Hyperspectral remote sensing monitoring of cyanobacteria blooms in a large South American reservoir: high- and medium-spatial resolution satellite algorithm simulation

Additional keywords: Dolichospermum, Microcystis, Salto Grande Reservoir.

Cyanobacterial blooms pose a threat to human health, as well as having detrimental effects on the economic and environmental value of many lakes and reservoirs worldwide. Currently, there is a focus on preventing, anticipating and mitigating bloom occurrence. Therefore, permanent monitoring of waterbodies at large spatial and temporal scales is needed. Traditional methods for monitoring cyanobacterial blooms include collection of water samples, followed by microscopic identification and counts of cells, colonies or filaments. This approach is time and labour intensive, and, despite the effort and costs involved, it often fails to represent the spatial and temporal heterogeneity of cyanobacterial blooms, particularly in large ecosystems. Remote sensing has long been recognised as a tool for providing complete and synoptic geographical coverage of water quality in freshwater systems (Hadjimitsis et al. 2010 and references therein). Indeed, remote sensing data can complement monitoring networks that exist in many parts of the world (Klemas 2012).

Remote sensing is increasingly being used to assess cyanobacterial bloom coverage (Kutser 2004; Alikas et al. 2010), frequency (Kahru et al. 2007), temporal trends (Stumpf et al. 2012) and toxin distribution (Shi et al. 2015), as well as to forecast cyanobacteria distribution (Wynne et al. 2013). Remote sensing still faces challenges in assessing cyanobacterial toxicity (Stumpf et al. 2016) because cyanotoxins lack spectrally detectable characteristics (Shi et al. 2015). In addition, it remains difficult to differentiate cyanobacteria from other phytoplankton, particularly in inland waters (Hunter et al. 2008). Moreover, for accurate and quantitative monitoring of cyanobacterial blooms, there is a need for field measurements.

A wide spectrum of satellites has been used to monitor cyanobacteria, with these satellites varying in their spectral, spatial and temporal resolution, as well as in terms of access (e.g. downloaded free of charge or at a cost). For monitoring cyanobacteria in large waterbodies, moderate-resolution multispectral satellites (MRMS), such as sea-viewing wide field-of-view sensor (SeaWiFS; Cannizzaro and Carder 2006; Kahru et al. 2007), Medium Resolution Imaging Spectrometer (MERIS; e.g. Alikas et al. 2010; Stumpf et al. 2012; Lunetta et al. 2015) and moderate-resolution imaging spectroradiometer (MODIS; Kahru et al. 2007; Hu et al. 2010), have been used. Sensors such as MODIS have both a high spectral resolution, allowing monitoring of cyanobacterial accessory pigments, and a high temporal resolution (daily revisits); yet, because the spatial resolution is low, the use of MODIS in inland waterbodies is limited. The monitoring of cyanobacteria in inland water bodies requires sensors with moderate or high spatial resolution, such as Landsat, SPOT, Sentinel 2 (launched in 2015–17) and WorldView, among others.

The Landsat series (5, 7, 8) of satellites has a low spectral resolution and is not optimal for many applications over inland waters (Palmer et al. 2015). Yet, these satellites have been frequently used for remote sensing in small inland waterbodies (Östlund et al. 2001; Wang et al. 2004; Olmanson et al. 2008; Torbick et al. 2008). Moreover, the Landsat missions have created the longest continuously acquired space-based, moderate-resolution data archive that is free and open access. Indeed, the Landsat program has been acquiring images for more than 30 years (e.g. Landsat 5 since 1984), hence allowing for long-term studies. Moreover, the Landsat program will continue with the Landsat 9 mission planned for launch in December 2020. These reasons make the Landsat series a good tool for use in situations where there is low funding of research programs, as well as for analysing temporal trends. The Sentinel2A and Sentinel 2B satellites are technologies that can bridge the limitations of the current Landsat satellites. Sentinel 2A and Sentinel 2B have wavelengths suitable for monitoring inland water phytoplankton (bands on the near-infrared edge; 700–800 nm), high spatial (1 pixel = 100 m²), temporal (approximately every 3 days) and spectral resolution, and they are free and open access. However, these satellites were launched in 2015 and 2017 respectively; hence, their use is limited to recent and future assessments. The SPOT series and WorldView satellites are very high-spatial resolution satellites that can provide detailed information, yet access to their images can be expensive for frequent monitoring programs.

Spectral remote sensing reflectance is defined as the ratio of water leaving radiance to downward radiance, both measured above the water surface and as a function of wavelength. The variation in reflectance along different wavelengths constitutes a quantitative measure of the spectral properties of a particular object, and is known as its spectral signature. Spectral signatures obtained from spectrometers in the field (Yacobi et al. 1995; Randolph et al. 2008; Cicerelli et al. 2017) or airborne spectrometers (Jupp et al. 1994; Östlund et al. 2001; Kallio et al. 2003) can be measured independently from the satellite and can be used to simulate satellite spectral band signatures, as done in several previous studies (Gitelson et al. 1993; Yacobi et al. 1995; Randolph et al. 2008). Reflectance (a geophysical variable) can be linked to variables of a studied object (e.g. chlorophyll, cyanobacteria cell numbers) by finding the radiometric and statistical relationship with the data measured in the field.

Semi-empirical algorithms used for monitoring cyanobacteria with remote sensing assess chlorophyll (Chl)-a (Tebbs et al. 2013; for a review, see Dörnhöfer and Oppelt 2016), accessory pigments particular to cyanobacteria such as phycocyanin (Simis et al. 2005; Randolph et al. 2008; Li et al. 2015) and cyanobacteria cell counts (Lunetta et al. 2015). Despite the wide use of Chl-a as a proxy for the analysis of cyanobacteria, Chl-a does not provide information about the type of organism present because this pigment is universal across phytoplankton. During cyanobacteria-dominated blooms, most Chl-a can be attributed to cyanobacteria; nevertheless, Chl-a may fail to represent the concentration of cyanobacteria due to changes in pigment allocation per unit of biomass. This underscores the need to develop algorithms to quantify cyanobacteria cell numbers. Such algorithms are limited to sensors with higher spectral but low spatial resolution (MERIS; Lunetta et al. 2015). Developing algorithms with cyanobacteria cell numbers in sensors with higher spatial resolution is of particular relevance because they can be used as a tool to assess risk thresholds posed by cyanobacteria, such as those established by the World Health Organization (WHO) (Chorus and Bartram 1999).

Using remote sensing imagery in cyanobacteria-monitoring programs in inland waters should ensure good and quick assessment of cyanobacterial biomass, have a high spatial and temporal cover, be easy to use (size of files) and be cost-effective. Although semi-empirical algorithms developed in previous studies can be a good initial approach for monitoring cyanobacterial blooms in new environments, there is a need to obtain local data in order to develop algorithms that reflect the optical properties of the waterbody being assessed. Obtaining hyperspectral signatures in the field is a core input for simulating satellite data in different sensors. Furthermore, hyperspectral signatures obtained in the field enable optical models and semi-empirical algorithms to be developed that can be then simulated for a wide range of satellites.

The aim of this study was to develop a program for cyanobacteria monitoring using remote sensing tools in a large South American reservoir (750 km²) that experiences recurrent massive blooms. Cyanobacterial blooms in this waterbody can vary markedly in space and time in both open waters and in the beach areas. The latter are the places where the densest cyanobacterial blooms occur (O’Farrell et al. 2012; Bordet et al. 2017) and the places that are visited by thousands of tourists every summer. The maximum width of the reservoir at the beach areas is ~2 km; hence, for monitoring cyanobacteria we selected satellites with moderate and high spatial resolution, including Landsat 7 Enhanced Thematic Mapper Plus (ETM+), Landsat 8 operational land imager (OLI) and SPOT-4/5 high resolution visual infrared (HRVIR) SPOT-6/7, WorldView 2 and Sentinel 2 (1 pixel = 4–900 m², depending on the satellite used). For each satellite we selected the best model for monitoring Chl-a under cyanobacteria-dominated conditions and cyanobacteria cell numbers. Finally, we discuss which satellite to use for monitoring cyanobacteria in the studied reservoir in terms of effectively covering the spatial and temporal heterogeneity of blooms, as well as cost.

Salto Grande is a large river-like reservoir (750 km²) located along the main channel of the Uruguay River (29.43–31.12°S, 57.06–57.55°W; Fig. 1). It has a main river-like area and several lateral arms with a dendritic morphology. The mean depth of the Salto Grande Reservoir is 6.4 m (maximum depth 35 m), mean flow ranges from 2800 to 5563 m³ s^–1, mean annual temperature is 19°C and mean annual rainfall is 1260 mm. Local winds have a predominating north-east direction, with mean monthly velocities ranging between 10 and 12 km h^–1 (Rojas and Saluso 1987). Total nitrogen ranges from 0.33 to 1.67 mg L^–1 and total phosphorus ranges from 0.01 to 0.1 mg L^–1 (Bordet et al. 2017), indicating eutrophic to hypereutrophic waters.

Fig. 1.  Sampling sites (dots) assessed during the 10 field campaigns in the Salto Grande Reservoir (Uruguay–Argentina) during the summer and autumn of 2012–16.

The Salto Grande Reservoir experiences recurrent cyanobacterial blooms (O’Farrell et al. 2012; Bordet et al. 2017) that can cover large areas. A cyanobacteria field monitoring program has been ongoing since 2003, ~20 years after the dam was constructed. Cyanobacterial blooms are more frequent in summer, spring and autumn, and less frequent in winter (Comisión Administradora del Río Uruguay 2016), and can be toxic (Bordet et al. 2017).

The reservoir is used for recreational activities, including bathing, sports and fishing. Its beaches have high recreational value and thousands of tourists visit its shores during the summer. Health risk in beach areas (the arm parts of the reservoir) is high, because the number of cyanobacterial cells and toxin (microcystin) concentrations frequently surpass the alert levels established by the WHO (Gangi 2016).

From 2012 to 2016, 10 field campaigns were undertaken in the Salto Grande Reservoir in seasons when cyanobacterial blooms frequently occur. The sampling dates were 26–27 March and 24–26 April 2012, 12–15 March, 23–25 April and 12–14 November 2014, 21–24 January, 11–13 March and 1–4 December 2015, and 14–16 and 29–31 March 2016. On each occasion, multiple sites were sampled along a length of ~30 km (Fig. 1). At each sampling point, samples were collected for laboratory analysis of Chl-a, phytoplankton (cyanobacteria and algae) composition and abundance, and turbidity. Also, at each sampling point in situ hyperspectral signatures were obtained (detailed below).

At each sampling site, subsurface water samples were collected: one fraction was kept fresh for assessment of Chl-a and total suspended solids (TSS), whereas the other fraction was fixed with 1% Lugol solution for phytoplankton quantification. The samples for Chl-a estimation were kept in the dark until processing in the laboratory, then filtered through glass microfibre filters (Whatman GF/F, GE Healthcare UK Limited, Little Chalfont, UK. Whatman is a trademark of GE Healthcare companies) and extracted with 90% acetone (Nusch 1980). Chl-a concentrations were calculated by measuring absorbance with a spectrophotometer before and after acidification (0.1 M HCl), then applying the equations of Lorenzen (1967). TSS were assessed following standard protocols (American Public Health Association 2005).

Phytoplankton (cyanobacteria and algae) were quantified and identified under an inverted microscope following Utermöhl (1958). Counting errors were estimated according to Venrick (1978), accepting a maximum of 20% error for the most frequent species. When an extremely high phytoplankton density was detected, the counting was performed using a Neubauer haemocytometer on an optical microscope after hot sodium hydroxide digestion (Reynolds and Jaworski 1978). Cyanobacteria abundance is expressed as the number of cells per millilitre, because this proxy is used for assessing risk for cyanobacteria exposure in humans (Chorus and Bartram 1999). The number of cyanobacteria cells was determined by counting the number of cells per colony or per filament in 30 independent individuals in each sample. Next, the mean number of cells per individual was used to convert individuals per millilitre to cells per millilitre.

Samples were selected in which the relative abundance of cyanobacteria dominated the phytoplankton assemblage, in which, Chl-a values were considered a proxy of cyanobacteria concentration. Whenever cyanobacteria absolute densities were >50 000 000 cells mL^–1 and Chl-a concentrations were >5000 µg L^–1, samples were discarded. In cyanobacteria-dominated situations linear regressions were run between Chl-a and: (1) TSS; and (2) cyanobacteria cell numbers.

At each sampling point in situ hyperspectral reflectance was measured as an input for simulating spectral signatures in seven satellites of medium and high spatial resolution (Table 1). Hyperspectral reflectance measurements were taken from a boat with a high-resolution hyperspectral spectrometer (Radiometer ASD Field Spec FR Serial number 6250, Malvern Panalytical Ltd, Malvern, UK). Measurements were taken in the range 350–950 nm with a spectral resolution of 1 nm and with a field of view (FOV) optical fibre of 25° pointing downward, retrieving the below-surface upward radiance at 40° from the nadir, as suggested by Mobley (1999). The tip of the optical fibre was kept at 1.5 m height above the water surface, maintaining an azimuth of 90° from the solar plane with respect to the sun (Dogliotti et al. 2014), using a 2-m-long, hand-held black pole. Interference in the light field from the boat and equipment was considered negligible due to the small diameter of the tip of the optical fibre and because the front part of the boat (wood) pointed in the direction of waves. Measurements were taken from 1000 to 1500 hours (solar zenith angles >30°), on sunny and calm days with wind velocity below 5 km h^–1 and waves not higher than 0.3 m. Cloud cover was always <20%. Total incident radiation was measured using a spectralon Labsphere (Labsphere Inc., North Sutton, NH, USA) as a reference plate. At each point, 30 consecutive scans were first performed of spectralon surface reflectance, obtaining solar radiation. Then, two water surface upwelling radiation spectra were measured, each averaging 30 consecutive scans, followed by two sky downwelling radiations, each averaging 30 consecutive scans. Finally, a median spectrum was computed for further analysis. This procedure was repeated twice with the aim of measuring the range of spectral variation at each site. The data obtained were processed using View Spec Pro (ver. 5.7, Malvern Panalytical Ltd). The radiances were used to estimate the reflectance and to obtain high-resolution spectral signatures of the water surface. Two spectral signatures that resulted in higher response in the 900 than 800 nm range were discarded, assuming interference from different components of the equipment, sampling method or environmental conditions. In addition, spectral signatures out of range because of boat drift were also discarded.

Table 1.  Characteristics of the high- and medium-resolution satellites used in this study
ETM+, Enhanced Thematic Mapper Plus; HRVIR, high resolution visual infrared; MSS, multi-spectral scanner; OLI, operational land imager

Gitelson et al. (1993) found that reflectance data could be related to constituent concentrations, with regression functions of the following form:

where Z is a normalised reflectance function that is maximally sensitive to a given constituent and least sensitive to disturbing effects (e.g. survey and irradiance conditions, equipment parameters, variations in other constituent concentrations), C is the constituent concentration and a and b are regression coefficients that usually do not vary within the system. Gitelson et al. (1993) also proposed detection algorithm (Z) development through a normalisation process that enables the specific relative contribution of the inherent optical properties of a constituent (i.e. absorption and scattering coefficients) to reflectance to be emphasised. The normalisation process is a relationship between the reflectance at a wavelength that is sensitive to constituent variation (R_active) and the reflectance at wavelengths that are insensitive to constituent variations (R_reference). Gitelson et al. (1993) proposed two normalisation approaches, the reflectance ratio (Eqn 1) and the reflectance difference ratio (Eqn 2):

Jensen (2000) also found that when using remotely sensed index, external effects such as sun angle, viewing angle and atmosphere can be minimised applying normalised difference indices.

Gitelson et al. (1993) found that in inland waters a high proportion of the upwelling irradiance in the blue region of the spectrum (430–500 nm) can be attributed to the solar-induced fluorescence of dissolved organic matter, which almost entirely masks the absorption peak of phytoplankton at 440 nm. Therefore, blue–green two-band ratio algorithms, which are popular for coastal and open ocean waters, are not appropriate for inland waters (Ruddick et al. 2001; Dall’Olmo et al. 2005). R_active usually involves red bands and red edge bands (700–730 nm), whereas R_reference usually involves green, optical and near-infrared (NIR) regions.

Moreover, Gitelson et al. (2008) developed a three-band semi-empirical model for homogeneous waterbodies to estimate Chl-a coefficient absorption at a wavelength, as follows (Eqn 3):

where Chl-a_abs is Chl-a absorbance, λ₁ is a wavelength highly sensitive to Chl-a concentration (red zone), λ₂ is a wavelength minimally sensitive to the absorption of Chl-a, tripton and coloured dissolved organic matter (CDOM) and is close to λ₁ (range 710–730 nm), λ₃ is a wavelength range unaffected by Chl-a, tripton or CDOM absorption and, instead, is related to water absorption (NIR range, wavelengths >730 nm) and α₁ is a regression coefficient.

In order to define R_reference and R_active for each of the seven high- and medium-resolution satellites (Table 1), we estimated the CV (calculated as (s.d. ÷ mean × 100)) along the spectrum. R_reference was selected as the spectral region(s) where the CV was minimal in relation to Chl-a concentration and other components (tripton, CDOM). R_active was selected as the spectral region that had a high CV relative to the Chl-a concentration. We then simulated spectral bands in the medium- and high-resolution satellites assessed in this study (Table 1) under conditions of cyanobacteria dominance. For this we resampled the hyperspectral signature by applying the specific filter to each satellite.

We applied Eqn 2 analysing CV on simulated spectral bands for each of the seven satellites assessed (Landsat 5, 7, 8, SPOT-4/5, SPOT-6/7, Sentinel 2 and WorldView 2) using two-band (Model 1) and three-band (Model 2) algorithms. In addition, we applied Eqn 2 to satellites WorldView 2 and Sentinel for bands in the edge of the red range (700–800 nm) that are absent in Landsat and SPOT satellites, including three (Models 3 and 5 for World View-2, Model 6 for Sentinel 2), four (Model 4 for World View-2) and six (Model 7 for Sentinel 2) band algorithms. Finally, we applied Eqn 3 for satellites WorldView 2 and Sentinel (Model 8).

Next we regressed Chl-a with models 1–8. For the best model for Chl-a for each satellite, we ran regressions with cyanobacteria cell numbers (Model 9). Also, for each satellite we ran a multiple regressions with the significant bands and the natural logarithm of cyanobacteria cell numbers (Model 10). Within each satellite, the best model for each variable (Chl-a, cyanobacteria abundance) was selected based on the significance of the regression (P < 0.05), the (higher) coefficient of determination and the (lower) root mean square error (RMSE).

As an example, and in order to evaluate the performance of the cyanobacteria algorithms on satellite data, we downloaded several images (Landsat, Sentinel 2 and SPOT) on dates close to the sampling campaigns. The L2 products for Landsat 5 (26 February 2011), Landsat 7 (17 March 2015) and Landsat 8 (9 March 2015) were downloaded and the cyanobacteria cell algorithm was applied. The Sentinel 2 (30 March 2016) image used was available only for a date after the field campaign ended; an atmospheric correction algorithm for VNIR-SWIR (visible and near-infrared–short-wave infrared) multi- and hyperspectral imagery was applied (Bernstein et al. 2005, 2006) and the cyanobacteria cell density algorithm was applied.

Cyanobacteria dominated in 75 of the 111 valid samples (68% of cases). The mean Chl-a concentration during cyanobacteria dominance (n = 75) was 278 µg L^–1, but ranged from very low to very high concentrations (4–4700 µg L^–1). Similarly, cyanobacteria absolute abundance ranged from almost 0 to 12 000 000 cells mL^–1, and showed high heterogeneity (mean 461 715 cells mL^–1; median 10 660 cells mL^–1). The cyanobacteria assemblage was mostly composed of the colonial species Microcystis aeruginosa and Microcystis wesenbergii, accompanied by filamentous species Dolichospermum circinale, Dolichospermum planctonicum and Raphidiopsis mediterranea. The mean concentration of TSS was 33.8 mg L^–1 (range 5–467 mg L^–1). There was a linear relationship between ln[Chl-a] and ln[TSS] (R² = 0.74, P = 0.0000, n = 75; Fig. 2a), as well as between ln[Chl-a] and ln(cyanobacteria cell density) (R² = 0.595, P = 0.0000, n = 75; Fig. 2b).

Fig. 2.  Regressions between (a) the natural logarithm of chlorophyll concentrations (ln[Chl-a]; mg L^–1) and the natural logarithm of total suspended solids (ln[TSS]; mg L^–1) and (b) ln[Chl-a] and the natural logarithm of cyanobacteria cell numbers (ln(Cyano); cells mL^–1).

Fig. 3a shows the 75 hyperspectral signatures obtained in the field. In the blue range, mean reflectance was low and the CV was ~30–40% (Fig. 3b). In the green and red range, mean reflectance was higher than in the blue region, and the CV was approximately 30%. In the red region a minimum of reflectance was observed at around 675 nm, coinciding with a peak in CV (40%) (Fig. 3b). Reflectance in the green region of the spectrum was independent of Chl-a concentrations (P > 0.05), whereas in the red region reflectance decreased with increasing Chl-a concentration (Spearman ρ near –0.5, P < 0.05). In the NIR, mean reflectance decreased, yet the maximum reflectance and CV increased markedly (~80–90%; Fig. 3b). In the NIR region, reflectance increased with Chl-a concentration (Spearman ρ near 0.5, P < 0.05).

Fig. 3.  (a) Hyperspectral signatures obtained in the field under conditions of cyanobacteria dominance (n = 75) and (b) CV and mean, maximum and minimum reflectance. Chl-a, chlorophyll-a absorption; PC, phycocyanin absorption; PE, phycoerythrin absorption; NIR, near-infrared.

Based on the CVs obtained from the hyperspectral signatures (Fig. 3), we assigned R_reference to the green part of the spectrum, because it had a low CV and hence was the region less affected by variations in water component concentrations, and R_active to the red and NIR parts of the spectrum because these parts had high CVs and hence were more affected by variations in water component concentrations (Table 2). For each of the seven satellites assessed the two- and three-band algorithms were:

Table 2.  Models explored for the Landsat 5, 7 and 8, SPOT 5 and 6 satellites for chlorophyll (Chl)-a and cyanobacteria abundance
The models in bold are those with the best fit for each variable within each satellite. Cyano, cyanobacteria abundance; ETM+, Enhanced Thematic Mapper Plus; OLI, operational land imager; RMSE, root mean square error; TM, Thematic Mapper; λ, centre of the satellite band width

where λ₅₅₀ is the centre of band width for the green range, λ₆₅₀ is the centre of band width for the red range and λ₈₀₀ is the centre of band width for the NIR range. In Models 1 and 2, because the red zone was inversely related to cyanobacteria concentrations, it was subtracted from the reference band. In Model 2, because the NIR zone was proportionally related to cyanobacteria concentrations it was added to the reference reflectance. The normalised indices for WorldView 2 (Models 3, 4, 5) and Sentinel 2 (Models 6,7) with bands in the edge of the red range (700–800 nm) resulted in:

For WorldView 2 and Sentinel 2 satellites the three-band model developed by Gitelson et al. (2008) (Eqn 3, Model 8, Table 2) was:

and

The results showed that for each of the seven satellites, all the algorithms applied were significant (Tables 2, 3). Sentinel 2 showed the highest coefficient of determination of all satellites (Model 6; R² = 0.80), followed by WorldView (Model 5; R² = 0.78; Tables 2, 3). For the Landsat (5, 7 and 8) and SPOT (5/6) satellites, the highest coefficient of determination for Chl-a was obtained with the three-band algorithm (Model 2; R² = 0.67–0.73), although in Landsat 8 both the two- and three-band models (Models 1 and 2) had the same R² (0.7; Table 2). When the best-fit indices for Chl-a were calculated for the natural logarithm of cyanobacteria cell abundance (Model 9, Tables 2, 3), the coefficient of determination either decreased slightly or remained the same, but the RMSE increased markedly (Tables 2, 3). Regarding algorithms of cyanobacteria abundance, for most satellites Model 9 showed a better fit than the multiple regression (Model 10; Tables 2, 3).

Table 3.  Models explored for WorldView 2 and Sentinel 2 satellites for chlorophyll (Chl)-a and cyanobacteria abundance
The models in bold are those with the best fit for each variable within each satellite. Cyano, cyanobacteria abundance; RMSE, root mean square error, λ, centre of the satellite band width

Fig. 4 exemplifies the best-performing cyanobacteria algorithm applied to Landsat 5, Landsat 7, Landsat 8 and Sentinel 2 images for the Mandisovy arm of the reservoir. The images show high heterogeneity in both space and time (Fig. 4). Concentrations of cyanobacteria in most arms (dendritic part, beach sites) were high, ranging between 100 000 and 1 000 000 cells mL^–1, and covered an extended area (kilometres; Fig. 4). However, in the most open water areas of the reservoir cyanobacteria densities were <5000 cells mL^–1 (Fig. 4).

Fig. 4.  Satellite images after applying the best semi-empirical algorithm for cyanobacteria abundance in: Landsat 5 (26 February 2011), Landsat 7 (17 March 2015), Landsat 8 (9 March 2015) and Sentinel 2 (30 March 2016).

The models developed here could be used as a preliminary approach for remote sensing of cyanobacterial blooms in the Salto Grande Reservoir. On most sampling occasions (~68%) phytoplankton was dominated by cyanobacteria, with blooms reaching a high density (up to 12 000 000 cells mL^–1). Blooms showed high heterogeneity, as reflected by differences between mean and median data. For example, mean cyanobacteria values were 461 715 cells mL^–1; this concentration is approximately fourfold higher than the Alert Level 2 for drinking and recreational use of water established by Chorus and Bartram (1999) of 100 000 cells mL^–1. However, median cyanobacteria abundance was 10 660 cells mL^–1; this concentration is below the guidelines established by Chorus and Bartram (1999) for water recreational use (Level 1 = 20 000 cells mL^–1), yet approximately fivefold higher than the Alert Level 1 for drinking water (2000 cells mL^–1). These results underscore the risk posed by cyanobacterial blooms in this ecosystem. Cyanobacterial blooms were primarily composed of scum-forming taxa, mostly M. aeruginosa and less frequently Dolichospermum spp., and rarely by dispersive bloom-forming taxa Raphidiopsis spp. The range of Chl-a concentrations in cyanobacteria-dominated scenarios was large (4–4700 µg L^–1), and TSS concentrations were highly and positively correlated with Chl-a concentrations. The latter suggests that the main driving factor controlling suspended matter concentration was cyanobacteria. The relationship between ln[Chl-a] (during cyanobacterial blooms) and ln(cyanobacteria cell numbers) was linear and significant (R² = 0.59), suggesting that Chl-a in cyanobacteria-dominated scenarios reflects cyanobacteria cell numbers. The dispersion between these variables was probably related to the differences in pigment concentration at similar cell densities, possibly due to differential physiological allocation of pigment in response to light availability.

The hyperspectral signatures obtained in this study were similar in shape and magnitude to reflectance spectra collected in inland waters in general, and of cyanobacteria abundance in particular (Zimba and Gitelson 2006; Randolph et al. 2008; Tebbs et al. 2013). The CV facilitated identification of the parts of the spectrum reflectance that varied. For example, blue regions experienced low reflectance values due to algal pigments and dissolved organic matter. In the green region, the CV was low, whereas in the red region the CV was higher and in the red edge and NIR regions the CV was very high. Minimum reflectance was found for cyanobacteria-dominated situations ~675 nm, coinciding with the red Chl-a absorption maximum. In the NIR (700–900 nm), reflectance was positively related to Chl-a concentrations. Indeed, the NIR region is affected by any kind of suspended particles (Yacobi et al. 1995; Dall’Olmo et al. 2005). The high reflectance probably relates to the influence of the dominant cyanobacteria species Microcystis, which is colonial and could increase scattering, and therefore NIR reflection.

For Landsat (5, 7, 8) and SPOT (4/5, 6/7) satellites, the three-band algorithm (Model 2; R_active in the red and NIR zones) was the one that performed best, highlighting the importance of the red and NIR zones in identifying cyanobacteria. Similarly, for WorldView 2 and Sentinel 2 satellites, the best performing algorithms were those using bands from the red edge. However, the algorithms developed by Gitelson et al. (2008) did not render the best fits, probably because the Salto Grande Reservoir has high heterogeneity in water optical properties, Chl-a concentrations and cyanobacteria abundance. Algorithms based on Chl-a had a high coefficient of determination among the satellites analysed, despite a large range in Chl-a concentrations (4–4700 µg L^–1). Many studies performed with algorithms using Chl-a to monitor cyanobacteria record Chl-a concentrations up to 500 µg L^–1 (for a review, see Dörnhöfer and Oppelt 2016), and few waterbodies have maximum Chl-a values within the range observed by us (e.g. German 2017).

Algorithms based on cyanobacterial cell numbers resulted in similar coefficient of determination to those estimated for Chl-a. This probably happened because the Chl-a algorithms were run in cyanobacteria-dominated scenarios. However, cyanobacterial cell abundance algorithms had a higher RMSE than Chl-a algorithms and the abundance algorithms could affect ~20% of the cell abundance estimate. Nevertheless, cell concentrations could be well allocated to the different alert level categories for cyanobacteria abundance (for drinking water, Alert 1 and Alert 2 correspond to 2000 and 100 000 cells mL^–1 respectively, whereas for the recreational use of water Guidance Level 1 is 20 000 cells mL^–1 and Guidance Level 2 is the same as Alert 2 for drinking water; Chorus and Bartram 1999). Based on the regressions, the algorithms of cyanobacterial cell numbers could detect average concentrations of 150 cells mL^–1 (for Sentinel and Landsat 8) and 1000 cells mL^–1 (for Landsat 7); these thresholds allow for the identification of vigilance levels (200 cells mL^–1; Chorus and Bartram 1999). The detection limit obtained in the present study was between one and two orders of magnitude lower than the detection threshold achieved by Lunetta et al. (2015) using MERIS because of differences in the spatial resolution of the satellites used.

The best-performing algorithm was obtained for Sentinel 2A and 2B. Its higher coefficient of determination, its high spatial (1 pixel = 400 m²), temporal (approximately every 3 days) and spectral resolution, plus the fact that its images can be downloaded free of cost make this technology the best for monitoring cyanobacteria. The launching of these satellites (2A and 2B in 2015 and 2017 respectively) occurred only during the last part of our field sampling program, and although the satellites will be certainly used for future monitoring, we cannot use them for assessing past temporal trends in the reservoir.

Regarding Landsat satellites, the complexity of inland waters, associated with the low spectral resolution of the Landsat satellites often makes it challenging to use these satellites for the accurate estimation of biomass or other pigments in inland waters. Despite their recognised limitations, Landsat satellites have been used in many studies (Wang et al. 2004; Olmanson et al. 2008; Torbick et al. 2008; Lin et al. 2018). Our results showed that algorithms for these satellites fit rather well compared with satellites with better spectral resolution. Hence, our results highlight how, with sufficient field data, it was possible to develop simple models to monitor cyanobacteria using Landsat satellites, as done by others (e.g. Ferral et al. 2018). The long-term continuity of the Landsat program could allow assessment of historical trends since 1984 (Landsat 5 launched) at high temporal resolution (weekly when combining Landsat satellites). SPOT 6/7 and WorldView 2 would be excluded from operational monitoring due to the large size of their files and the high cost incurred in accessing their images. However, they could be used for analysing particular sites of high sanitary interest (e.g. popular beaches) or when a high spatial resolution is required in events such as fish kills, or when high toxicity is detected in beach areas.

When estimating the densities of cyanobacteria through the application of algorithms to real satellite data for dates close to the sampling dates, the estimated magnitudes were in accordance with the values obtained in the field in this study. These values were also within the ranges obtained previously in the same reservoir by O’Farrell et al. (2012) and Bordet et al. (2017). The algorithms correctly reflected the magnitude (the minimum and maximum values observed for field data) and spatial heterogeneity of cyanobacteria for monitoring purposes and health management in the Salto Grande Reservoir.

Beach areas and arms in the reservoir (with high residence time and less suspended solids) had the largest blooms. Despite temporal and spatial variations in water conditions (with differences in suspended matter, cyanobacteria distribution and density), we found absence of false positives.

Continuous algorithm development and evaluation will help increase our ability to monitor and forecast the occurrence of cyanobacterial blooms. Validation of the algorithms in the Salto Grande Reservoir is challenging due to its great spatial gradient (~30 km) with differences in morphology and hydrology (e.g. arms, beaches, open water), as well as because of daily temporal variations in the magnitude of cyanobacterial blooms. Because there is a high fluctuation in cyanobacterial bloom accumulation on the surface at different times of the same day, this may result in errors for field validation driven by the natural degree of change and patchy structure of cyanobacterial blooms. Moreover, the high spatial range over which blooms occur poses an inevitable gap between the passing of the satellite and the field sampling (seconds v. hours), as field sampling can imply hours of difference between measurements. The up-scaling of discrete in situ measurements (point stations) to the spatial measurement of a sensor (depending on the spatial resolution of the satellite) is an essential problem in validating remote sensing indicators (Giardino et al. 2010). Yang et al. (2013) highlighted that the spatial and temporal mismatches between satellite observations and validation datasets must be considered.

Further improvement of the models developed here could allow better precision in quantifying cyanobacterial cell numbers. Yet, it must be highlighted that for assessing hazards and health risks posed by cyanobacteria, the thresholds used by the WHO were well identified.

The authors declare that they have no conflicts of interest.

This research was funded by the Comisión Administradora del Río Uruguay (CARU) Project numbers 133/13 and 134/13.