03Fairview 9.8 −26.16 31.66 297 693 33.3 41.3 Mar −6.9 −6.8 1.7 Apr −0.5 – – –

Vaalkop 9.7 −25.31 27.47 987 752 108.5 154.4 Jan −12.4 −14 7.0 Apr 2.3 – – –

Kwena 9.6 −25.37 30.37 1185 746 30.2 74.3 Jan −14.2 −14.9 1.1 Oct −0.8 0.0 Jul –

Roodekoppies 8.9 −25.41 27.59 1018 750 147.0 205.2 Dec −10.7 −9.6 13.1 Mar 1.6 0.5 Mar 0.11Witbank 8.7 −25.90 29.31 1507 745 117.5 276.8 Feb −7.2 −6.0 0.7 Jan −0.1 0.3 Jan −0.09Lake Msingazi 8.7 −28.76 32.10 5 553 5.6 7.1 Dec 0.4 0.4 0.0 Jun −0.1 – – –

Bronkhorstspruit 7.7 −25.90 28.69 1435 747 105.9 163.1 Feb −10.5 −10.9 0.4 Aug −0.1 – – –

Jericho 7.7 −26.64 30.48 1470 696 18.7 30.4 Dec −0.1 −0.3 0.1 Dec 0.02 – – –

Mokolo 7.3 −23.98 27.75 912 708 31.4 54.0 Dec 0.7 1.5 0.0 Sep 0 – – –

Inanda 6.7 −29.69 30.87 151 510 27.3 50.6 Mar −1.1 0.2 3.1 Feb 0.5 0.2 Mar 0.00

163M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

the graph processing framework in BEAM the files were subsetted ac-cording to a bounding box for each of thewater bodies, corrected for ra-diometric effects using the Radiometry Processor V. 1.0.1 (Bouvet &Ramoino, 2009), and corrected for gaseous absorption using thebottom-of-Rayleigh reflectance (BRR) processor V. 2.3 (ACRI, 2006;Santer, Carrere, Dubuisson, & Roger, 1999) (see Matthews et al., 2012for further details). The output is a subsetted BRR product for eachreservoir in the scene. Clouds were detected using the Cloud ProcessorV. 1.5.203 (© ESA, FUB, and Brockmann Consult, 2004). The MERISBRR subset files were then used to compute the MPH products usingNumpy (V.1.5.1) in Python. The output products for each reservoirwere plotted usingmatplotlib tools (V.1.0.1) and saved to a hierarchicaldata format (HDF) file for further processing.

Conventional methods used to distinguish between land and water,such as empirically-based land–water separation algorithms, often faildue to the high reflectance values associated with cyanobacterialblooms and surface scums. Therefore water pixels were extractedusing shapefiles drawn for each of the water bodies from high resolu-tion (30 m) archived Landsat data. Shapefiles for the minimum andmaximum water extents were produced with weekly water level data(DWA, 2013a). In certain cases several shapefiles were needed due tofluctuating water levels. Where water level data were not available(only 5 natural lakes) the minimum water extent was determined byexamining archived Landsat data between 2002 and 2012. This rigorousprocedure ensured that land pixels were not classified aswater and viceversa.

Fig. 1. Schematic processing chain for MERIS data. Inputs are shaded in blue, outputs in green and processes in red. See text for details.

164 M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

2.2. Time series methods

Time series products for chl-a, cyanobacteria and surface scum arealcoverage were computed for each water body. The average chl-a valueof each image was determined as the median chl-a value of validwater pixels. Given the relatively small size of thewater bodies, a singleaverage value for each scene is generally appropriate. The median waschosen as it is less sensitive to outliers than the arithmeticmean. Imagesinwhich less than 5% of the lake areawas visible due to cloud and/or po-sition of the image swath were removed. The average chl-a value foreach scene was used in subsequent time series analyses. The seasonal(monthly) cycle for each water body was computed as the monthlymean using the entire time series of observations. The phase,ϕ, and am-plitude, A, of the seasonal cycle was determined as the month with thegreatest chl-a value, and the difference between the months with thesmallest and largest chl-a values, respectively. The chl-a anomalies, de-noted chl-a′, were computed by subtracting themonthly chl-a averagesfrom chl-a values in that month— this effectively removes the seasonalcycle for calculation of the trend. The trend of chl-awas computed usingthe anomalies and linear regression analysis according to y(t)= at+ b;the regression coefficient a was determined as the linear trend withrespect to time t. The statistical significance of the trend was computedusing the Student t-test statistic according to:

tr ¼rffiffiffiffiffiffiffiffiffiffiffiN−2

p

1−r2

where r, the correlation coefficient, is computed as:

r ¼ 1N−1

XNt¼1

x tð Þ−xσx

� t−tσ t

where N is the number of observations, x is the variable of interest, t istime, and σ is the standard deviation.

Yearly chl-a averages were computed by calculating the mean of allobservations within a given year. The trend of the yearly averages wascomputed using linear regression analysis as for the seasonal anomalies.Calculation of trends was limited to between 2005 and 2011 due tosome missing data between 2002 and 2004 and excluding incompleteyears, e.g. 2012. The mean value of all chl-a observations for the entire

time series was used to determine the overall trophic state of the reser-voir. Trophic status was defined using the Organisation for EconomicCo-operation and Development chl-a thresholds applicable to inlandwaters: oligotrophic, 0 to 10 mg m−3; mesotrophic, 10 to 20 mg m−3;eutrophic, 20 to 30 mg m−3; and hypertrophic, N30 mg m−3.

The area coverage of cyanobacteria, Acy, and surface scum, Asc, in per-centwere computed as the number of pixels identified as cyanobacteriaor surface scum normalised by the number of water pixels visible inthe image. Surface scum is defined as pixels which are identified ascyanobacteria and have chl-a N 350 mg m−3. The overall mean areacoverage in percent for cyanobacteria and surface scumwere computedfor the entire time series for each reservoir. The seasonality ofcyanobacteria and surface scum area coverage was determined usingmonthly means, as for chl-a, and the phase (or timing) determined asthe month with the greatest coverage. The trends of cyanobacteriaand surface scum area coverages were determined, when appropriate,using the annual means between 2005 and 2011, as for chl-a.

2.3. In situ validation

Validation of satellite chl-a estimates was performed using in situchl-a data collected by the NEMP (DWA, 2013b). Unfortunately, theexact time and location of the discrete-point sampleswere not recorded(pers. comm.Michael Silberbauer, 2013). Nevertheless, amatchup anal-ysis was performed where sufficient data were available, but should beconsidered preliminary. Only satellite and in situ data collected on thesame day were used, and chl-awas determined following the methodsof Sartory andGrobbelaar (1984)with samples being transported on iceto the laboratory. Pixels contained in a small area (10 pixels ≈ 1 km2)corresponding to the likely location of the in situ sample point were ex-tracted from the time series of satellite data and themean and standarddeviation was computed. The correlation between satellite and in situmeasurements were calculated using linear regression analysis. Suffi-ciently good matchups were obtained for the following reservoirs: Al-bert Falls, Bronkhorstspruit, Hartbeespoort, Klipvoor, Midmar, Inandaand Vaalkop.Water bodieswith very fewmatchups or poor correlationslikely due to amismatch between the location of the sampling point andsatellite observations were excluded. The overall algorithm perfor-mance was evaluated for the combined matchup dataset, as well asfor waters identified as cyanobacteria and eukaryote dominant. The

165M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

comparison was limited to ch-a b 350mgm−3 because of the very highspatial patchiness associated with surface scums and extremely highbiomass blooms. The goodness of fit was determined using the rootmean square error (RMSE) which was calculated as:

RMSE ¼P

chl‐asatellite−chl‐ainsituð Þ2N−2

!1=2

:

The bias (average residual) was calculated as:

Bias ¼X

chl‐asatellite−chl‐ainsituð ÞN

:

The mean absolute percentage error (MAPE) was calculated by:

MAPE ¼ 100N

X chl‐asatellite−chl‐ainsituchl‐ainsitu

where N is the number of observations.

2.4. Error estimates and uncertainties

In situ and satellite derived chl-a estimates are subject to uncer-tainties from a variety of sources. The discrete-point sample error(Matthews et al., 2012) for in situmeasurements results in large uncer-tainty that tends to increase with trophic state due to inhomogeneity inthe horizontal and vertical distribution of phytoplankton. It is expectedthat this error alonemay result in large discrepancies between averagedsatellite and discrete in situ estimates. The uncertainty for chl-a deter-mined spectrophotometrically from the horizontal variability and stan-dard error of measurement typically ranges between 20 and 30% inSouth African inland waters (Matthews et al., 2012). The uncertaintyin chl-a estimated using the MPH algorithmmeasured by the mean ab-solute percentage error is near 60% for eukaryote-dominant waters and30% for cyanobacteria-dominantwaters (for details see Matthews et al.,2012).

Additional sources of uncertainty affecting individual pixels arisefrom bottom/shoreline effects; contamination by atmospheric aerosolssuch as smoke; errors from cloud shadows, fog and mist; and geo-location errors resulting in incorrect classification of land and waterpixels. Contamination by these sources has been reduced by cloud andaerosol flagging (already described), by extracting water pixels usingshapefiles, and by rigorous examination of the products for geo-location and other errors. MERIS scenes with more than 30% cloud orinwhich less than 5% of the lake surface areawas visiblewere excluded.Asmost SouthAfricanwater bodies are optically deep, shallowwater ef-fects are not often likely to be encountered for themajority of thewaterbodies. The precise affect that turbidwaters rich in suspended sedimentand optically shallow waters (lake bottom or submerged aquatic vege-tation) have on the MPH algorithm is still to be determined, but studiesshow that these can significantly alter the signal in the red and NIR partof the spectrum (Hill, Zimmerman, Bissett, Dierssen, & Kohler, 2014).This could be a significant source of error for some turbid water bodies.However, it is likely that for most water bodies uncertainty resultingfrom these sources has small, or negligible effects.

3. Results and discussion

The findings of the analysis from 10 years of observations from theMERIS archive are summarised in Table 1. The table contains thenames, surface areas and geographical location of each of the waterbodies. The number of MERIS scenes with less than 30% cloud coverand in which more than 5% of the lake area is visible is indicated by N,which ranges between 863 (Loskop) and 176 (Theewaterskloof). Thecoverage of the MERIS archive is poor in the south and south west ofSouth Africa resulting in fewer acquisitions for water bodies in these

regions, such as Theewaterskloof. For well covered areas, there are upto 90 scenes a year, or 8 scenes per month which have less than 30%cloud cover. This observation frequency is suitable for drawing conclu-sions on monthly or biweekly time-scales, and is therefore suitable forcapturing the seasonal (monthly) phenology of bloom initiation andshorter bloom events occurring on the scale of weeks to months.

Table 1 indicates the great variety in chl concentrations existing inSouth African water bodies, which range between 3.3 mg m−3 for theclearest oligotrophic waters (Sterkfontein) and turbid hypertrophicMthatha (478 mg m−3).

Examples of the time series products for Hartbeespoort, Loskop andSterkfontein Dams are shown in Fig. 2. These reservoirs exhibit differenttrophic states and algal bloom types: Hartbeespoort Dam is representa-tive of a hypertrophic system dominated by frequent and persistentMicrocystis aeruginosa cyanobacterial blooms and surface scum condi-tions (Matthews & Bernard, 2013); Loskop is ameso/eutrophic reservoirtypically dominated by dinoflagellate Ceratium hirundinella bloomswithoccasional cyanobacterial blooms which occur seasonally and periodi-cally (Oberholster, Myburgh, Ashton, & Botha, 2010); SterkfonteinDam is a high-altitude oligotrophic reservoir with seasonally timed in-creases in chl-a from low-biomass cyanobacterial blooms in the summer(van Ginkel et al., 2000). The seasonal periodicity of phytoplankton bio-mass is clearly observed for each of these systems which is stronglyphased with summer months. The pulsing of cyanobacterial bloomscoinciding oftenwithwarming spring and summermonths is clearly ob-served for Hartbeespoort Dam, with coincident scum events. Infrequentsummer-time cyanobacterial blooms are also observed in Loskop Dam,however most of the blooms appear to be from eukaryote species. Incontrast, Sterkfontein Dam exhibits typically very low biomass in win-ter, with pulses during summer months which are difficult to identifyas cyanobacteria due to the low biomass. Results from the MPH algo-rithm are consistent with the known trophic status and bloom occur-rence of these reservoirs, and provide new insights into their phenology.

The corresponding chl-a anomalies, monthly and yearly averages,and trends are shown in Fig. 3. The chl-a anomalies are typicallywithin one standard deviation window of the monthly averages.Anomalous events occurring outside of these windows are also occa-sionally visible, but typically occur during summer when productionis highest. An anomalously high bloom event is visible for Loskop inthe summer of 2007 to 2008. Occasional low anomalies are also visibleat Hartbeespoort Dam in the winters of 2005 and 2011. The shapes ofthe seasonal signals are typically bi-modal, with a summer peak andanother smaller peak towards autumn. The trend computed fromthe yearly averages and from the anomalies are similar showing thatthe variability at the shortest time scale (days to weeks) is consistentwith that of the longest (years). For Hartbeespoort Dam the trendsare close to −1 mg m−3 year−1. This represents a small or negligiblechange over time when considering the mean chl-a value forHartbeespoort Dam is near 90 mgm−3. A small positive trend is appar-ent for Loskop Dam, while for Sterkfontein Dam the trend is close tozero.

3.1. Status and seasonality

The overall status of the water bodies was determined by the meanvalues for chl-a, and cyanobacteria or surface scum area coverage.The highest chl values were associated with Mthatha, Ncora, Erfenis,Krugersdrift, Ntshingwayo, and Allemanskraal Dams. In situ data con-firms that this group of reservoirs exhibits very high turbidity likelyresulting from mineral components, and likely represent a specialcase. The performance of the MPH in high mineral conditions has notyet been assessed and is subject to potential errors arising from highmineral scattering affecting red/NIR signals. Results from these reser-voirs should be treated with some caution before more validation dataand modelling studies are undertaken. Aside from these reservoirs,the highest chl values were near 150 mg m−3 for Grassridge, Klipvoor

Fig. 2. Time series of chl-a, cyanobacteria coverage and surface scum coverage for A) Hartbeespoort, B) Loskop and C) Sterkfontein Dams. Note differences in y-axis scales and x-axis years.

166 M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

and Bloemhof Dams. The lowest values were from Sterkfontein,Pongolapoort, and Midmar Dams (chl less than 5 mg m−3). The chlvalueswere used to classify thewater bodies according to trophic status(Fig. 4A). An additional class accounting for highly turbid water bodieswas included for chl N 300 mg m−3. The results indicate that of the 50water bodies observed, 14% were oligotrophic, 8% were mesotrophic,6% were eutrophic, and 62% were hypertrophic of which 14% hadchl N 300 mg m−3. Therefore the majority of South Africa's largestwater bodies were hypertrophic between 2002 and 2012.

Fig. 4C shows that 27 (54%) of the water bodies had less than 10%cyanobacteria coverage over the time series. These water bodies hadinfrequent to occasional cyanobacterial blooms with insignificant orlittle area coverage. Examples of these are Loskop and SterkfonteinDams seen in Fig. 2. Eighteen water bodies had between 10 and 30%average cyanobacteria coverage. These water bodies had common andregular cyanobacterial blooms with medium coverage and includeSouth Africa's two largest reservoirs, Gariep and Vaal. Spitskop,Hartbeespoort, Koppies, Barberspan and Lake Chrissiesmeer had greater

Fig. 3.Upper panels: monthly chl-a averages (red dots)with standard deviation envelope (thin dashed line) showing the seasonal signal of chl-a. Lower panels: chl-a anomalies (red line)and yearly averages (black dots)with fitted trend lines (solid and dashed, respectively). Shown for A)Hartbeespoort, B) Loskop and C) Sterkfontein Dams. Note differences in y-axis scales.

167M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

Fig. 4.Histograms for A) trophic status class, B) phase of chl-a seasonal cycle, C) cyanobacteria area coverage, D) cyanobacteria phase, E) scum are coverage and F) scumphase of seasonalcycle, for 50 South African water bodies. Oligo = oligotrophic, Meso = mesotrophic, Eutro = eutrophic, Hyper = hypertrophic.

168 M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

than 30% cyanobacteria area coverage which indicated frequent, persis-tent and extensive cyanobacterial blooms. This list includes two of SouthAfrica's largest natural lakes, Barberspan and Lake Chrissiesmeer, ofwhich Barberspan had the most extensive cyanobacteria coverage.

Cyanobacterial surface scum conditions existed in at least 26 (or52%) of the 50 water bodies (Fig. 4E). The majority (88%) of these hadless than 1% mean areal scum coverage which means that surfacescum events were infrequent. Spitskop and Darlington Dams had be-tween 1 and 5% cover which is indicative of occasional scum eventswhich covered an intermediate or small area. Hartbeespoort Dam hadan average scum coverage of 10.2% (see Fig. 2) which indicates regularand extensive scum events, which is widely attested by other studies

(e.g. Matthews & Bernard, 2013). It should be noted that the thresholdfor scum was for chl-a values larger than 350 mg m−3. However, thisexcludes many high risk cyanobacterial bloom scenarios which occurfor chl-a values larger than 50 mg m−3 (Chorus et al., 2000). Thescum coverage estimates from these data should therefore be treatedas serious events which posed a severe health threat from densecyanobacterial surface accumulations (Chorus et al., 2000).

The seasonal cycles of chl-a for each of the water bodies are plottedtogether in Fig. 5 grouped by various chl thresholds to facilitate compar-ison. The shapeswere typically sinusoidal although frequently containedmore than one peak at various times of the year, including in exceptionalcases, winter (June to August). The seasonal amplitude represents the

Fig. 5. Seasonal cycle of chl-a for 50 South Africanwater bodies. The data are grouped in different panels by thresholds formean chl-a to aid visibility. Reservoir names are abbreviatedwithfirst three letters except for Vaal = Va1, Vaalkop = Va2, Spioenkop = Sp1 and Spitskop = Sp2. Note differing y-axis scales.

169M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

range of variability present in any reservoir throughout the year. Oligo-trophic reservoirs such as Sterkfontein, Pongolapoort, Sibayi andMidmarhad small seasonal signals less than 10mgm−3 and little year round var-iability. The largest seasonal amplitudes were associated with the groupof highly turbid reservoirs previouslymentioned. Aside from these reser-voirs, large seasonal variation of magnitude around 250 mg m−3 wereapparent for Grassridge, Gariep and Koppies Dams. These had verystrong seasonal signals and displayed the greatest variability in produc-tivity throughout the year. The phase (timing) of the seasonal cycle wasstrongly biased to summer months (see Fig. 4B) January and Februarywere the most productive months: 64% of the maxima were in summer,22% were in autumn, 10% were in spring, and only 4% were duringwinter.

Cyanobacteria coverage was less seasonally biased than chl-a(Figs. 4D and 6). The majority of cyanobacterial blooms appeared toreach their maximum coverage during autumn with the peak occur-rence during April and then unexpectedly in June. On close inspectionhowever, the data does not imply that cyanobacteria were most prolificin South Africanwater bodies duringwinter. Rather it is indicative of theseasonal cycle of maximum coverage for individual water bodies. Forexample most of the water bodies with winter maxima coverage hadsmall Acy magnitudes (six are close to 0%), while comparatively the

average value of Acy for water bodies with autumn or summer maximawas larger than 10%. The cumulative % coverage for autumn wasalmost 5 times that for winter. Thus while several water bodies hadwinter maxima, these are often insignificant (refer to Table 1). Theonly reservoir with a noticeably peaked winter maximum was LakeChrissiesmeer. Removing water bodies with insignificant coverage (14water bodies with less than 1% cover) 50% (18) had autumn maxima;22% had summermaxima; 17% hadwintermaxima; and 11% had springmaxima. Therefore cyanobacteria were most widespread during theend of the summer season in themonths ofMarch and April. The shapesof the seasonal cycles for water bodies with greater than 5% Acy (Fig. 6)varied from distinct sinusoidal curves (e.g. Hartbeespoort Dam andLake Chrissiesmeer) to bimodal curves with two maximums at dif-ferent times of the year (e.g. Barberspan and Theewaterskloof). The oc-currence of winter maxima cover is a surprising result, but wintercyanobacterial blooms have been recorded for some water bodiessuch as Midmar (Oberholster & Botha, 2007). What might be drivingwinter cyanobacterial blooms is mostly uncertain, however the re-sults here suggest that it might be a fairly common occurrence inSouth African water bodies. Therefore temperature alone does notseem to control cyanobacterial blooms in some South African waterbodies.

Fig. 6. Seasonal cycle of cyanobacteria area coverage for South African water bodies with Acy N5%. The data are grouped in different panels by season of the phase. Reservoir names areabbreviated as in Fig. 5. Note differing y-axis scales.

170 M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

As expected, the seasonal timing of surface scumwas strongly biasedtowards summer and autumnmonths with themaximum coverage oc-curring during late summer (March to April) (Fig. 4F).

There was a significant correlation between chl and the amplitude ofthe seasonal cycle Achl (r2=0.8) (Fig. 7). As expected, water bodieswiththe lowest chl values had the least seasonal variability or amplitude andvice versa. Various clusterings of thewater bodies are visible in Fig. 7. To-wards the origin, a group of oligotrophicwater bodieswith the least sea-sonal variation is visible with chlb50 mg m−3 and Achl b 100 mg m−3,including Pongolapoort, Sterkfontein, Midmar and Lake Msingazi. Asecond cluster is apparent with chl between 50 and 150 mg m−3 andAchl between 100 and 200 mg m−3. Loskop, Vaal, Hartbeespoort andLake Chrissiesmeer belong to this grouping of eutrophic water bodies,which have greater seasonal variability. A third cluster is apparentfor chl around 150 mg m−3 and Achl between 200 and 300 mg m−3.Gariep, South Africa's largest reservoir, Koppies and Bloemhof belongto this cluster. These water bodies have more extreme seasonal varia-tions. A separate cluster then exists for extremely hypertrophicreservoirswithchlN200 mg m−3 and Achl N 400mgm−3. These reser-voirs are those identified previously as outliers with high suspendedsediment concentrations. If the data are accurate then these reservoirsare most severely affected by hypereutrophication and subject to ex-treme seasonal variability. It is not clear whether there is any ecologicalsignificance to the apparent clustering. In any case, it is apparent that

various water bodies have similar seasonal responses, and possess sim-ilar production characteristics.

3.2. Trends

There was a significant correlation between the trend of chl-a yearlyaverages and the trend calculated from chl-a anomalies (Fig. 8, Table 1).This indicates the robustness of the trend calculations, since the lowestfrequency of observation (yearly) is consistent with the highestfrequency (anomalies). Only significant trends may be used to drawconclusions regarding whether eutrophication is worsening or improv-ing in a reservoir over the time period. There were 20 negative and ninepositive significant anomaly trends. For significant annual trends therewere three positive and 15 negative trends. There were 18water bodieswhere both the anomaly and annual trendswere significant, three beingpositive and 15 being negative. Based on the number of negative trendsit would seem that within the 50 water bodies eutrophication becameless severe between the years 2005 and 2011. The trend estimates,while robust, are limited by the short time interval (only seven yearsof data) and are potentially affected by anomalous bloom eventsand outliers, e.g. from water level fluctuations. These trend datashould therefore not be used to draw conclusions for managementwithout close examination of the data to ascertain the degree towhich they are affected by outliers or anomalous events/seasons. Such

Fig. 7. The mean versus the amplitude of the seasonal cycle for chl-a for 50 South Africanwater bodies. Data are plotted using abbreviations of reservoir names. The insets showzoomed areas towards the origin and for large outliers. Reservoir names are abbreviatedas in Fig. 5.

Fig. 8. Trend coefficients of chl-a between 2005 and 2011 for yearly averages and anomalies fowhile significant negative anomalies trends are in blue. Italic data indicates significant yearly trare zoomed near the origin to ensure all data is visible (note scales). Reservoir names are abbr

171M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

an examination is not feasible here given the number of water bodiesand difficulty of displaying in-depth results in short publication form.

The trendsmust be considered in view of themean and seasonal var-iability (amplitude) of chl-a in each of the water bodies to contextualisetheir significance and severity. Normalising the trend by the amplitudeof the seasonal cycle gives the percentage relative to the expected sea-sonal variability at which eutrophication is worsening or improving.Plotted against the mean chl-a value enables comparison with otherwater bodies with similar trophic status (Fig. 9). An annual increase inchl-a near 10% relative to the expected seasonal variability was foundfor Heyshope, Vanderkloof, Xonxa, Rustrontein and Darlington Dams.These reservoirs seem to have experienced the most rapid increase inchl-a during the time period and likely require themost urgentmanage-ment attention. Lake Msingazi and Heyshope Dam are potentiallyvulnerable to a regime shift from oligo/mesotrophic to eutrophic condi-tions. Water bodies that had the greatest improvement between 2005and 2011 include Kuhlange, Kwena, Fairview and Voëlvlei which hadan annual decrease in chl-a of nearly 20% of the seasonal signal. Thehigh number of significant negative trends could be an indication ofreduced nutrient input to several water bodies.

Examining the significant trends of cyanobacteria areal coverage be-tween 2007 and 2011 reveals that there were 9 water bodies whichshowed a decrease in coverage, and 7 which showed an increase. Thesetrends are independent of those determined for chl-a. The largest posi-tive trends of 2.3 and 2.1% y−1 were found for Vaal and Ntshingwayo.The largest negative trend was for Barberspan (−6.3% y−1), which alsohad the highest average cyanobacteria areal coverage. The sum of thesignificant negative trends was −16.4% y−1 while that for positive

r 50 South African water bodies. Significant positive anomalies trends are indicated in redends. The font size indicates the value of the correlation coefficient, r. The insets A, B and Ceviated as in Fig. 5.

Fig. 9. Chl versus the anomalies trend normalised by the amplitude of chl-a for 50 SouthAfrican water bodies. Significant positive trends are red while significant negative trendsare blue. Italic data indicates significant yearly trends. The font size indicates the valueof the correlation coefficient, r. Reservoir names are abbreviated as in Fig. 5.

Fig. 10. Mean area coverage of cyanobacteria versus the trend normalised by the ampli-tude of the seasonal cycle for 50 South African water bodies. Significant positive trendsare red while significant negative trends are blue (italics). The font size indicates thevalue of the correlation coefficient, r. Reservoir names are abbreviated as in Fig. 5.

ig. 11. Surface scum area coverage mean versus the trend for 50 South African waterodies. Significant positive trends are red while significant negative trends are bluetalics). The font size indicates the value of the correlation coefficient, r. Reservoirames are abbreviated as in Fig. 5.

172 M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

trends was 6.1% y−1. The trends for most of the water bodies (68%)were insignificant, indicating no significant changes in cyanobacterialblooms over the time period. The trend normalised by the seasonalamplitude indicates that Loskop, Vaal and Klipfontein Dams had themost significant increase in coverage relative to the seasonal variability,while Goedertrou and Kwena Dams experienced the highest relativedecrease in coverage (Fig. 10). The significant increase in coverage inLoskop is verified by recent studies showing increasing incidence ofM. aeruginosa blooms (Dabrowski, Oberholster, Dabrowski, Brasseur, &Gieskes, 2013).

For surface scum, four water bodies, namely Loskop, LakeChrissiesmeer, Vaal and Roodekoppies, had significant positive trends,while only Vanderkloof had a significant negative trend (Fig. 11). Over-all, positive trendswere found for 12water bodies, while 8 had negativetrends. The highest significant change was found for Roodekoppies,which had an increased scum coverage of only 0.1% per year. Therefore,there is some evidence that surface scum events became slightly moreextensive, although perhaps these changes are small enough not to besignificant.

3.3. Comparison with in situ measurements

Therewas a significant positive correlation between satellite derivedchl-a estimates and those measured in situ (Figs. 12, 13, Table 2).The highest correlation coefficients (significant at 95% confidencelevel) were determined for Bronkhorstspruit (r = 0.72), Albert Falls(r = 0.64), Vaalkop (r = 0.54) and Hartbeespoort Dams (r = 0.56).The correlation for the combined dataset was 0.64 (p = 0.00,RMSE = 55.7, see Fig. 13). The slope of 0.97, intercept of 17.6 mg m−3

and bias of 16.6mgm−3 indicate that theMPH algorithm had a tenden-cy to overestimate chl-a relative to in situ measurements. For non-cyanobacteria dominant (eukaryote) waters the fit was r = 0.72(p = 0.00, N = 174, RMSE = 53.4 mg m−3). This is comparable tothe original derivation of the MPH algorithm in Matthews et al.(2012) which had an r value of 0.84. The algorithm's tendency to

overestimate for eukaryotes was slightly higher with m = 1.27, andbias of 19.9 mg m−3. The small sample size (N = 8) fromcyanobacteria-dominant waters make statistical calculations less reli-able for these data, and likely explain the poor correlation. Further

Fb(in

Fig. 12. Scatter plots of in situ versus satellite estimated chl-a for sixwater bodies. Solid line shows linear regression fits, while dotted line is 1:1 fit. The error bars are a 30% estimated errorfor in situ measurements and the calculated standard deviation of the satellite pixels. Matchups were limited to 350 mg m−3. Note differences in x and y-axis limits.

173M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

data should be obtained for analysis of algorithm performance incyanobacteria-dominant waters.

The poorer r values for Inanda and Midmar show the difficulty ofproviding accurate quantitative estimates of chl-a in oligotrophicwater bodies from a TOA fluorescence signal. This could result from awide range of possible errors including atmospheric variability, straylight influence in small water targets, possible variation of fluorescencequantum yield at 681 nm caused by species composition, seasonal

Fig. 13. Scatter plot of in situ versus satellite estimated chl-a. Dotted line is 1:1 fit. Chl-aconcentrations greater than 350 mg m−3 were excluded to avoid matchup errors due tolarge horizontal gradients in extremely high biomass conditions.

changes or other physiological factors. Therefore quantitative detectionof chl-a at low biomass (b20 mg m−3) was the most challenging case.

The following conclusions regarding algorithm performance can bedrawn from Fig. 13: the largest uncertainty was for chl-a values lessthan 20mgm−3 with considerable scatter in the data; for chl-a greaterthan 20 mg m−3 performance was improved, however there was aslight tendency to overestimate chl-a; for chl-a less than 10 mg m−3

there appeared to be a tendency to underestimate chl-a. These conclu-sions illustrate the need to train the MPH algorithm using a high-quality matchup in situ dataset.

The average statistics from the in situ and satellite derived chl-a timeseries shows how the MPH algorithm was suitable for providing grosstrophic status estimates through chl-a (Table 2). This was true for hy-pertrophic waters such as Hartbeespoort Dam (mean in situ estimatewas 57.3 mg m−3 versus 44.1 mg m−3 from satellite) and oligotrophicwaters such as Midmar (mean in situ estimate was 6.0 mg m−3 versus4.2 mg m−3 estimated from satellite). In many cases the median andstandard deviations were also close, indicating that the range of vari-ability of the satellite estimates was consistent with in situ measure-ments. The satellite estimate tended to underestimate chl-a for someoligotrophic waters such as Albert Falls. However, this could have re-sulted from a mismatch in sample location or time. For the combineddataset the median estimates were identical. It may be that the under-estimates for oligotrophic waters and overestimates for high biomasswaters cancelled out to provide a close comparison. The similarity ofthe statistical results confirms that theMPH provided reliable estimatesof chl-a over a wide trophic range, and had only slightly higher errorranges (standard deviations) than in situ measurements, confirmingthe robustness of the approach.

3.4. Synthesis

The mean and trends determined for chl-a, cyanobacteria and scumarea coverage are plotted on the map of South Africa (Fig. 14). Somespatial patterns are clearly visible from the maps. The clustering of the

Table 2Average statistics for time series of in situ and satellite derived chl-a estimates, and results of linear regression analysis for matchups. Themean, median and std. dev of satellite estimatesare presented in brackets. Slope = m, intercept = c, p-value = p. Units for RMSE and bias are mg m−3, MAPE is %.

Reservoir Mean Med. St. dev. r m c n p RMSE Bias MAPE

Albert Falls 12.2 (3.8) 11.8 (1.8) 7.4 (5.0) 0.64 0.43 −1.47 18 0.004 10.7 −8.4 66Bronkhorstspruit 59.1 (88.1) 53.1 (59.9) 40.3 (80.8) 0.72 1.44 2.83 42 0.000 67.2 28.9 82Hartbeespoort 57.3 (44.1) 23.3 (27.7) 62.2 (46.8) 0.56 0.42 20.00 29 0.002 56.6 −13.2 81Inanda 8.5 (14.0) 8.3 (8.3) 6.2 (24.8) 0.38 1.53 0.93 23 0.072 24.9 5.5 134Klipvoor 59.4 (130.2) 38.3 (114.2) 49.6 (64.1) 0.47 0.60 94.37 30 0.009 96.1 70.8 208Midmar 6.0 (4.2) 4.5 (1.9) 3.7 (4.5) −0.08 −0.09 4.76 24 0.722 6.6 −1.7 103Vaalkop 33.3 (40.9) 16.0 (40.0) 32.5 (30.9) 0.54 0.51 23.90 16 0.032 33.6 7.6 123Cyanobacteria 123.5 (67.6) 106.1 (56.1) 68.4 (40.1) 0.21 0.12 52.62 8 0.622 105.0 −55.9 47Eukaryote 34.6 (54.5) 16.7 (16.0) 39.8 (69.8) 0.72 1.27 10.54 174 0.000 53.4 19.9 117All 38.5 (55.1) 18.0 (18.0) 45.3 (68.8) 0.64 0.97 17.59 182 0.000 55.7 16.6 114

174 M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

water bodies towards the eastern half of South Africa is representativeof the summer rainfall pattern which is heavily bias to the east, whilethe western half of the country is mostly arid. Most of the oligo/mesotrophic water bodies with correspondingly little cyanobacteriacoverage are clustered towards the east. This is consistent with thesteep gradients and high rainfall associated with the mountainousDrakensberg escarpment region, and coastal lakes. Water bodies in thecentral region typically have the highest cyanobacteria coverage andchl-a values, and higher incidence of surface scum. The high altitudeplateau region (or Highveld) is dominated by agricultural uses andcropland consisting of the catchments for the Vaal and Orange riverswhich flow towards the west. The data indicates that this region ismost impacted and vulnerable to eutrophication and cyanobacterialblooms. The south western region of South Africa has a winter rainfallpattern and only a few large water bodies which have significantcyanobacteria coverage. Scum events appear to most severely affect agrouping of water bodies including Hartbeespoort, Klipvoor andRoodekoppies Dams in the central northern region. These results seemto corroborate with previous attempts to examine the spatial patternsof eutrophication and cyanobacteria in South Africa (van Ginkel et al.,2000) as well as the geographical incidence of animal poisonings fromcyanotixins (Oberholster, Botha, & Myburgh, 2009). A detailed discus-sion of the impact on specific catchments is beyond the scope of thepresent publication.

The same data are plotted on each of the axes in three dimensionalFig. 15. The clustering enables the identification of various types ofwater bodies and extreme cases. Hartbeespoort Dam is an obviousextreme case with very high incidence of cyanobacteria and surfacescum coverage as previously discussed. The group of highly turbidreservoirs identified as outliers earlier, including Mthatha and Ncora,are closely grouped together (coloured red) and typically have interme-diate cyanobacteria and surface scum coverage. The hypertrophic waterbodies in the 100 to 200mgm−3 chl-a range (coloured orange) includ-ing Koppies, Lake Chrissiesmeer and Bronkhorstspruit, have the mostvariable cyanobacteria and scum composition as seen by considerablespread on the plot. Further towards the origin, the meso/eutrophicwater bodies (coloured green) have noticeably lower cyanobacteriaand scum coverage. Barberspan is a stark exception in this groupwhich has very high cyanobacteria cover, but lower incidence of scumthan Hartbeespoort Dam. Flag Boshielo and Theewaterskloof Damsalso appear to be outliers to this group, with higher than average cover-age of cyanobacterial blooms. Closer to the origin, the oligotrophicwater bodies expectedly have the least cyanobacteria and scum cover-age (coloured blue), with the exception of Khulange.

Excluding the extreme cases and outliers identified above, there is asignificant positive correlation between chl-a and cyanobacteria cover-age (r= 0.68, n = 38) (Fig. 16). This seems to confirm that eutrophica-tionmeasured by chl-a or phytoplankton biomass is positively correlatedwith the incidence of cyanobacterial blooms in South African waterbodies. Therefore, the higher the average chl-a concentration, the higher

the likelihood of severe and extensive cyanobacterial blooms. This couldbe thefirst comprehensive evidence of this relationship verified from sat-ellite data for a sample of water bodies.

4. Conclusions

This study has presented a time series of chl-a, cyanobacterialblooms and surface scums for the 50 largest South African water bodiesbetween 2002 and 2012 using the MERIS archive and the MPH algo-rithm. This information can be used to prioritise management and mit-igation strategies, in order to reduce health risks. The main findings ofthis research are as follows:

• 62% of South Africa's 50 largest water bodies were hypertrophic withmean chl-a in excess of 30 mg m−3.

• Cyanobacterial blooms occurred in all 50 of thewater bodies, with ex-tensive coverage in 10% of the water bodies.

• Cyanobacterial blooms were strongly seasonal reaching maximumcoverage during autumn (March–April).

• Cyanobacterial surface scum conditions which pose a high health riskexisted in 26 (or 54%) of the water bodies over the time period.

• Water bodies affected most severely by cyanobacterial surfacescum with intermediate to extensive coverage were Hartbeespoort,Darlington and Spitskop. The relevant managing authorities shouldtake immediate measures to reduce the risk of exposure to surfacescums at these water bodies.

• Surface scums reached their maximum extent in themonths ofMarchand April.

The seasonality of chl-a production peaked during the summermonths of January and February. The trend in chl-a suggests that overalleutrophication became less severe in the 50water bodies between 2005and 2011. The detection of several water bodies possessing wintercyanobacteria coverage maxima suggests that this is a phenomenonoccurring in South African water bodies, and could be related to warmwinter temperatures or unknown causes. The overall trend in cyano-bacteria coverage is uncertain, with an approximately equal numberof significant positive and negative trends. Therefore the change incyanobacteria occurrence in South African water bodies remainsuncertain.

Validation of chl-a estimates using an independently acquired in situdataset showed that theMPH algorithm provided accurate estimates ofgross trophic status in both cyanobacteria and eukaryote-dominantwaters through chl-a. The correlation was significant for individualwater bodies but was weakened by a lack of information on the preciselocation and time of sampling for the in situ dataset. The overall resultsshowed that oligotrophic andmesotrophic waters (chl-a b 20mgm−3)are themost challenging, requiring further work to elucidate and quan-tify the possible sources of error for estimating chl-a using theMPH ap-proach. It also indicated the tendency of the MPH to overestimate chl-a

Fig. 14. Status and trends of eutrophication and cyanobacteria in 50 South African water bodies. A) The mean chl-a concentration, B) the trend coefficient of chl-a anomalies, C) meancyanobacteria areal coverage, D) trend of cyanobacteria coverage, E) mean cyanobacterial scum areal coverage, F) trend of scum coverage. All trends are significant at the 95% confidenceinterval of Student t-test.

175M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

establishing the need for further calibration of the algorithm coefficients(see Matthews et al., 2012).

The study has demonstrated both the power and efficiency ofsatellite-based water quality monitoring performed on a sub-continental scale. Time series of satellite-based chl-a estimates such asthose derived here may be used in ecosystem models in order to

establish the role played by lakes in global biogeochemical cycles. Themethods used in this study can also contribute to new avenues ofresearch into catchment-scale climatological drivers of eutrophica-tion and cyanobacterial blooms, as well as to models for predictingphytoplankton biomass and surface scum formation from meteoro-logical and hydrological variables (e.g. Soranno, 1997). It has also

Fig. 15. Three dimensional plot showing chl,Acy, andAsc for 50 South Africanwater bodies. The insets on the right hand side are zoomed in towards the origin for visibility. The colour scaleshows the values for chl. Reservoir names are abbreviated as in Fig. 5.

176 M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

demonstrated how satellite-based monitoring systems can play a vitalrole in supplementing traditional in situ monitoring networks to fill in-formation gaps, especially in the developing world.

Eutrophication and cyanobacterial blooms remain issues of criticalconcern for water quality in South Africa, and cyanobacterial surface

Fig. 16. Chl versusAcy for 50 South African water bodies showing best fit linear regressionline (r= 0.68). Extreme and special cases excluded from the regression analysis are plot-ted in red. The insets are zoomed in towards the origin. Reservoir names are abbreviated asin Fig. 5.

scums present in many of the water bodies constitute a health risk re-quiring ongoing attention.

Acknowledgements

I would like to gratefully acknowledge Andy Rabagliati for imple-mentation of the MERIS FR database and its integration with the MPHPython code. Credit is also due to Julie Deshayes who assisted with themethods for time series analysis, and my PhD supervisor StewartBernard. I am also grateful to the South African Department of WaterAffairs Hydrological Services for shapefiles, and Resource QualityServices, Erica Erasmus and Triana Louw for providing the in situ chlo-rophyll dataset. Dr. Michael Silberbauer is thanked for reviewing anearly draft, and the three anonymous reviewers are gratefully acknowl-edged. The University of Cape Town and the Council for Scientific andIndustrial Research funded my PhD research. Brockmann Consult andthe European Space Agency provided the MERIS FR dataset.

References

ACRI (2006). MERIS level 2 detailed processing model. Tech. rep.Bouvet, M., & Ramoino, F. (2009). Equalization of MERIS L1b products from the 2nd

reprocessing. Tech. rep. European Space Agency.Carvalho, L., Poikane, S., Lyche Solheim, A., Phillips, G., Borics, G., Catalan, J., et al. (2012,

Nov.). Strength and uncertainty of phytoplankton metrics for assessing eutrophica-tion impacts in lakes. Hydrobiologia, 704(1), 127–140 (URL http://link.springer.com/10.1007/s10750-012-1344-1).

Chorus, I., Falconer, I. R., Salas, H. J., & Bartram, J. (2000). Health risks caused by freshwatercyanobacteria in recreational waters. Journal of Toxicology and Environmental HealthPart B(3), 323–347 (URL http://www.ncbi.nlm.nih.gov/pubmed/11055209).

Crétaux, J. -F., Jelinski, W., Calmant, S., Kouraev, A., Vuglinski, V., Bergé-Nguyen, M., et al.(2011, May). SOLS: A lake database to monitor in the Near Real Time water level andstorage variations from remote sensing data. Advances in Space Research, 47(9),1497–1507 (URL http://linkinghub.elsevier.com/retrieve/pii/S0273117711000287).

Dabrowski, J. M., Oberholster, P. J., Dabrowski, J. M., Brasseur, J. L., & Gieskes, J. (2013).Chemical characteristics and limnology of Loskop Dam on the Olifants River (SouthAfrica), in light of recent fish and crocodile mortalities. Water SA, 39(5), 675–686.

DWA (2013a). Hydrological services data extracted between 1990-01-01 and 2013-01-01. Pretoria: Hydrological Services, Department of Water Affairs (http://www.dwaf.gov.za/hydrology/. URL http://www.dwaf.gov.za/hydrology/).

177M.W. Matthews / Remote Sensing of Environment 155 (2014) 161–177

DWA (2013b). National Water Management System data extracted between 2002-01-01and 2012-04-28. Pretoria: Department of Water Affairs (http://www.dwaf.gov.za/iwqs/report.aspx).

Harding,W. R., & Paxton, B. R. (2001). Cyanobacteria in South Africa: A review (TT 153/01).Tech. rep. Water Research Commission.

Hart, R. C., & Wragg, P. D. (2009). Recent blooms of the dinoflagellate Ceratium in AlbertFalls Dam (KZN): History, causes, spatial features and impacts on a reservoir ecosys-tem and its zooplankton. Water SA, 35(4), 455–468.

Hill, V. J., Zimmerman, R. C., Bissett, W. P., Dierssen, H., & Kohler, D. D. R. (2014). Evaluat-ing light availability, seagrass biomass, and productivity using hyperspectral airborneremote sensing in Saint Josephs Bay, Florida. Estuaries and Coasts, 1–23.

Hu, C., Lee, Z., Ma, R., Yu, K., Li, D., & Shang, S. (2010, Apr.). Moderate Resolution ImagingSpectroradiometer (MODIS) observations of cyanobacteria blooms in Taihu Lake,China. Journal of Geophysical Research, 115(C4), 1–20 (URL http://www.agu.org/pubs/crossref/2010/2009JC005511.shtml).

Jupp, D. L. B., Kirk, J. T. O., & Harris, G. P. (1994). Detection, identification and mapping ofcyanobacteria: Using remote sensing to measure the optical quality of turbid inlandwaters. Australian Journal of Marine and Freshwater Research, 45(5), 801–828.

Matthews, M. W. (2011, Nov.). A current review of empirical procedures of remote sens-ing in inland and near-coastal transitional waters. International Journal of RemoteSensing, 32(21), 6855–6899 (URL http://www.tandfonline.com/doi/abs/10.1080/01431161.2010.512947).

Matthews, M., & Bernard, S. (2013). Characterizing the absorption properties for remotesensing of three small optically-diverse South African reservoirs. Remote Sensing, 5,4370–4404.

Matthews, M.W., Bernard, S., & Robertson, L. (2012, Sep.). An algorithm for detecting tro-phic status (chlorophyll-a), cyanobacterial-dominance, surface scums and floatingvegetation in inland and coastal waters. Remote Sensing of Environment, 124,637–652 (URL http://linkinghub.elsevier.com/retrieve/pii/S0034425712002350).

Matthews, M. W., Bernard, S., & Winter, K. (2010, Sep.). Remote sensing of cyanobacteria-dominant algal blooms and water quality parameters in Zeekoevlei, a small hypertro-phic lake, using MERIS. Remote Sensing of Environment, 11(49), 2070–2087 (URLhttp://linkinghub.elsevier.com/retrieve/pii/S0034425710001264).

Michalak, A. M., Anderson, E. J., Beletsky, D., Boland, S., Bosch, N. S., Bridgeman, T. B., et al.(2013, Apr.). Record-setting algal bloom in Lake Erie caused by agricultural and me-teorological trends consistent with expected future conditions. Proceedings of theNational Academy of Sciences of the United States of America, 110(16), 6448–6452(URL http://www.pubmedcentral.nih.gov/articlerender.fcgi?).

Oberholster, P. J., & Ashton, P. J. (2008). State of the nation report: An overview of the cur-rent status of water quality and eutrophication in South African rivers and reservoirs.Tech. rep. Pretoria: Republic of South Africa.

Oberholster, P. J., & Botha, A. M. (2007). Use of PCR based technologies for risk assessmentof a winter cyanobacterial bloom in Lake Midmar, South Africa. African Journal ofBiotechnology, 6(15), 1794–1805.

Oberholster, P. J., & Botha, A. M. (2010). Use of remote sensing and molecular markers todetect toxic cyanobacterial hyperscum crust: A case study on Lake Hartbeespoort,South Africa. African Journal of Biotechnology, 95(1), 8791–8799.

Oberholster, P. J., Botha, A. M., & Cloete, T. E. (2005). An overview of toxic freshwatercyanobacteria in South Africa with special reference to risk, impact and detectionby molecular marker tools. Biokemistri, 17(2), 57–71.

Oberholster, P. J., Botha, A. -M., & Myburgh, J. G. (2009). Linking climate change and pro-gressive eutrophication to incidents of clustered animal mortalities in different geo-graphical regions of South Africa. Journal of Biotechnology, 8(21), 5825–5832.

Oberholster, P. J., Myburgh, J. G., Ashton, P. J., & Botha, A. -M. (2010, Mar.). Responses ofphytoplankton upon exposure to a mixture of acid mine drainage and high levelsof nutrient pollution in Lake Loskop, South Africa. Ecotoxicology and EnvironmentalSafety, 73(3), 326–335 (URL http://www.ncbi.nlm.nih.gov/pubmed/19896711).

Odermatt, D., Gitelson, A., Brando, V. E., & Schaepman, M. (2012, Mar.). Review of constit-uent retrieval in optically deep and complex waters from satellite imagery. RemoteSensing of Environment, 118, 116–126 (URL http://linkinghub.elsevier.com/retrieve/pii/S0034425711004081).

Olmanson, L. G., Bauer, M. E., & Brezonik, P. L. (2008). A 20-year Landsat water clarity cen-sus of Minnesota's 10,000 lakes. Remote Sensing of Environment, 112(11), 4086–4097.

Paerl, H. W., & Huisman, J. (2009, Feb.). Climate change: A catalyst for global expansion ofharmful cyanobacterial blooms. Environmental Microbiology Reports, 1(1), 27–37(URL http://www.ncbi.nlm.nih.gov/pubmed/23765717).

Prigent, C., Papa, F., Aires, F., Rossow,W.B., &Matthews, E. (2007, Jun.). Global inundation dy-namics inferred frommultiple satellite observations, 1993–2000. Journal of GeophysicalResearch, 112(D12107), 1–13 (URL http://doi.wiley.com/10.1029/2006JD007847).

RQS (2004). Dams and some lakes of South Africa. Pretoria: Resource Quality Services(http://www.dwa.gov.za/iwqs/gis_data/river/rivs500k.html. URL http://www.dwa.gov.za/iwqs/gis_data/river/rivs500k.html).

Santer, R., Carrere, V., Dubuisson, P., & Roger, J. (1999). Atmospheric correction over landfor MERIS. International Journal of Remote Sensing, 20(9), 1819–1840.

Sartory, D. P., & Grobbelaar, J. U. (1984). Extraction of chlorophyll a from freshwater phy-toplankton for spectrophotometric analysis. Hydrobiologia, 114(3), 177–187.

Silberbauer, Michael (2013). Pretoria, South Africa (Personal communication).Soranno, P. A. (1997). Factors affecting the timing of surface scums and epilimnetic

blooms of blue-green algae in a eutrophic lake. Canadian Journal of Fisheries andAquatic Sciences, 54, 1965–1975.

Stumpf, R. P., Wynne, T. T., Baker, D. B., & Fahnenstiel, G. L. (2012, Jan.). Interannualvariability of cyanobacterial blooms in Lake Erie. PLoS ONE, 7(8), e42444 (URLhttp://www.pubmedcentral.nih.gov/articlerender.fcgi?artid=3409863&tool=pmcentrez&rendertype=abstract).

Van Der Walt, N. (2011). Investigation into the occurrence of the dinoflagellate, Ceratiumhirundinella in source waters and the impact thereof on drinking water purification.(Ph.D. thesis). Potchefstroom: North-West University.

van Ginkel, C. E. (2011). Eutrophication: Present reality and future challenges for SouthAfrica. Water SA, 37(5), 693–701.

van Ginkel, C. (2012, Apr.). Algae, phytoplankton and eutrophication research and man-agement in South Africa: Past, present and future. African Journal of Aquatic Science,37(1), 17–25 (URL http://www.tandfonline.com/doi/abs/10.2989/16085914.2012.665432).

van Ginkel, C. E., Hohls, B. C., & Vermaak, E. (2001). A Ceratium hirundinella (O.F. Müller)bloom in Hartbeespoort Dam, South Africa. Water SA, 27(2), 269–276.

vanGinkel, C. E., Silberbauer,M. J., &Vermaak, E. (2000). The seasonal and spatial distributionof cyanobacteria in South African surface waters. Verhandlungen der InternationalischenVereinigung für Theoretische und Angewandte Limnologie, 27, 871–878.

Williamson, C. E., Saros, J. E., Vincent, W. F., & Smol, J. P. (2009). Lakes and reservoirs assentinels, integrators, and regulators of climate change. Limnology and Oceanography,54(6, part 2), 2273–2282.