Spectrally active characteristics of water can be detected by using satellite-based remote sensing techniques and have wide applications in limnological studies. Specifically, many recent water quality monitoring programs were supported by remote sensing (Chen and Han 2018; Najafzadeh, Homaei and Farhadi 2021; Prasad et al. 2015; Tian et al. 2023; Topp et al. 2020). Presently, the models developed by using satellite-imaging and remote sensing have become an attractive alternative to traditional algae monitoring programs (Sellner, Doucette and Kirkpatrick 2003).
Although satellite sensors do not measure pigments directly, empirical or semi-analytical models can be applied to derive pigment levels from satellite observations of water-leaving radiances (Lee and Carder 2004). In previous studies, chlorophyll-a was most widely used proxy for the detection of cyanobacterial blooms because it is the most abundant light harvesting pigment in cyanobacteria. However, in Peridinium and many other dinoflagellates other than chlorophylls carotenoid pigments play a major role in light harvesting. Thus, carotenoids can serve as a unique biomarker for Peridinium and other dinoflagellates containing carotenoids. The spectral signatures of carotenoids allow specific detection carotenoid containing phytoplankton in a mix population of phytoplankton where chlorophylls function as the major photosynthetic pigment. Therefore, in this study, we utilized specific spectral signatures of carotenoids to predict Peridinium dinoflagellates in freshwaters.
Sentinel-2 sensor is the most preferred sensor for monitoring blooms in inland waters due to several advantageous feature such as high spatial resolution (10 m pixel), high spectral resolution with cluster of narrow spectral bands, high signal to noise ratio, frequent revisit time and 13-channel spectral range (Caballero et al. 2020; Gascon et al. 2017; Laneve et al. 2021). Further, free public accessibility to Sentinel-2 images with fine-scale mapping has widened its applications. Therefore, Sentinel-2 sensors would replace most previously used sensors and may provide more accurate monitoring options for bloom monitoring (Gunawardana et al. 2022; Klemas 2012).
Accuracy of a prediction largely depends upon the algorithm used in the empirical model, while the strength of the algorithm is a function of selected bands and band ratios. Further, the accuracy of the prediction model also depends on the selected band combination and the way that band incorporated into the algorithms. Compared to the other band combinations, the reflectance ratio derived in the present study by using B2 (490 nm) and B3 (560 nm) was highly correlated to the measured carotenoid concentration (R2 = 0.93, p < 0.001). Typically, carotenoid pigments absorb wavelengths in the range of 400nm to 550nm (Merzlyak, Solovchenko and Gitelson 2003; Sai et al. 2019). Since wavelength range of Sentinel-2 band 2 and band 3 fall within this range, those two bands performed well in the estimation of carotenoid pigment concentration.
Detection range of carotenoid pigment is an important factor which enhances the model performance. If the measured carotenoid pigment concentration varies in a narrow range, the accuracy of the model will also be limited to that range. As the measured carotenoid data set used in the present study spanned a wide range (19.39–91.86 µg/L), our empirical estimation has a wider applicability. Specifically, if the model sensitivity is restricted to high concentrations of carotenoids, model will only be able to identify high carotenoid concentrations that occur due to dense blooms. Such models would not benefit to take remedial actions, as the bloom has already been established. In contrast, if the model sensitivity spans over reliable estimation of low carotenoid concentrations, it will provide an early warning on future bloom consequences and assist managers to take remedial actions to suppress further bloom development. Further, we used a separate measured data set for model validation without using the same data set used to model development. Therefore, we were able to confirm the accurate prediction of carotenoid concentrations even if they were beyond to the range of initial data set that were used in model development.
Reduced time window and reduced cloud cover are another two key factors which directly affect to the prediction accuracy of a model. According to the time window factor, reduced time gap between field data measurements and the date of satellite image enhances the performance of the algorithm (Boucher et al. 2018). Spatial variation between in-situ sampling and the satellite image acquisition can increase the uncertainty of the algorithm outputs (Toming et al. 2016). 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). During the present study, for model development we were able to obtain satellite images captured within 1 day time gap to the in-situ sampling. Further, for model validation we were able to obtain satellite images captured on the same day of in-situ sampling. Therefore, errors that could be arisen due to time window is minimum and more reliable predictions could be made using the present model. Further, analysis on satellite revisit time prior to the in-situ sampling is an important good practice when developing a model.
Model developing season is also an important factor since it may cause several errors in photosynthetic pigments estimation including underprediction of high pigment concentrations and difficulty in quantifying low pigment samples (Boucher et al. 2018). Due to the changing composition of phytoplankton assemblages throughout the year, optical properties of blooms would be changed. Hence, season-specific algorithms are required for particular regions (Ligi et al. 2017). Due to these seasonal effects, algorithms are need to be tuned by season. Here, only certain bands may be subject to more seasonal variation than others depending on the types of lakes (Boucher et al. 2018). Therefore, algorithm tuning by season is an extensive process which requires broad understanding on both remote sensing and environmental variations. As a tropical country, Sri Lanka does not experience much intense seasonal variations. However, dry and wet seasons could have significant difference in carotenoid concentrations. We observed comparatively high carotenoid pigment concentration and Peridinium cell density in the first sampling which was carried out during a prolonged dry weather whereas, both carotenoid concentration and Peridinium cell density were less in the second sampling which was carried out few weeks after a rainy weather.
Humic substances such as coloured dissolved organic matter (CDOM) are naturally occurring water-soluble organic substances that are yellow to brown in colour (Conmy 2008). Organic matter in water absorbs UV and visible light at the blue end of the spectrum (McKnight et al. 2001). Increasing CDOM result in low reflectance which is often below 500 nm (Menken, Brezonik and Bauer 2006) whereas absorbance of carotenoids is maximum at 490nm (Yacobi, Gitelson and Mayo 1995). In this study we considered reflectance at 490 nm and 560 nm. Therefore, we trust that the selected band region is strong enough to distinguish the reflectance of carotenoid pigments from the reflectance of CDOM.
Our literature survey showed that the majority of existing HAB monitoring models have been designed to monitor cyanobacterial blooms and less attention has paid on dinoflagellates although they are capable of exerting similar risk. Specifically, Peridinium dinoflagellate blooms have been recorded in worldwide freshwater systems along with several environmental, ecological and health consequences. But well establish framework to monitor Peridinium blooms with sufficient spatial and temporal resolution are currently lacking. In this context, the model algorithm developed in the present study for predicting Peridinium blooms with high accuracy has (> 85% accuracy) has significant importance in ecological and public health. It is an efficient, inexpensive, and reliable manner approach of remote sensing of Peridinium blooms. Probably, this would be the first such attempt to monitor and quantify Peridinium blooms in freshwaters using remote sensing approach. Not only that, species identification through remote sensing approach has rarely been successful in global scale (Kudela et al. 2015). Therefore, our findings are of high importance in regional and global scale. Further, this model can be applied for the identification of other freshwater dinoflagellate blooms which contain carotenoid pigments. However, quantification of cell density is specific to the Peridinium dinoflagellate since the model was developed using the real-time data where Peridinium was successfully colonized. Therefore, model cannot be used for quantification of other dinoflagellate blooms. Further, the present model was not validated to outside of the region. Therefore, before applying this model in other areas, further tunning based on the regional atmospheric effects is recommended for better performance.
Overall, results obtained in this study implicate the importance and capacity of remote sensing in dinoflagellate bloom monitoring. Remote sensing of Peridinium provides a versatile tool for routing monitoring and surveillance of bloom development. Further, the present model will facilitate timely implementation of public alerts on potential health risks. Not only that, this approach can be adopted to freshwaters elsewhere in the world with simple recalibration and validation. Hence, it is recommended to focus future work on extending applications to monitoring different dinoflagellate blooms other than Peridinium.