Satellite assessment of eutrophication hot spots and algal blooms in small and medium-sized productive reservoirs in Uruguay’s main drinking water basin

Intensive agricultural activities favor eutrophication and harmful phytoplankton blooms due to the high export of nutrients and damming of rivers. Productive watersheds used for water purification can have multiple reservoirs with phytoplankton blooms, which constitutes a high health risk. In general, water quality monitoring does not cover small- and medium-sized reservoirs (0.25–100 ha) of productive use due to their large number and location in private properties. In this work, the in situ trophic state of fourteen reservoirs was simultaneously assessed using Sentinel-2 images in the Santa Lucía River Basin, the main drinking water basin in Uruguay. These reservoirs are hypereutrophic (0.18–5.22 mg total P L−1) with high phytoplankton biomasses (2.8–4439 µg chlorophyll-a L−1), mainly cyanobacteria. Based on data generated in situ and Sentinel-2 imagery, models were fitted to estimate satellite Chl-a and transparency in all the basin reservoirs (n = 486). The best fits were obtained with the green-to-red band ratio (560 and 665 nm, R2 = 0.84) to estimate chlorophyll-a and reflectance at 833 nm (R2 = 0.73) to determine transparency. The spatial distribution of the trophic state was explored by spatial autocorrelation and hotspot analysis, and the variation in spatial patterns could be determined prior and subsequent to a maximum cyanobacteria value in water treatment plant intakes. Therefore, reservoirs with greater potential for phytoplankton biomass export were identified. This work provides the first fitted tool for satellite monitoring of numerous reservoirs and strengthens the country’s ability to respond to harmful phytoplankton blooms in its main drinking water basin.


Introduction
Eutrophication is one of the most widespread environmental problems affecting surface waters worldwide (Michalak et al. 2013;Burford et al. 2020), since it affects the trophic web, biodiversity, and water quality (Carpenter et al. 1998;Chapin et al. 2011). This excessive nutrient enrichment process has a detrimental effect not only on ecosystem conservation and recreation but also on water supply for animal and human drinking, to name a few. Indeed, this gives rise to particular concern in terms of human health and economic costs (Conley et al. 2009;Dodds et al. 2009).
Increased nutrient concentration, high water residence time, increased temperature, and low turbidity create suitable conditions for phytoplankton growth (Reynolds 2006). In lentic aquatic systems, excessive growth of phytoplankton biomass is often dominated by cyanobacteria (Burford and O'Donohue 2006). These can produce toxins detrimental to aquatic and terrestrial biota, among which the most common is microcystin (Chorus and Welker 2021), represented by several genera, most notably Microcystis (Carmichael 2001;Leflaive and Ten-Hage 2007;Pearson et al. 2016). In recent decades, the occurrence of cyanobacterial blooms has been exacerbated on a global scale in terms of frequency,  (Huisman et al. 2018;Burford et al. 2020). Traditional monitoring methods require obtaining samples from numerous water bodies and subsequent laboratory analysis, thus entailing high economic, time, and skilled efforts (Navalgund et al. 2007;Lins et al. 2017). As an alternative, numerous indicators have been developed to quantify phytoplankton pigments and water transparency based on remote sensing reflectance (Rrs) and information collected in-situ (Ha et al. 2017;Watanabe et al. 2018;Giardino et al. 2019). Regarding inland waters, there is no consensus on the use of specific indicators for estimating chlorophyll-a (Chla) concentrations (Nechad et al. 2015). In general terms, approaches to successfully estimate Chl-a use reflectance near the red edge (~ 700 nm), where Chl-a exhibits its lowest absorption coefficient (maximum Rrs), and in the red (~ 665 nm) edge where it exhibits the highest absorption coefficient (minimum Rrs) (Gitelson 1992). The Normalized Difference Chlorophyll Index (NDCI), which is based on the difference between Rrs ~ 700 nm and Rrs ~ 665 nm (Mishra and Mishra 2012), is worth noticing. Also, some research studies have shown high accuracy for estimating Chl-a based on the red (~ 665 nm) to green (~ 560 nm) reflectance ratio (Ioannou et al. 2014), both in environments with low (4.6 µg L −1 ) (Ha et al. 2017) and high (65.1 µg L −1 ) Chl-a concentrations (Oliveira et al. 2016). It should be underlined that these indicators should be adjusted with in situ information gathered from the environments to be monitored (Yang et al. 2017).
The concentration of suspended solids, Chl-a, and dissolved organic carbon determine the transparency of inland waters, which has historically been measured in situ with the Secchi disk. Likewise, several research studies have estimated transparency from surface reflectance in the visible region (Zhang et al. 2021), near-infrared region, and by band combination (Ouma et al. 2020). This diversity of approaches is based on the fact that suspended solids and Chl-a increase water brightness, and, therefore, reflectance throughout the visible/NIR spectrum (Topp et al. 2020).
Uruguay has been facing issues related to eutrophication and phytoplankton blooms in several ecosystems for decades (Bonilla et al. 2015). They are mainly associated with excess nutrient inputs from agribusiness expansion, urban, and agroindustrial initiatives, as well as reservoir construction (Conde et al. 2002;Bonilla et al. 2006;Chalar et al. 2011;Ernst and Siri-Prieto 2011;Aubriot et al. 2017). Even though great efforts have been undertaken to assess eutrophication in various water bodies distributed throughout the national territory (Bonilla et al. 2015;Haakonsson et al. 2017), the lack of information on the trophic state of small-and medium-sized reservoirs for productive purposes (irrigation, agro-industrial, and animal consumption) raises our concern.
In general, agricultural reservoirs receive significant nutrient inputs for being located in watersheds where intensive agriculture is practiced. This, along with the high residence time of water, creates favorable conditions for the development of cyanobacterial blooms (Bowling et al. 2013;Beaver et al. 2014). These environments frequently suffer overtopping due to excess rainfall, flushing the bloom downstream and leading to their potential transport to the water treatment plant intakes. This scenario was evaluated in the Santa Lucia River Basin (CRSL), where high flows (and rainfall) were associated with the presence of harmful phytoplankton in the river, typically found in lentic and eutrophic environments, with the potential to affect water purification in the main treatment plant in the country (Aguas Corrientes) (Somma et al. 2022).
A large number of reservoirs have been registered in Uruguay as of 2015 (1363 reservoirs that supply domestic, commercial, industrial, irrigation, and other agricultural uses; MVOTMA 2017). The number of reservoirs amounts to approximately 7800 if small-and medium-sized (0.25-10 ha) unregistered reservoirs are included (Achkar et al. 2016). Reservoirs are expected to increase in number in this country due to the enforcement of policies that promote their construction (Act 16.858/1997(Act 16.858/ modified in 2017. Given their large number and the fact that they are located in private properties in most cases, conducting an in situ evaluation of water quality is not always feasible. CRSL is the main source of drinking water in Uruguay, accounting for 60% of the national supply. Considering that harmful phytoplankton blooms restrict water treatment, tools are required to enhance monitoring practices at a temporal and spatial level in order to improve the current management interventions and strengthen the country's response capacity (Aubriot et al. 2020). In this context, the present work aims at fitting optical indicators for satellite estimation of Chl-a concentration and transparency in order to monitor all the agricultural reservoirs in CRSL before and after high cyanobacterial concentrations in the Planta Potabilizadora de Aguas Corrientes (PPAC) (National Administration of State Sanitary Works, OSE). This work seeks to develop continuous monitoring tools for reservoirs and generate a rapid water quality diagnosis to improve the identification of eutrophication hot spots, as well as to strengthen the country's ability to respond to harmful phytoplankton blooms in its main drinking water basin.

Methodology
The methodological strategy is divided into four sections. The first one deals with the survey of all the reservoirs (0.25-100 ha) in the CRSL upstream PPAC, reservoir selection, and obtaining the authorizations needed to carry out in situ sampling. The second delves into the acquisition of in situ information, for which three sampling campaigns were carried out in fourteen agricultural reservoirs. Physicochemical parameters, total nitrogen and phosphorus (TN and TP, respectively), phytoplankton pigments, transparency, and suspended solids were measured. Surface level Rrs was also calculated using Sentinel-2 images obtained ± 4 h apart from in situ sampling. In the third stage, spectral signatures were analyzed and linear models were developed to monitor satellite Chl-a concentration and transparency on two specific dates. Finally, the spatial distribution patterns of these variables were explored (Fig. S1).

Study area
The study area includes the basin that drains to the water treatment plant intakes in the PPAC, as well as the basins of the Santa Lucía and Santa Lucía Chico rivers and La Virgen and Canelón Grande streams (Fig. 1), covering a total area of 9131 km 2 . Various productive activities are carried out in this area. In the southern region, horticulture, fruit growing, wine production, and pig and poultry farming prevail, while in the central region, dairy farming stands out. In the east, livestock, forestry crops, and mining dominate (Achkar et al. 2012) (Fig. 1). Increasing trends in land use intensity have been identified in the basin from 2000 to 2017, mainly linked to the cultivation of cereals in monoculture, mostly soybean, but also corn, and sorghum (Gazzano et al. 2019).

Sampling
To select the reservoirs to be sampled in situ, in the first place, potential reservoirs were chosen using surrounding land use as a criterion, according to the aerial imagery provided by the Spatial Data Infrastructure Agency of Uruguay (IDEUY). The first field trip was carried out in July 2019 in order to obtain access permits to establishments under different land use types, such as pastures, planted grasslands, fruit trees, dairy farms, and dryland farming. Fourteen reservoirs were finally chosen according to accessibility (Fig. 1).
Three in situ samplings were performed in fourteen reservoirs on 07/11/2019, 02/12/2019, and 16/11/2020), and a total n of 42 samplings were obtained, made up of three replicates per reservoir (126 samples). In each reservoir, physicochemical parameters (temperature, pH, conductivity, dissolved oxygen, and turbidity) were measured using Horiba U-52G probe, and transparency with Secchi disk. Three replicates of 1 L were taken for the analysis of phosphorus and nitrogen (total), chlorophyll-a, total solids, and organic and inorganic suspended solids. Cyanobacterial Chl-a (Chl-a Cy) was determined in each replicate by in vivo fluorometry (Aquafluor Turner Designs) according to Cremella et al. (2018).

Laboratory analysis
Each water replicate was filtered with two filters (MGF, Munktell): one previously burned at 450 ºC and weighed for suspended solids analysis, and the other for Chl-a extraction. Total phosphorus (TP) and total nitrogen (TN) were measured in the unfiltered sample volume (Valderrama 1981;APHA 2005). In parallel, the concentration of suspended solids and suspended organic matter was determined by the gravimetric/calcination method (APHA 2005). The Chl-a concentration was carried out using the hot ethanol extraction method (Nusch 1980).

Satellite imagery processing
The images captured with the MSI on-board Sentinel-2 were downloaded from https:// scihub. coper nicus. eu/ dhus/, which provides free and open access to MSI/S-2 L1C imagery. Level-1C processing includes radiometric and geometric corrections including ortho-rectification of reflectance at the top-of-the atmosphere and spatial registration on a global reference system (Fletcher 2012). Prior to the acquisition of spectral information, L1C images were converted to surface reflectance estimates using ACOLITE software (v20181210.0), which performs Rayleigh scattering and aerosol correction (Ansper 2018).
Reflectance values coincident with in situ sampling were obtained from grids of 3 × 3 pixels (90 m 2 ) with a minimum distance to shore of 10 m. This procedure was performed since data may not be accurate for a single pixel (Clark et al. 2017). The average surface reflectance of each grid was extracted for all bands from the MSI/S-2 images. Image processing and extraction of spectral information were carried out in the Sentinel Application Platform (SNAP) of the European Space Agency.

Data analysis
Basic exploratory data analyses were conducted. Linear regression (log-log) analyses with total Chl-a and Chl-a Cy were used as the response variable. Differences between reservoirs were evaluated for physicochemical and pigment variables using Kruskal-Wallis.
In order to estimate satellite Chl-a and transparency, band indicators and band ratios were fitted using linear models (LM). Subsequently, the residuals generated were evaluated using Shapiro-Wilk normality test and Breusch-Pagan homoscedasticity test. First, the spectral signatures of the reservoirs sampled in situ were obtained and the surface reflectance of each Sentinel-2 band was correlated with the variables in situ. From this analysis, the most appropriate band for estimating transparency was chosen. Chl-a estimation models used the reflectance ratio red (665 nm) to red edge (705 nm) bands and green (560 nm) to red (665 nm) edge bands as predictor variables. Chl-a concentration and transparency logarithms were considered (Kayastha et al. 2022) as response variables.

Satellite monitoring
To monitor the study basin, in the first place, all the reservoirs with a surface area > 0.25 ha were identified, using the information on cutwaters and water bodies provided by the Spatial Data Infrastructure Agency of Uruguay. Since this coverage includes the entire floodable area as reservoir surface and vectorization was used to extract spectral information from the water bodies, the polygons were manually adjusted to cover only the surface area corresponding to water.
In order to diagnose the trophic state of the reservoirs at two contrasting times according to precipitation metrics and evaluate biomass transport downstream to PPAC water intakes, Sentinel-2 MSI images were analyzed on a date before and after a high precipitation event (> 20 mm daily). Once the dates were identified and the images downloaded, the accumulated precipitation for the thirty days prior to each image was calculated using data from Las Brujas weather station, National Institute of Agricultural Research, located 17 km south of Aguas Corrientes (Fig. 1). The records of cyanobacterial concentration in the Santa Lucía River taken at PPAC were also assessed ( Fig. S2).
The images were acquired at L1C level, and the surface Rrs was calculated using ACOLITE software. The reservoir vectors previously elaborated were then used to calculate the median Rrs values for each band. Based on said information, the equation resulting from the Chl-a and transparency estimation models was applied. Once Chl-a was estimated, the reservoirs were classified according to their trophic state (Cunha et al. 2013).
In order to assess Chl-a spatial distribution and transparency in all the reservoirs under study, spatial autocorrelation (SA) was used. This tool refers to the degree to which one object is similar to other nearby objects. Moran's I, a measure of spatial autocorrelation, was applied. It is a Pearson correlation coefficient with a spatial location weighting matrix (Moran 1948).
To detect patterns of Chl-a distribution and transparency among reservoirs, hot spot analyses were performed with information from satellite monitoring (Torbick et al. 2013;Coffer et al. 2021). Said analyses identify statistically significant clusters called hot or cold spots. In this study, the Getis-Ord Gi* (Getis and Ord 2010) statistic was used to determine hot spot incidence. For a feature to be recognized as a statistically significant hot spot, it must have a high value and must also be surrounded by other high-value features. The opposite applies to cold spots.
SNAP software was used to extract Rrs information and ArcMap software version 10.4.1 for the spatial autocorrelation and hot spot analyses.

Results
All the reservoirs in the basin with a surface area of 0.25-100 ha were identified and vectorized to select the reservoirs where in situ sampling was carried out. A total of 486 reservoirs were recorded, ranging in size from 0.25 to 60 ha (median 0.93 ha), with the first to third quartiles located at 0.6, 0.9, and 1.8 ha respectively. In order to simplify the subsequent analyses, the size was categorized in increasing order between 1 and 4 according to the aforementioned quartiles. Fourteen reservoirs were then selected according to land use type and the corresponding access permits were obtained (Fig. 1). The reservoirs sampled in situ presented sizes ranging from 0.4 to 54 ha (Table 1).
In the three samplings, rainfall varied substantially depending on the accumulated precipitation in the previous thirty days; 02/12/19 was the driest day with 25 mm, followed by 16/11/20 when 81 mm were computed, and 07/11/19 with 204 mm of accumulated precipitation.
Total phosphorus (TP) was in all cases within the category range of hypereutrophic reservoirs (≥ 0.078 mg TP L −1 , Cunha et al. 2013) (range: 0.18-5.22 mg TP L −1 , Milk production Agribusiness, irrigation reservoir G and I, respectively), with extremely high cases such as reservoirs I and J (Fig. 2). These two reservoirs also accounted for high total nitrogen (TN) values. TN ranged from 0.70 to 23.56 mg TN L − 1 (F and I, respectively). Reservoirs Chl-a also presented a wide range of concentrations (2.8-4439 µg L −1 , B and I, respectively) (Fig. 2). The trophic state index of Cunha et al. (2013) for chlorophyll-a exhibited a gradient from mesotrophic (4 cases) to hypereutrophic (8 cases). The percentage of Chl-a corresponding to cyanobacteria yielded high values in reservoirs categorized as mesotrophic (B, C, E, and G), eutrophic (A), and hypereutrophic (D, F, H, I, J, K, M, and N). Significant correlations were obtained between most of the variables evaluated in situ (Table 2). Those with the highest magnitude were identified between turbidity and total suspended solids (TSS) and transparency and suspended organic matter (SOM) (Rs = 0.93, − 0.87, and 0.81, respectively). SOM was positively associated with Chla, TSS, and TN (Rs = 0.89, 0.84, and 0.81, respectively). Likewise, Cyanobacterial Chl-a (Chl-a Cy) was strongly correlated with total Chl-a and SOM (Rs = 0.91 and 0.84, respectively). On the other hand, size was associated with all evaluated variables. The highest magnitude of correlation was detected in SOM followed by transparency (Rs = − 0.60 and 0.59, respectively) ( Table 2). Even though Chl-a was significantly associated with TP, the correlation showed a high dispersion mainly in the reservoirs with higher turbidity (A, B, C, and G).

Spectral signatures
In the signatures with concentrations above 20 µg Chl-a L −1 , a valley in the blue region (492 nm), a peak in the green one (560 nm), followed by a valley in the red region (665 nm) can be observed, with its lowest Rrs values in the cases with the highest Chl-a concentration. After this valley, a sharp increase can be noticed (Fig. 3). In general terms, the spectral signatures with higher magnitude in the near-infrared region (˃ 704 nm) were those with higher Chl-a concentrations. Conversely, spectral signatures corresponding to reservoirs with concentrations below 20 µg Chl-a L −1 presented a progressive increase from 443 to 665 nm (Fig. 3).
At wavelengths ˃ 740 nm, Rrs was associated with all trophic state variables determined in situ, among which,  (Table 3).

Linear models
Two linear models were developed for satellite estimation of Chl-a concentration based on the Chl-a logarithm observed in situ as the response variable. The normalized differences of the red edge-red Rrs (NDCI) and the green-red Rrs (VR) were used as explanatory variables. For model fitting, the maximum value of Chl-a in situ (4500 µg L −1 ) was excluded as an outlier. Therefore, a total of n = 41 cases were used.
The linear model based on the NDCI indicator (adjusted R 2 = 0.81) showed no evidence of lack of normality in the distribution of the Shapiro-Wilk test results of the residuals (p = 0.44), nor was there evidence of non-homocedasticity according to the Breusch-Pagan test (p = 0.22). Therefore, it is an acceptable model for Chl-a estimation (Fig. 4).

Satellite monitoring
Two Sentinel-2 MSI images were obtained with contrasting precipitation values. The first one corresponds to 12/12/16 with 50 mm of accumulated precipitation on the previous thirty days, while the second was captured on 11/01/17 with 264 mm of accumulated precipitation on the previous thirty days. Among the dates with images available, a maximum value of cyanobacterial abundance (3581 cells mL −1 ) was detected on 01/01/17 in PPAC (Fig. S2).
Monitoring using the VR indicator fitted model revealed that most of the reservoirs are hypereutrophic. On 12/12/16, 225 reservoirs were classified as hypereutrophic, 86 as eutrophic, and 133 as mesotrophic. In addition, the area with Chl-a > 20 µg L −1 according to each reservoir median values reached 861 ha. On 11/01/17, a larger number of hypereutrophic environments (335) and a decrease in eutrophic and mesotrophic ones were detected, 64 and 65, respectively (Fig. 7). Also, the area covered by Chl-a concentrations > 20 µg L −1 was 1044 ha.
On 12/12/16, the median Chl-a value for all reservoir sizes categorized them as super-eutrophic, with the exception of those of 0.6-0.9 ha (quartile 2), which yielded the highest Chl-a concentrations and were categorized as hypereutrophic. It is worth mentioning that the smallest reservoirs Table 3 Spearman correlation magnitude between reflectance (Rrs) of Sentinel-2 bands and trophic state variables determined in situ ISS, inorganic suspended solids (mg L −1 ); SOM, suspended organic matter (mg L −1 ); TSS, total suspended solids (mg L −1 ); Chl-a, chlorophyll-a (µg L −1 ) In bold: cases where a significant correlation was identified according to the p value ˂ 0.05 Rrs (nm) ISS (mg L −1 ) SOM (mg L −1 ) TSS (mg L −1 ) Chl-a (µg L −1 ) Secchi (cm) Chl-a Cy (µg L −1 )  On the other hand, on 11/01/17, the median Chl-a value for all reservoir sizes categorized them as hypereutrophic regardless of their size, the smallest reservoirs also presented the lowest median values while those from quartile 2 accounted for the highest (Fig. 8). On the two dates evaluated, the standard deviation of Chl-a was higher in the smallest reservoirs, followed by the reservoirs from quartile 2 on 12/12/16 and the largest ones on 11/01/17. On the two dates monitored, transparency was similar and tended to increase with size (Rs = 0.38 and Rs = 0.40, 12/12/16 and 11/01/17, respectively). The minimum values were recorded in the smallest size reservoirs (quartile 1), while the largest sized (quartile 4) accounted for the maximum values according to the median (Fig. 8). Chl-a was not significantly associated with size (Spearman test), although it was positively associated with transparency both in December (p < 0.05; Rs = 0.28) and January (p < 0.05; Rs = 0.25).

Hotspot analysis
Spatial clustering patterns were identified in the distribution of Chl-a and transparency of the monitored reservoirs. According to Moran's I test, the highest autocorrelation intensity (Z) was detected for transparency with values of Z = 5.83 in December and Z = 4.88 in January. Likewise, Chl-a was distributed in clusters with higher intensity in December (Z = 2.06) than in January (Z = 1.86).
On 12/12/16, Chl-a formed clusters of high values. The highest intensity clustering occurred in the southern region, followed by the clusters in the eastern, northern, and western areas. The central area was characterized by a large cluster of reservoirs with low Chl-a concentrations. In contrast, on 01/11/17, Chl-a was distributed forming a cluster of reservoirs with low concentrations in the southern area, while the central and western areas did not present a specific clustering pattern. On the other hand, the high-value cluster in the northern area was maintained over time as was the low-value cluster in the center of the basin (Fig. 9).
In December, transparency was distributed forming a cluster with the clearest reservoirs in the southern area and another in the central-eastern area. On the other hand, the most turbid reservoirs close to each other were found mainly in the central, northern, and western areas. In contrast, in January, the southern area presented the largest cluster of  (Fig. 9).

Discussion
Linear models for satellite estimation of Chl-a and transparency based on in situ information allowed to monitor all the reservoirs in the Santa Lucía River Basin for the first time in Uruguay. These results represent the first approximation to the trophic state of 486 small-and medium-sized (0.25-100 ha) agricultural reservoirs located in the basin that supplies drinking water to 60% of the population of this country. Despite the fact that large basin databases on water quality monitoring systems have been developed in Uruguay, they only cover the main river courses and the two largest reservoirs (Aubriot et al. 2017;Somma et al. 2022). Therefore, prior to this work, little was known about the trophic state of more than a hundred productive reservoirs in the basin. This study allowed us to determine that most of these water bodies are in a hypereutrophic state and present high levels of phytoplankton biomass, mostly composed of cyanobacteria, and with a high potential for downstream transport of this biomass. Therefore, either due to their overtopping and release resulting from high rainfall or the possible artificial opening of dams during low water flow, this large number of reservoirs with a detected high biomass of potentially toxic phytoplankton undoubtedly represents a health threat.
The great impact of agricultural activities on the export of nutrients and organic matter in CRSL is manifested in the high degree of eutrophication of productive reservoirs. The percentage of the area covered by crops in the drainage basin has a direct impact on the eutrophication of reservoirs, particularly when they are shallow and have a larger basin size relative to the surface area of their water body (Knoll et al. 2015). In our study, the smallest-sized reservoirs presented the highest concentrations of Chl-a, TP and TN in the three in situ sampling sites. These results are in agreement with Brainwood et al. (2004) who suggested that smaller agricultural reservoirs have higher nutrient and Chl-a concentrations than larger ones. The reservoirs with the highest TP and Chl-a concentrations in our study were located in dairy farms, crop farms, and pig farms. Likewise, the maximum TN values were obtained from sites located on livestock farms (Fig. 2 and Table 1). These results are in line with Chalar et al. (2011), who postulated that dairy and pig farms are one of the main sources of nutrient pollutants in the Santa Lucía River Basin. Modeling of nitrogen and phosphorus in CRSL watercourses, with projections of Fig. 8 Satellite-estimated chlorophyll-a concentration and transparency for 12/12/16 and 11/01/17 clustered by reservoir size quartile productive intensification, predicts a scenario of increased eutrophication under current agricultural practices (Díaz et al. 2021). In this sense, the satellite indices applied in this work are an input for the diagnosis of eutrophication evolution in CRSL.
In the sites sampled in situ, phytoplankton development was found to fit better to TN concentration variability, while a lower TP correlation was observed (Table 2). This behavior agrees with the findings obtained by O' Farrell et al. (2021) who monitored Pampean shallow lakes monthly for three years and found that as the trophic state increases, the growth of cyanobacteria can be limited by N. High TP concentrations in CRSL are generally ascribed to high concentrations of soluble reactive phosphorus (> 88%), mainly associated with surface runoff and soil phosphorus saturation in the basin (Goyenola et al. 2015;Aubriot et al. 2017;Barreto et al. 2017).
Most suspended solids corresponded to suspended organic matter, which is strongly associated with phytoplankton biomass (Table 2). Nonetheless, some reservoirs, despite being hypereutrophic according to TP and TN levels, and having suspended solids mainly made up of inorganic origin, did not show high phytoplankton development. This is consistent with some postulations that suggest that inorganic solids limit phytoplankton growth due to the decrease in radiation penetrating the water column (Reynolds 2006;Dzialowski et al. 2011). It is worth noting that most sites with this behavior are located in the southern portion of the basin (Fig. 1), where clay soils predominate (MGAP 1976) which could increase inorganic suspended solids.
The results obtained in situ show a close proportionality of total Chl-a to cyanobacterial Chl-a (Fig. 2, Table 2). This finding raises the level of health risk in the watershed due to cyanobacterial blooms toxicity (Chorus and Welker 2021). Moreover, they support the statement by Somma et al. (2022) who inferred that the small-and medium-sized reservoirs in CRSL are environments conducive to the development of cyanobacterial blooms and that they can be transported downstream as evidenced by the increased cyanobacterial biomass recorded during high flow periods in Fig. 9 Estimated chlorophyll-a and transparency hotspots in the 486 satellite-monitored medium-and small-sized reservoirs according to the Getis-Ord hotspot analysis the main channel. The evaluation of hot spots of high phytoplankton biomasses in the basin, and their spatial and temporal evolution contributes to eutrophication management and to anticipate high-risk scenarios of biomass transport to the water treatment plant.

Satellite estimation models
In this study, two indicators for satellite estimation of Chl-a were compared. The best results were obtained from the VR indicator that measures the surface reflectance ratio between the green (560 nm) and red (665 nm) bands with an R 2 = 0.85 in the range 3-2992 µg L −1 (Fig. 5). This result is in line with Ha et al. (2017) who used the green-to-red ratio to monitor Chl-a from Sentinel-2 images and obtained successful results even at minimum concentrations < 6 µg Chl-a L −1 , as did Avdan et al. (2019). Likewise, Oliveira et al. (2016) tested said ratio to monitor coastal waters and obtained fits up to R 2 = 0.71 over a wide range from 1 to 974 µg Chl-a L −1 . According to the information obtained in situ in this study, the only case with > 20 µg Chl-a L −1 and no green reflectance peak was the K reservoir in the first sampling. As a consequence, this is an exceptional case worthy of further validation of the fitted model. Likewise, a validation with in situ information of the transparency model is required, mainly due to the inconsistency found in the satellite monitoring where high concentrations of Chl-a and high transparency are found (Fig. 9). This discrepancy could be due to the underestimation of organic solids, which could be related to the high spectral response in the infrared that both organic and inorganic solids in the reservoirs (Neil et al. 2019).
The spectral signatures of the reservoirs sampled in situ presented characteristics of productive inland waters as reported by Spyrakos et al. (2018) and were differentiated into two major groups according to Chl-a concentration. Those reservoirs in which concentrations > 20 µg L −1 were recorded, presented a green reflectance peak between the minimum values recorded in the blue (⁓492 nm) and red (⁓665 nm) bands (Fig. 3). This behavior is expected since, according to Gitelson (1992), the energy absorption maxima of Chl-a is close to the blue and red bands, while the minimum is close to the green (⁓560 nm) band.
The NDCI indicator yielded good results as expected according to Neil et al. (2019) who achieved high NDCI fits in several water bodies. Likewise, Uudeberg et al. (2020) reported fits for NDCI of up to R 2 = 0.89 specifically using Sentinel-2 MSI imagery in very organic and inorganic turbid waters. However, in some of the in situ sampling sites with concentrations < 20 µg Chl-a L −1 , positive slopes between the red and red-edge bands (705 nm) were detected that could lead to Chl-a overestimation. In these cases, reflectance increases progressively from the blue (⁓490 nm) to the infrared (> 705 nm) bands. According to the review by Giardino et al. (2019), such behavior could be explained by the high content of inorganic suspended solids as they are characterized by maximum absorption in the blue and exponential decrease towards the infrared. Therefore, it could be affirmed that in the sampling sites reflectance in the blue and red-edge bands was influenced by the optical behavior of both Chl-a and suspended solids, while the reflectance in the red band was mainly dominated by Chl-a.

Satellite monitoring
In this study, all the reservoirs in CRSL were monitored by fitting the linear models with in situ information. Even though this is the first experience in the country, monitoring plans for agricultural reservoirs integrating satellite estimation tools in a complementary manner have already been developed worldwide. In this sense, Papenfus et al. (2020) and McCullough et al. (2012) found advantages in their application such as spatial scope, high sampling frequency, and the savings in monitoring costs.
In this sense, Papenfus et al. (2020) and McCullough et al. (2012) found advantages in its application such as spatial scope, high sampling frequency, and savings in monitoring costs.
In spite of the fact that less hypereutrophic environments and less surface area covered by high concentrations of Chl-a were detected prior to rainfall (12/12/16) than after rainfall (11/01/17; austral summer), PPAC displayed a postrainfall maximum of cyanobacteria, which is consistent with the results reported by Somma et al. (2022) who associated cyanobacterial abundance in the PPAC with high rainfall. Several investigations suggest that the effect of precipitation can act both as an exacerbating factor of cyanobacterial blooms, through the transport of nutrients from the watershed, and as a temporal disrupting factor through washing and dilution of biomass in precipitation events (Reichwaldt and Ghadouani 2012;Haakonsson et al. 2017;Sinha et al. 2017). The overall increase in biomass recorded in most reservoirs would indicate a growth-stimulating effect due to nutrient input by rainfall and the early summer temperature effect. However, the maximum bloom of cyanobacteria observed in the river channel (PPAC) post-rainfall would not be linked to an effect of the overall biomass flushing, but to possible specific contributions from certain groups of reservoirs. Further research should be conducted to elucidate which reservoirs release significant toxic cyanobacterial biomass to PPAC.
As in the study by Coffer et al. (2021), we identified sets of elevated Chl-a values that transcended watershed and eco-region boundaries, representing essential information for targeting management efforts and risk forecasting. Trophic state mapping reveals the increase in hypereutrophic reservoirs in early summer (Fig. 7). While the changes detected in Chl-a distribution patterns before and after heavy rainfall reveal contrasting situations that could be linked to cyanobacteria transport to PPAC. Prior to rainfall, the main hotspots of Chl-a are reported in the southern and eastern areas of the basin. After rainfall, the distribution is reversed and these reservoirs become coldspots (Fig. 9). Therefore, the reservoirs located in the southern and eastern areas of the basin could be the main contributors to the cyanobacterial maximum recorded in PPAC.

Conclusions
All the reservoirs in the CRSL were remotely monitored for the first time in Uruguay. Most of these water bodies were classified in a hypereutrophic state, with high levels of phytoplankton biomass, mostly composed of cyanobacteria. The linear models fitted for satellite estimation of Chl-a and transparency achieved high fits compared to those reported in the literature for turbid and small environments. This is the first experience at a national level in which chlorophylla concentration and transparency were monitored in all the reservoirs of CRSL, after fitting a linear model with information obtained in situ. The relevance of advancing the research agenda on the validation of the fitted models based on new information should be emphasized.
The results obtained in situ allowed us to identify two sets of environments. On the one hand, one with high concentrations of Chl-a, phosphorus, and total nitrogen, and on the other, one that, despite having high concentrations of nutrients, does not develop phytoplankton biomass, and turbid conditions caused mainly by inorganic solids in suspension prevail.
The reservoir cluster located in the southern and eastern areas of CRSL presented the highest risk of high cyanobacterial biomass contribution according to Chl-a distribution before and after rainfall, which would have led to biomass transport to the water treatment plant intakes. The geo-statistical analyses allowed to evaluate all the reservoirs on a synoptic basis and thus identify behaviors and differentiate areas at risk of phytoplankton biomass export. An in-depth study of biomass transport to the drinking water treatment plant should be conducted, based on the analysis of the Sentinel-2 series (2016-present) available for CRSL and the joint analysis of the water quality dataset recorded in PPAC.
Author contribution BZ participated in the field work and in the laboratory analyses, processed the satellite images and data analysis, and contributed to the interpretation and discussion of the results obtained.
LA directed the study and participated from the in situ data acquisition to the discussion of the results. HO participated in field work and laboratory analyses, performed chemical analyses, and contributed to data interpretation and discussion. MA guided the construction of satellite estimation models and geostatistical analyses and contributed to the interpretation and discussion of data. All authors actively participated in the writing of the manuscript and gave final approval for publication.
Funding The research was funded by the National Agency for Research and Innovation under the code POS_NAC_D_2020_1_163887 and the Sectoral Commission for Scientific Research of the University of the Republic through the project I + D (2018) No. 354.

Data availability
The datasets used during the current study are available from the authors on reasonable request.

Declarations
Ethics approval and consent to participate This is not applicable.
Consent for publication This is not applicable.

Competing interests
The authors declare no competing interests.