Quantitative prediction of toxin-producing Aphanizomenon cyanobacteria in freshwaters using Sentinel-2 satellite imagery

Deterioration of water quality in lakes, streams and reservoirs is a major global environmental and socioeconomic concern. In particular, freshwaters experience the unprecedented influence of nutrient enrichment, global warming and inorganic and organic pollution, and consequently become conducive for cyanobacterial bloom formation. Some of the cyanobacterial blooms are considered harmful (collectively referred to as harmful algal blooms, HABs) as they produce cyanotoxins. In addition, HABs reduce dissolved oxygen, alter the taste and odour of drinking water, reduce aesthetic value and impede the recreational value of the waterbody. Therefore, the presence of HABs stands as an indicator of water quality deterioration. Hence, frequent monitoring of their growth and development towards bloom formation is essential for the detection of the status of bloom and prevention of associated health risks.

Aphanizomenon is a filamentous, nitrogen-fixing cyanobacterial genera belonging to the order Nostocales. It is geographically widespread and commonly found in freshwaters at a wide range of temperatures (15–30 °C) (Cires & Ballot 2016). Globally, Aphanizomenon has been detected in several toxic cyanobacterial blooms including North American (Carmichael et al. 2000), European (Carrasco et al. 2007; Brient et al. 2009) and Asian reservoirs (Liu et al. 2006; Wu et al. 2010). Nostocales cyanobacteria are considered to be potentially invasive due to their inherent advantageous characteristics such as the presence of heterocyst for atmospheric nitrogen fixation, ecological plasticity, and development of resting and dispersal structures (akinetes) under adverse conditions (Sukenik et al. 2012). Aphanizomenon produces a range of cyanotoxins, namely anatoxin-a (ATX), cylindrospermopsin (CYN), microcystins (MCs) and saxitoxins (STX) (Table 1).

Among these toxins, neurotoxic ATX is the most dangerous form of cyanotoxin which can cause death within minutes or hours and is considered a Very Fast Death Factor (VFDF) (Trost & Oslob 1999). Importantly, ATX was found airborne during a harmful algal bloom where Aphanizomenon flos-aquae and Cuspidothrix issatschenkoi were present (Sutherland et al. 2021). This is an alarming new route of human exposure to ATX that has not been revealed before. Therefore, a predominance of multiple toxin-producing cyanobacterial species such as Aphanizomenon in the freshwaters need thorough attention. Under this context, early detection and timely public awareness have prime importance in minimizing potential health risks. Although the global warming and climate changes are considered as the prime factors for the high prevalence of algal blooms in general, actual triggers that drive toxic Aphanizomenon bloom formation remain unclear to date (Cires & Ballot 2016). Hence, frequent and long-term monitoring of water bodies are essential for understanding the dynamics of bloom formation, which will help to decide management criteria to control algal blooms. In this context, novel and robust water quality assessment techniques including satellite image-based modelling coupled with mathematical modelling would be better alternatives for the traditional water monitoring techniques.

In situ data collection and subsequent laboratory analysis are the most common approaches of monitoring water quality and algal blooms. However, owing to the laborious nature of currently practising traditional methods, nowadays in situ monitoring of water quality in combination with remote sensing is becoming a popular tool in global water quality monitoring programmes (Arabi et al. 2020; Chu et al. 2021). Remote sensing is a tool that can acquire a large amount of data with a sufficient spatial and spectral resolution to cover large geographical areas along with a high temporal coverage (Matthews et al. 2012). Furthermore, remote sensing can predict the dynamics of commonly used indicators of both cyanobacteria and HABs. Some of these indicators are chlorophyll-a, phycocyanin and water turbidity (Kudela et al. 2015). Chlorophyll-a (chl-a) is a spectrally active photosynthetic pigment present in both green algae and cyanobacteria. Therefore, chl-a is the most preferred pigment of bloom monitoring over phycocyanin which is an accessory pigment specific to freshwater cyanobacteria. Over the past few years, different algorithms have been developed to estimate the chl-a concentration using remote sensing (Boucher et al. 2018). Although significant relationships between in situ chl-a concentration with reflectance have been derived (Allan et al. 2011; Bresciani et al. 2011; Keith et al. 2012), the applicability of such relationships found limited accuracy outside the original area of the study and season of the year (Moses et al. 2012; Ligi et al. 2017; Boucher et al. 2018). Applicability of previously developed six models on 192 freshwater lakes across two states in the USA, New Hampshire and Maine, were tested for remote sensing of chl-a (Boucher et al. 2018). Those models were originally developed to apply in oceans, coastal and estuarine waters and large freshwater lakes. According to their findings, remotely predicted chl-a weakly correlated to the in situ measured chl-a while the strength of the correlation varied with the season and chl-a threshold. Therefore, regional-scale models which are developed by using best performing remote sensing algorithms for the implementation of efficient monitoring programmes is particularly important. As reviewed by Sebastiá-Frasquet et al. (2020), most of the existing remote sensing studies of algal blooms have an origin of either the USA, China or UK. However, this is an emerging research field where novel tools are being applied to achieve a higher accuracy. Recently, a novel interactive cloud-based dashboard called ‘CyanoKhoj’ has been developed by using Sentinel-3 Ocean and Land Colour Instrument (OLCI) data for rapid monitoring of cyanobacterial blooms across India (Maniyar et al. 2022). Other than that, remote sensing of algal blooms is not practised in the South Asian countries including Sri Lanka except for some publications originated from sporadic and short-term research programmes (Ha et al. 2017; Poddar et al. 2019; Wagle et al. 2019; Caballero & Navarro 2021; Gunawardana et al. 2021; Chusnah & Chu 2022; Tham et al. 2022).

Most of the existing cyanobacterial prediction algorithms fail to discriminate between cyanobacterial genera (Kutser et al. 2016). However, a separate monitoring of cyanobacterial genera is very important since some of these genera are more likely to be toxic than others (Paine et al. 2018). Even though there is scientific validity in this regard, only a limited effort has been made so far in this context. For example, Kudela et al. (2015) developed a model by using MODIS/ASTER (MASTER) Airborne sensor to distinguish Aphanizomenon and Microcystis in Pinto Lake, California. However, they were unable to do quantitative (cell density) prediction of either species, and suggested employing a different sensor having high spectral resolution and appropriate atmospheric corrections for separate monitoring of Aphanizomenon and Microcystis blooms. Apart from the limitations drawn by the airborne sensor used in the above study, the developed algorithms were applied only in eutrophic water, sparing a gap in the sensitivity of algorithms in monitoring Aphanizomenon in oligotrophic and mesotrophic waters. Therefore, here we highlight the necessity of developing Aphanizomenon index-derived algorithms that are capable of applying across a broad range of trophic status in water with high accuracy.

Among the different algorithms developed by using different sensors, empirical algorithms that directly correlate remote sensing data with in situ measured data have been successfully applied not only to monitor algal blooms but also to quantify algal pigments, such as chl-a, cyanobacteria-specific pigment, phycocyanin and turbidity (Chang et al. 2004; Ruiz-Verdú et al. 2008; Petus et al. 2010). Different satellite imagers, such as WorldView-2, Sentinel-2, Landsat-8, MODIS and MERIS have been widely used for remote sensing of algal bloom with a varying degree of success and those were reported in the literature with their limitations (Gitelson et al. 1993; Mayo et al. 1995; Duan et al. 2008; Chavula et al. 2009; Moses et al. 2009; Beck et al. 2016; Poddar et al. 2019; Sebastiá-Frasquet et al. 2020). Specifically, many of those sensors are less preferred at present due to low spatial and spectral resolution, low signal-to-noise ratio and inadequate repeat cycles for terrestrial satellites (Duan et al. 2008). Sentinel-2A and 2B launched in June 2015 and July 2016 by the European Space Agency complement many of those issues. Sentinel-2 has a high spatial resolution (10 m pixel), weekly revisit time and a cluster of narrow spectral bands making it an excellent option to monitor cyanobacterial blooms (Gascon et al. 2017; Caballero et al. 2020). Furthermore, the opportunity given by Sentinel-2 for fine-scale mapping with free public access has widened its applications. Therefore, Sentinel-2 sensors would replace most previously used sensors and may provide more accurate monitoring options for cyanobacterial blooms.

Many empirical algorithms have been derived either by using various bands or band ratios of the reflectance spectrum of the Sentinel-2 with their own sensitivity and detection range (Sebastiá-Frasquet et al. 2020). However, the majority of those algorithms were rarely evaluated outside the original place of development (Sebastiá-Frasquet et al. 2020), while those were restricted to a narrow range of chl-a concentrations (Ha et al. 2017; Yadav et al. 2019; Germán et al. 2021). The sensitivity of algorithms appeared dependent upon the selected spectral band combinations and ratios. For instance, Chen et al. (2017) obtained relatively low accuracy (R² = 0.49) when B4 and B5 bands were used while Ha et al. (2017) obtained high accuracy (R² = 0.68) with band ratio of B3 and B4, Yadav et al. (2019) obtained high accuracy (R² = 0.76) with B2, B3, B4 and B8 band combinations and Germán et al. (2021) obtained even higher accuracy (R² = 0.77) with B8/B4 band ratio. On the other hand, only a few Sentinel-2 algorithms specifically detect a particular species of bloom-forming cyanobacteria such as Microcystis (German et al. 2020) while few other algorithms focused on total cyanobacterial count or total biovolume of cyanobacteria (Beck et al. 2017; Pompêo et al. 2021). None of the Sentinel-2 algorithms were developed for the specific detection and quantification of Aphanizomenon density to date. Hence, there will still be a large room for novelty and improvement of existing models for robust algorithms to keep pace with the rapidly evolving field of remote sensing and detection of cyanobacterial blooms.

Our aim in this study was to develop an empirical model to predict spatial distribution of Aphanizomenon cell density by coupling in situ data with Sentinel-2 reflectance data to support the developing efficient monitoring strategies of algal blooms in the future.

Field samplings were carried out on 7 December 2020 and 30 November 2021. The in situ chl-a concentrations were measured using a chlorophyll-a sensor (TAL-PC, YSI, USA). Overall, 90 data records from 30 sampling sites on 7 December 2020 and 36 data records from 18 sampling sites on 30 November 2021 were collected. Coordinates at each sampling site were recorded with a portable GPS (eTrex 22x, Garmin, UK).

For phytoplankton analysis, approximately 100 L of water sample was filtered through a plankton net (30 μm, Hydro-Bios, Germany) and planktons were collected into 500-mL amber high-density polyethylene bottles at each sampling site. Samples were observed with the aid of a bright-field microscope equipped with a digital camera (Nikon, Japan) to identify dominant phytoplankton. Their cell density was estimated following the method described in Intergovernmental Oceanographic Commission (IOC) Manuals and Guides No.55.

Terrain-corrected Sentinel-2 MSI images of the study area were downloaded from the USGS (United States Geological Survey) server using the Earth Explorer platform (http://glovis.usgs.gov/) from the nearest date to the field sampling. Accordingly, images dated 7 December 2020 and 3 December 2021 were downloaded. All images were selected specifically to have <20% cloud cover, and sampling sites near to a cloud were excluded. ESRI ArcGIS 10.6.1 (Esri Inc., USA) and QGIS 3.16 software were used to process acquired satellite data.

Dark object subtraction (DOS), which is a quick and simple method for atmospheric correction was used to minimize atmospheric effects (Giardino et al. 2001; Urbanski et al. 2016). The scattering was removed by subtracting the darkest pixel value from every pixel in the band (Ruddick et al. 2000; Boucher et al. 2018).

For developing empirical models with dual-model strategy, regression analysis was carried out on the combined dataset of chl-a taken from the 30 sampling sites on 7 December 2020. Several empirical regression models were developed between in situ chl-a values and same day reflectance values. Direct and log transformed chl-a data were applied with reflectance values of various spectral band combinations of B2, B3, B4, B5, B6, B7 and B8. To determine the agreement of developed models, coefficients were determined using linear regression analysis and the best performing model was selected based on the regression coefficients of determination (R²) at p < 0.001.

Furthermore, based on the developed relationship between Aphanizomenon cell density and chl-a concentration, Aphanizomenon cell density of the Ridiyagama reservoir was predicted. Sensitivity of the model was determined according to the regression analysis. RMSE and percentage RMSE were also calculated to evaluate the model performance.

It is well accepted that chl-a could be considered as a proxy of eutrophication in the lentic systems, where the chl-a concentration above 9 μg/L indicates eutrophication (Nürnberg 1996; Mamun et al. 2020). Accordingly, Ridiyagama reservoir on 7 December 2020 was eutrophic while it was oligotrophic on the other sampling date (Table 2). Aphanizomenon density was nearly 2 times higher in eutrophic water and in the presence of bloom compared to oligotrophic water (Table 2).

For modelling, 90 records of chl-a were used as the response variable and spectral bands of the visible light portion (B2 [blue], B3 [green], B4 [red]) and Vegetation Red Edge portion (B5, B6, B7) and NIR (Near Infra-Red [B8]) as predictor variables. Algorithms were developed for the estimation of chl-a concentration by considering all possible band regions of the reflectance spectrum. In total, 130 algorithms with different band combinations were developed and algorithms that were statistically significant (p < 0.001) and gave high R² are listed in Table 3.

Remote sensing is a promising technology for tracking large-scale high-frequency data compared to traditional water quality monitoring surveys which are often time-intensive, expensive, laborious and unable to provide data on a wider scale. In the present study, we showed successful application of Sentinel-2 data to predict chl-a and Aphanizomenon cell density in a freshwater reservoir. Although the developed models were derived using data collected from eutrophic (chl-a > 9.0 μg/L) condition, the same model could also be applicable for oligotrophic waters to predict chl-a concentrations and Aphanizomenon cell density with a high reliability.

Application of Sentinel-2 image data in water quality assessments is becoming popularly used in limnologic studies (Sòria-Perpinyà et al. 2020) as there is high spatial resolution compared to other remote sensing applications such as Landsat-8, MODIS, etc. The spectral range of Sentinel-2 bands spans over a wide range where B1, B2, B3 and B4 are coastal aerosol, blue, green and red bands, respectively. B5, B6 and B7 are vegetation red edge bands while B8 is the NIR band. Algorithms based on the red and NIR bands are more successful in predicting chl-a concentration (Tebbs et al. 2013). The absorption peak at the short wavelength region (450–500 nm) is given due to the combination effect of chl-a, carotenoids and dissolved organic matter (Gitelson et al. 2000). Reflectance of green peak (543–578 nm) represents the minimum pigment absorption of chlorophyll, phycocyanin and carotenoids and scattering by non-organic particles and phytoplankton cell walls (Gitelson et al. 2000; Schalles & Yacobi 2000). Small peak at 620 nm represents the reflectance caused by phycocyanin and the peak at 705 nm (red band) is caused by the suspended matter in water, including algal cells and absorption of water. (Gitelson et al. 2000). The next three bands in the spectrum are vegetation red edge bands and those bands are the newly included bands in order to enhance the applications of Sentinel-2 sensor. Therefore, algorithms developed using these red (B4), vegetation red edge (B5, B6 and B7) and NIR (B8) bands have promising applications in inland water studies.

The available empirical algorithms derived by Sentinel-2 data represent a varying degree of accuracy. The detection range of those empirical models depend upon the selected spectral bands and band ratios. Obtaining a high correlation between the predicted chl-a using remote sensing tools and in situ measured chl-a concentrations along with a broad detection range is a challenging task. Even with the same spectral band combinations, two algorithms are not identical as there are differences in the way that these data were incorporated into the algorithm. Similar to the present study, Chen et al. (2017) also used both B4 and B5 bands, while the level of accuracy of the present study (R² = 0.81) was comparatively higher than the former study (R² = 0.49). Although the same bands were selected in both studies, the applied band ratios were different. The ecological models are continuously modified to achieve better predictions together with a high accuracy either by replacing or adding new variables to the existing models and claims novelty. Chl-a and Aphanizomenon have a significant absorbance at λ₆₆₅ compared to λ₇₀₅ where both have a significant reflectance (Paine et al. 2018). Instead of applying original reflectance, the present study scaled up the reflectance at the significant absorbance phase of λ₆₆₅ (1 + R₆₆₅) while it down at the significant reflectance phase of λ₇₀₅ (1−R₇₀₅) to enhance the intended ratio of the two bands as a new approach. Modified reflectance ratios were well regressed compared to using their original values in our study (Table 3). On the contrary, moving more towards the NIR region hardly attain spectral distinction between species (Paine et al. 2018) and it was ascertained that B4 and B5 bands worked well in the present empirical model rather than conventional λ₆₆₅ (B4) to λ₈₄₂ (B8) ratio.

Generally, single reflectance band algorithms are preferred for in situ radiometry, when atmospheric correction is not an issue, whereas band ratios are preferred when atmospheric factors that affect reflectance are present such as sun glint and other atmospheric factors (Tebbs et al. 2013). Compared to single bands, a two-band ratio algorithm [(1 + R₆₆₅)/(1 − R₇₀₅)] derived from B4 and B5 of the spectrum performed best for the prediction of chl-a concentrations. Furthermore, the selected B4 and B5 regions of the reflectance spectrum permit separate detection of cyanobacteria from green algae that would interfere distinct monitoring of cyanobacterial blooms (Paine et al. 2018). According to the time window factor, decreasing the time between in situ chl-a measurement and the date of satellite image is better for the performance of the algorithms (Boucher et al. 2018). The most robust algorithms are those that were developed with matching time between in situ sampling and remote sensing data acquisition (Kabbara et al. 2008; Keith et al. 2012). In this study, we were able to capture satellite images within 3 days of in situ sampling and thereby increased the performance of predictive models. Being a tropical country, cloud cover is a major concern when obtaining satellite images especially for temporal data retrieval. In the present study, Sentinel images of sampling sites disturbed by cloud cover were excluded.

Different forms of algorithms have been developed worldwide and reported in the scientific literature to predict chl-a concentrations by using remote sensing techniques (Matthews et al. 2010, 2012; Tebbs et al. 2013; Keith et al. 2018). Some of those were only applicable to detect chl-a contents of oligotrophic waters (chl-a concentration <9 μg/L). Compared to those previously established models, present study was able to predict chl-a concentrations with a high accuracy (R² = 0.83) together with wider detection range of chl-a (approximately 3–100 μg/L) extending its applicability for both oligotrophic and eutrophic waters. On the other hand, validation at a wide range of in situ data would minimize the false estimation. Therefore, the empirical algorithms developed in the present study would permit their applicability in freshwaters in other geographic regions with simple recalibration and validation.

Furthermore, we established a relationship between Aphanizomenon cell density and reflectance values to predict Aphanizomenon cell density. As some species of the genera are capable of producing cyanotoxins, frequent and timely monitoring is highly important to prevent health risks associated with algal blooms. In this context, remote sensing of Aphanizomenon provides a versatile tool for routing monitoring and surveillance of bloom development. It will facilitate timely implementation of public alerts on potential health risks. Previously, Aphanizomenon was remotely predicted in Pinto Lake in California using the data MASTER Airborne sensor (Kudela et al. 2015), and their detection was based on the spectral-shape algorithms which lacked Aphanizomenon index-derived quantification. Therefore, we attempted to predict Aphanizomenon cell density in freshwaters through the model.

As the formation of cyanobacterial blooms is increasing worldwide, empirical models can provide a significant contribution to minimize anticipated health impact by facilitating early warnings. Routing availability of such data would allow more reliable decision-making by authorities towards efficient management plans to ensure public health. Remote sensing of multiple toxin-producing Aphanizomenon is particularly important for Ridiyagama reservoir because of its multipurpose nature. In the future, applications could further be extended by coupling other proxies such as turbidity, dissolved oxygen, nitrogen, and phosphorus in the aquatic environment, for example, to predict trophic status, identification of bioindicators of water nutrient pollution, etc.

We developed robust algorithms for monitoring spatial variations in the distribution of chl-a and Aphanizomenon cell density in freshwaters. The study was based on the in situ measurements of chlorophyll-a that were correlated with reflectance data of Sentinel-2 images relevant to the study area. Our results indicate that chl-a and cyanobacteria density can be detected and quantified with significantly high accuracy by hyperspectral remote sensing. The best algorithm was obtained from the reflectance band ratio of (1 + R₆₆₅)/1−R₇₀₅) with an R² of 0.81 (p < 0.001). The relationship between in situ and predicted chl-a values displayed a correlation (R²) of 0.83, RMSE of 6.13 with a very high significance (p < 0.001), confirming the reliable prediction of chlorophyll-a concentration. Furthermore, an empirical model was developed to predict cell density of Aphanizomenon which is a cyanotoxin-producing cyanobacterial genus that has global distribution. The models developed here could be successfully used to predict quantitative and spatial distribution of chl-a and Aphanizomenon and its potential for bloom formation in both eutrophic and oligotrophic water in Ridiyagama reservoir, and this approach can be adopted to freshwaters elsewhere in the world with simple recalibration and validation.

This research was supported by the Accelerating Higher Education Expansion and Development (AHEAD) Operation of the Ministry of Higher Education Sri Lanka funded by the World Bank. Furthermore, the support given by the National Aquaculture Development Authority of Sri Lanka during field data collection is highly acknowledged.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.