Using Hyperspectral Remote Sensing to Monitor Water Quality in Drinking Water Reservoirs

Abstract

At the Blankaart Water Production Center, a reservoir containing 3 million m${}^3$ of raw surface water acts as a first biologic treatment step before further processing to drinking water. Over the past decade, severe algal blooms have occurred in the reservoir, hampering the water production. Therefore, strategies (e.g., the injection of algaecide) have been looked at to prevent these from happening or try to control them. In this context, the HYperspectral Pointable System for Terrestrial and Aquatic Radiometry (HYPSTAR), installed since early 2021, helps in monitoring the effectiveness of these strategies. Indeed, the HYPSTAR provides, at a very high temporal resolution, bio-optical parameters related to the water quality, i.e., Chlorophyll-a ($Ch{l}_a$) concentrations and suspended particulate matter ($SPM$). The present paper shows how the raw in situ hyperspectral data (a total of 8116 spectra recorded between 2021-02-03 and 2022-08-03, of which 2988 spectra passed the quality check) are processed to find the water-leaving reflectance and how $SPM$ and $Ch{l}_a$ are derived from it. Based on a limited number of validation data, we also discuss the potential of retrieving phycocyanin (an accessory pigment unique to freshwater cyanobacteria). The results show the benefits of the high temporal resolution of the HYPSTAR to provide near real-time water quality indicators. The study confirms that, in conjunction with a few water sampling data used for validation, the HYPSTAR can be used as a quick and cost-effective method to detect and monitor phytoplankton blooms.

1. Introduction

The increasing occurrence of extreme summers, urban growth and socio-economical pressure challenges the safe-guarding and availability of surface and ground water quality. In 2021, the European Commission reported the need for actions to ensure the availability and sustainable use of freshwater resources [1]. Hence, there is an urgent need in extensive temporal and spatial water quality monitoring systems to provide scientific guidance for water resources management. Traditionally, water quality monitoring relied on sampling the water with a container and analyzing the water sample in the laboratory. However these discrete in time and spatially coarse sampling methods provide insufficient frequency and spatial coverage to capture fluctuations and trends in water quality or to investigate the effectiveness of remediation programs [2]. In recent years, significant advances have been made in online technologies to measure physical and chemical parameters using autonomous monitoring systems [3,4]. A large number of online sensors exists, each providing advantages and limitations.

Among the automated optical-based technologies, above-water visible and near-infrared optical radiometers have already shown to be a cost-effective solution for monitoring water quality and aquatic biogeochemical processes [5]. With adequate processing [6], the measured reflected light is processed as the water-leaving reflectance, also referred to as the water color. Variations in spectral features seen in the water-leaving reflectance (i.e., in magnitude and shape) thus reflects variations in the inorganic, organic, particulate and dissolved materials present in the water, and, subsequently, provide insight into the water quality. With the development of hyperspectral radiometry, even finer-scale spectral features can now be detected [7,8], pushing the ability of optical sensors to address aquatic biodiversity, such as phytoplankton composition and abundance [9,10,11]. Palmer et al. [12] provide an overview of studies using hyperspectral data focusing on various inland water characteristics ranging from Chlorophyll-a ($Ch{l}_a$) concentrations and the presence of organic and inorganic material to water turbidity/transparency and the identification and quantification of cyanobacteria blooms (e.g., [13,14,15]).

When mounted on automated pointing systems, above-water sensors provide water reflectance data at high temporal resolution and with very low manpower requirements. The development of such systems was mostly driven for the validation of satellite data [16,17,18]. However, with their very high temporal resolution data, autonomous systems may also be used for effective water quality monitoring. Cullen et al. [19] showed how one multispectral radiometer buoy could aid in monitoring and characterizing biological variability in surface waters over scales from minutes to months. Muller-Karger et al. [20] confirmed that long-term trends analysis requires measurements of short-term variability over long periods of time. Similarly, Song et al. [21] showed the feasibility of using hyperspectral remote sensing techniques as a rapid assessment tool for determining cyanobacterial abundance in drinking water systems.

In this context, the HYperspectral Pointable System for Terrestrial and Aquatic Radiometry (HYPSTAR) mounted on a pointable device, installed since early 2021, helps in monitoring the water quality over the Water Production Center (WPC) at Blankaart. The Blankaart WPC reservoir, located in Belgium 20 km from the Belgian coast (50.9888°N–2.8352°W), holds 3 million m${}^3$ of water and acts as a first biologic treatment step before downstream processing. The Blankaart WPC is one of the main sources of drinking water for the coastal province of Belgium. During the last decade, the WPC has experienced several algal blooms. Due to their difficult removal from the water (more chemicals such as flocculants needed, clogged filters, etc.), algal blooms affect the production process of drinking water and force the water managers to rely on non-optimal water production planning. Knowing that Belgium is among the most water-scarce regions in Europe [1], preventing such situations is essential to ensure a sustainable drinking water supply.

The aim of the present study is (1) to demonstrate the capabilities of the HYPSTAR to provide reliable low-maintenance long-term reflectance measurements at very high temporal resolution, and (2) to investigate how the operational use of these continuous reflectance measurements may help in assessing the water quality. Here, we focus on two major water quality indicators, $Ch{l}_a$ and suspended particulate matter ($SPM$). The $Ch{l}_a$ concentrations provided by De Watergroep (i.e., water managers of the Blankaart WPC) from water sampling points allows us to validate and further improve the algorithms. In addition, to demonstrate the ability of hyperspectral sensors to retrieve more advanced water quality parameters, we investigated the potential of the HYPSTAR to detect phycocyanin, PC, an accessory pigment unique to freshwater cyanobacteria. However, the retrieval of PC is based on a limited number of match-ups between water sampling and HYPSTAR data and is therefore presented here as a feasibility study. Two extra measurements taken at Blankaart in 2019 and 2020 with the hyperspectral sensor PANTHYR [16] also complement the study to increase the number of validation data.

The Section 1 describes the HYPSTAR sensor and measured data as well as the processing used to estimate water-leaving reflectance. The water sampling data used for the validation of the $Ch{l}_a$ algorithms are described in Section 2.2. Section 2.3 describes the algorithms used for the retrieval of the water quality products, i.e., $Ch{l}_a$ and $SPM$. Improvements to those $Ch{l}_a$ algorithms are presented in Section 3.2. The retrieved long-term $Ch{l}_a$ and $SPM$ time series together with the water sampling data and some concurrent satellite images are shown in Section 3.3. In Section 3.4, we investigate the potential to detect phycocyanin with the HYPSTAR data. Finally, in Section 4, we discuss the possible improvements to further exploit the data and how the HYPSTAR water quality products could be used for operational water quality assessment in conjunction with satellite images.

2. Materials and Methods

2.1. Above Water Reflectance Data

2.1.1. HYPSTAR

A Hyperspectral Pointable System for Terrestrial and Aquatic Radiometry, HYPSTAR (www.HYPSTAR.eu, accessed on 14 October 2022), was installed at Blankaart WPC reservoir in January 2021 and continuously measures (except during short periods of downtime for maintenance) the upwelling and downwelling radiance and downwelling irradiance during daylight (7 a.m. till 6 p.m. local time). The HYPSTAR installed at Blankaart (Figure 1) measures the reflected light within the visible and near-infrared (VNIR) spectral range (380–1020 nm) with a spectral sampling and resolution of 0.5 and 3 nm, respectively, and a field of view of 2°. The HYPSTAR is mounted on a pan and tilt unit allowing commands to be received from the host system and to return measurements according to the defined measurement protocol (i.e., viewing geometries, number of scans, radiance or irradiance measurements, etc.). The system also includes a GPS and embedded camera to visualize the target and sky conditions for visual inspection. A true color Sentinel-2B image and the panchromatic GEOEye image acquired over the reservoir show the octagonal shape of the 800 m-diameter drinking water reservoir (see Figure 1). The reservoir has a concrete wall of 400 m built from its north-east edge to its center. The HYPSTAR is deployed on this wall, 30 m from the external border. The HYPSTAR has been developed in the frame of the H2020 HYPERNETS project (www.hypernets.eu, accessed on 14 October 2022), and is part of the HYPERNETS network [22].

2.1.2. Acquisition Protocol and Data Processing

The above-water approach includes measurements from above-water radiometers for $L_d(\lambda ,180-\theta ,\Delta \varphi ,{\theta }_0)$, $L_u(\lambda ,180-\theta ,\Delta \varphi ,{\theta }_0)$ and $E_d\left(\lambda \right)$, the downwelling radiance, upwelling radiance and downwelling irradiance, respectively. In the absence of sun glint, (near-) simultaneous $L_d(\lambda ,180-\theta ,\Delta \varphi ,{\theta }_0)$ and $L_u(\lambda ,180-\theta ,\Delta \varphi ,{\theta }_0)$ measurements allow the evaluation of the water-leaving radiance, $L_w\left(\lambda \right)$ as follows:

$$L_w(\lambda ,180-\theta ,\Delta \varphi ,{\theta }_0)=L_u(\lambda ,180-\theta ,\Delta \varphi ,{\theta }_0)-{\rho }_F(\lambda ,\theta ,\Delta \varphi ,{\theta }_0,ws,E)L_d(\lambda ,180-\theta ,\Delta \varphi ,{\theta }_0)$$

(1)

The term ${\rho }_F$ is the air–water interface reflectance coefficient expressed as a function of viewing geometry, i.e., $\theta$, $\Delta \varphi$, and sun zenith angle ${\theta }_0$, and environmental factors, E. When the water surface is perfectly flat, ${\rho }_F$ is the Fresnel reflectance and the environmental factor E only depends on the relative refractive index of the air–water interface. When the water is not perfectly flat, ${\rho }_F$ needs to account for the geometric effects of the wave facets created by the roughened water surface. Wave facets change the water surface geometry and thus (1) the upwelling rays by changing the relative geometry between the illumination and viewing direction, as well as (2) the contribution of the sky and sun glint to $L_u(\lambda ,180-\theta ,\Delta \varphi ,{\theta }_0)$. After normalization by $E_d\left(\lambda \right)$, $L_w\left(\lambda \right)$ becomes the remote sensing reflectance, $R_{rs}\left(\lambda \right)$ (expressed in $s{r}^{-1}$), or, when multiplied by $\pi$, the water-leaving reflectance, ${\rho }_w\left(\lambda \right)$ (dimensionless). These are the standard input for most optical models used to derive bio-geochemical products.

Numerous measurement protocols exist for the correction of the air–water interface reflectance for above-water acquisitions (e.g., [6] and references therein). However, the present study does not intend to provide an inter-comparison of these methods and follows, therefore, the approach adopted by NASA [24] and IOCCG [25] for the measurement protocol and data processing. Radiometer measurements are taken with the HYPSTAR in a defined set of geometries called a sequence. Each geometry in a sequence is called a series, composed of a set of repeated scans that are averaged. Thus, a single sequence results in one measurement of water-leaving reflectance. A sequence was taken every 15 to 30 min during daylight and contained the following series:

• Three scans of $E_d(\lambda ,\theta ={180}^{\circ },\Delta \varphi =+/-{90}^{\circ })$;
• Three scans of $L_d(\lambda ,\theta ={140}^{\circ },\Delta \varphi =+/-{90}^{\circ })$;
• Six scans of $L_u(\lambda ,\theta ={40}^{\circ },\Delta \varphi =+/-{90}^{\circ })$;
• Three scans of $L_d(\lambda ,\theta ={140}^{\circ },\Delta \varphi =+/-{90}^{\circ })$;
• Three scans of $E_d(\lambda ,\theta ={140}^{\circ },\Delta \varphi =+/-{90}^{\circ })$.

2.1.3. Data Processing

Data are processed using the Hypernets Processor [22], which is fully accessible via (https://github.com/HYPERNETS/hypernets_processor, accessed on 14 October 2022). The first step in the processor is to check the quality of the raw radiance and irradiance data and the blacks. Next, raw scans are corrected for darks and calibrated to irradiance and radiance scans. For the computation of the water reflectance, the HYPERNETS processor includes the RHYMER component (i.e., “Reliable processing of HYperspectral MEasurement of Radiance”). RHYMER provides all the required functions to process the above-water measurements and is written as such that it can easily welcome any additional look-up tables, processing functions, and/or quality flags. Before averaging per series, RHYMER first assigns flags to each scan showing a temporal jump, with the previous and following scan, of more than 25 at 550 nm. Next, RHYMER parses the cycle, i.e., it separates upwelling and downwelling radiance, investigates if all required angles are present and if there are coincident upwelling and downwelling radiance measurements for the retrieval of the water-leaving radiance. Since the irradiance and radiance measurements have a slight shift in wavelength, a spectral interpolation is first performed between radiance and irradiance spectra. Next, $E_d\left(\lambda \right)$ and $L_d\left(\lambda \right)$ scans are averaged per series and temporally interpolated to $L_u\left(\lambda \right)$ scans. For the computation of the water-leaving reflectance, the reflectance factor ${\rho }_F$ is retrieved from the look-up table of [26] using as input the solar zenith angle, viewing zenith angle, relative azimuth angle between sun and sensor, and the wind speed, at 10 m above the sea surface.

In the present study, wind speed is taken from a nearby station, i.e., the VLINDER station located at Blankaart. The VLINDER data, (www.vlinder.ugent.be, accessed on 14 October 2022) [27], provide near-real-time weather data including temperature, rainfall, wind speed, and wind direction.

To ensure quality reflectance data and discard any suspect spectra (e.g., noisy or non-water spectra, high variation in illumination during the sequence, and/or inconsistency between radiance and irradiance measurements during the sequence), three additional quality checks are used. A spectrum is discarded if one of these conditions are met:

• Filter 1: the ratio of $L_u\left(\lambda \right)/E_d\left(\lambda \right)$ between 800 and 950 nm is greater than 0.025 sr${}^{-1}$;
• Filter 2: the NIR similarity correction factor, $ϵ(720,780)$, exceeds a given threshold (with $ϵ(720,780)$ defined by Equation (8) in [28]);
• Filter 3: negative reflectance values between 400 and 900 nm.

The first filter relies on the fact that the water radiance is expected to be very low in the 800–950 nm spectral range as most of the water absorbs in these wavelengths. As observed by Groetsch et al. [29], applying filter 1 removes most of the spectra affected by foam-caps or surface scum. We found that this empirical threshold also removed most of the suspect spectra in our dataset without considering inaccuracies in the data processing.

The second filter was used to remove spectra for which we suspect erroneous processing. According to Ruddick et al. [30], the shape of the NIR spectrum is mainly determined by the pure water absorption and is thus invariant. Hence, for waters following this theory, $ϵ(720,780)$ is the residual error when the data processing results in an inaccurate correction for the presence of high sun glint. Despite the 90° relative azimuth between sun and sensor, $\Delta \varphi$, during acquisition attempts to avoid sun glint, over wind-roughened surfaces, sun glint may still be present in the field of view of the sensor. Therefore, spectra for which the estimated $ϵ(720,780)$ exceeds a given threshold are considered as inaccurately processed and removed for further data analysis. The threshold (see Section 3.1) of $ϵ(720,780)$ is defined based on a visual inspection of the removed and remaining spectra.

The last filter removes the negative reflectance spectra. Negative reflectance often results from an inaccurate removal of the air–water interface reflectance. The estimation of ${\rho }_F$ in Equation (1) is indeed often a source of error when, among other things, the wave slope distribution at the surface is unknown and/or there are above-water optical perturbations [6]. The filters suggested here present a good compromise between the number of data to retain (i.e., allowing a high temporal resolution time series) and the autonomous quality assessment of the data (i.e., essential for autonomous systems recording very high amounts of data).

2.2. Validation of Data

2.2.1. $Ch{l}_a$ Concentrations

Surface water was sampled with a 5 L bucket on irregular days by the water managers for monitoring the water quality. Although it was not meant to be used for the validation of the water quality products from the HYPSTAR, it appeared to be very relevant as the sampling was conducted from the same platform where the HYPSTAR was installed. After sampling, $Ch{l}_a$ concentrations were determined with a spectrophotometric method, 10200 H [31].

In total, we found 43 $Ch{l}_a$ concentrations sampled during the time interval of the HYPSTAR deployment (i.e., since January 2021).

2.2.2. Counting Procedure for Cyanobacteria and Diatoms

For the manual counting of the algae, the sampling protocol and preparation followed the “Water Sampling, Measurement and Analysis Compendium (WAC)” from the Flemish Government [32] and the European standard EN 15204 [33].

2.3. Algorithms for Water Quality Products

2.3.1. $Ch{l}_a$

Both multispectral [14] and hyperspectral [34] red-edge approaches are tested and compared with the water sampling retrieved $Ch{l}_a$. The red-edge spectral region is well-suited for detecting $Ch{l}_a$ in turbid and eutrophic waters thanks to the combination of the high $Ch{l}_a$ absorption around 672 nm and its strong scattering around 705 nm [35]. Blue-green band ratio algorithms and quasi-analytical algorithms, often referred to as QAA [36], or the generalized inherent-optical-properties inversion [37] are less suited for these inland waters because of the significant influence of the dissolved and particulate material in blue and green bands. The quasi-analytical algorithms were generally developed for clear waters to moderately turbid waters.

Among the wide variety of red-edge-based algorithms, the algorithm proposed by Simis et al. [14] (referred to hereafter as the SIMIS algorithm) was selected because it is very close to the semi-empirical algorithm introduced by Gons [38]. This algorithm is also associated with a phycocyanin algorithm which is assessed here to monitor cyanobacteria blooms. The algorithm suggested by Ruddick et al. [34], also referred to as the “Chlorophyll a Retrieval using an Adaptive Two-band” algorithm (referred to hereafter as the CRAT algorithm), has been selected to investigate the benefits of the hyperspectral resolution. Although the CRAT algorithm is only based on two spectral bands, the spectral position of the second band varies from one spectra to another and can only be determined from hyperspectral data. Both algorithms assume (1) a constant back-scattering coefficient, $b_b$, in the red and NIR, (2) a negligible absorption from colored dissolved organic matter (CDOM) and non-algal particles compared with water absorption ($a_w$) at 672 nm, 704 nm and 750 nm and to $Ch{l}_a$ absorption at 672 nm, and (3) negligible absorption by phytoplankton at 704 nm and 750 nm.

With the SIMIS algorithm, $Ch{l}_a$ concentrations are estimated from the $Ch{l}_a$ absorption at 665 nm, $a_{Chl}\left(665\right)$, as follows:

$$a_{Chl}\left(665\right)=\frac{\frac{{\rho }_w\left(709\right)}{{\rho }_w\left(665\right)}(a_w\left(709\right)+b_b)-b_b-a_w\left(665\right)}{\gamma }$$

(2)

where $\gamma$ is a factor applied to correct underestimation due to the hypothesis of negligible CDOM and non-algal particulate absorption. Simis et al. [14] estimated $\gamma$ to be equal to 0.68 for the Ijselmeer. However, because this parameter is very regional, it is set to 1 in the present study, as suggested by Gons [38]. $b_b$ is calculated from ${\rho }_w\left(779\right)$ as proposed in Gons et al. [39] and the pure-water absorption coefficients, $a_w\left(\lambda \right)$, are retrieved from Buiteveld et al. [40]. $Ch{l}_a$ absorption is then converted into $Ch{l}_a$ concentration using the $Ch{l}_a$-specific absorption coefficient, $a^{*}ph\left(665\right)$:

$$Ch{l}_a^S=\frac{a_{Chl}\left(665\right)}{a^{*}ph\left(665\right)}$$

(3)

$a^{*}ph\left(\lambda \right)$ is the absorption coefficient of phytoplankton per unit of $Ch{l}_a$ concentration. Simis et al. [14] suggests a constant $a^{*}ph$(665) equal to 0.0153 m${}^2$m/g for the IJselmeer.

With the CRAT algorithm, $Ch{l}_a$ is calculated differently for each spectrum as:

$$Ch{l}_a^C=\frac{a_w\left({\lambda }_2\right)-a_w\left({\lambda }_1\right)}{a\ast ph\left(672\right)}$$

(4)

where ${\lambda }_2$ is selected individually in the 704–740 nm interval as the wavelength at which the reflectance is equal to the reflectance at the red $Ch{l}_a$ absorption peak (i.e., ${\rho }_w\left({\lambda }_2\right)$ = ${\rho }_w\left({\lambda }_1\right)$). ${\lambda }_1$ is equal to 672 nm [34].

Compared with the SIMIS algorithm (and similar algorithms, see references in [14]), the CRAT is based on a subtraction which avoids the calculation of $b_b$. In addition, residual absorption from CDOM and non-algal particles are also partially corrected [34].

$a*ph\left(\lambda \right)$ values are known to vary as a function of a variety of factors, such as phytoplankton-specific pigment composition or the water trophic state [41]. As mentioned by Bramich et al. [42], and references therein, the use of a varying $a^{*}ph\left(\lambda \right)$ is therefore expected to improve the $Ch{l}_a$ concentration estimates, in particular for the lowest and highest concentrations. Therefore, Bricaud et al. [41] recommended using a power function of the form:

$$a^{*}ph\left(\lambda \right)=ACh{l}_a^{-B}$$

(5)

where A and B are wavelength-specific coefficients (A reflects $a^{*}ph\left(\lambda \right)$ per $Ch{l}_a$ unit while the B exponent reflects the deformations of $a^{*}ph\left(\lambda \right)$ with an increase in $Ch{l}_a$ concentrations). Locally tuned A and B coefficients have already been used by numerous studies (e.g., [43,44]). We therefore also analyzed how using a varying $a^{*}ph\left(\lambda \right)$ could improve the accuracy of $Ch{l}_a^S$ and $Ch{l}_a^C$ by deriving a linear regression between the logarithm of $a^{*}ph\left(\lambda \right)$ and logarithm of $Ch{l}_a$ from the water sampling, $Ch{l}_a^W$. Hence, the slope of the linear regression is equal to B in Equation (5) and the intercept to log(A).

Ultimately, we compared the accuracy of $Ch{l}_a^S$ and $Ch{l}_a^C$ with a constant $a^{*}ph\left(\lambda \right)$ and with a varying $a^{*}ph\left(\lambda \right)$ (as in Equation (5)). To evaluate their performance, we used the following metrics.

• The Root Mean Square Error (RMSE), which measures the scatter of the data from the regression line (units in $\mathsf{\mu }$g/L);
• The slope (S) and intercept (I) of the least squares regression to detect systematic multiplicative or additive biases;
• The Mean Absolute Percentage Error ($MAPE$, unsigned and units in %) between the HYPSTAR-retrieved $Ch{l}_a$, $Ch{l}_a^C$ or $Ch{l}_a^S$, and water-sampling-retrieved $Ch{l}_a$ and $Ch{l}_a^W$, for a total number of samples n. $MAPE$ is calculated as follows:

  $$MAPE=100\times \frac{1}{n}{\Sigma }_{i=1}^{i=n}\frac{|Ch{l}_a^{C,S}-Ch{l}_a^W|}{Ch{l}_a^W}$$

  (6)
• The $BIAS$, to assess the systematic errors in the algorithm outputs (units in $\mathsf{\mu }$g/L), calculated as follows:

  $$BIAS=\frac{1}{n}{\Sigma }_{i=1}^{i=n}(Ch{l}_a^{C,S}-Ch{l}_a^W)$$

  (7)

2.3.2. $SPM$

For the computation of $SPM$, the algorithm from Nechad et al. [45] was used with the measured water-leaving reflectance at 700 nm as follows:

$$SPM=\frac{A^{N2010}{\rho }_w\left(700\right)}{1-\frac{{\rho }_w\left(700\right)}{C^{N2010}}}$$

(8)

where $A^{N2010}$ and $C^{N2010}$ are constant values (i.e., 445.11 and 1.13, respectively) taken from

Table 1. Constant A and B values estimated in the present study with the SIMIS algorithm and found in the literature [41,44] for $\lambda$ equal 665 and 672 nm.

	$\mathit{\lambda }$ = 665 nm		$\mathit{\lambda }$ = 672 nm
	A	B	A	B
SIMIS	0.0537	−0.3897	–	–
Paavel	0.0311	−0.2711	0.0339	−0.2488
Bricaud	0.0150	−0.1333	0.0197	−0.1530

in [45]. This algorithm has already been used and validated by numerous studies (e.g., [46,47]).

2.3.3. Cyanobacteria

Since the match-ups between water sampling points and HYPSTAR reflectance data were not sufficient to properly validate or design a cyanobacteria algorithm (only two match-ups), we limited our results to the evaluation of several indexes. We also complemented the dataset with two extra PANTHYR measurements [16] taken at Blankaart during large cyanobacteria blooms in 2019 and 2020.

Several indexes have been developed from cyanobacteria spectral specificities due to PC absorption and cyanobacteria fluorescence anomalies. Using an analytical algorithm, Simis et al. [14] proposed calculating the concentration in PC from its absorption at 620 nm and after correction for water and $Ch{l}_a$ absorption (i.e., estimated with the SIMIS algorithm in Section but using the 620 nm reflectance band). This first index, referred to as $CI$1, is estimated as follows:

$$CI1=\frac{\left[\frac{{\rho }_w\left(709\right)}{{\rho }_w\left(620\right)}(a_w\left(709\right)+b_b)\right]-b_b-a_w\left(620\right)}{\delta }-{ϵ}^{665-620}\times a_{Chl}\left(665\right)$$

(9)

with ${ϵ}^{665-620}$ being the conversion factor from $a_{Chl}\left(665\right)$ to $a_{Chl}\left(620\right)$ (equal to 0.24, [14]), and $\delta$ being the correction factor similar to $\gamma$ in Equation (2). In the present study, $\delta$ is set to 1 for the same reasons as mentioned above.

The second index, referred to as $CI$2, is the band ratio between ${\rho }_w\left(600\right)$ and ${\rho }_w\left(700\right)$ as suggested by Mishra et al. [15]:

$$CI2=\frac{{\rho }_w\left(700\right)}{{\rho }_w\left(600\right)}$$

(10)

These bands were chosen for their low variability (minimum reflectance at 700 nm) and for the absence of influence of $Ch{l}_a$ absorption at 600 nm while phycocyanin is still absorbing. Hence, the higher the index, the higher the probability of cyanobacteria presence.

The third index, referred to as $CI$3, is proposed by Wynne et al. [48] and is similar to a fluorescence line height equation with a focus on the spectral shape around 681 nm. According to the authors, a deficit of fluorescence at 681 nm is expected in the presence of cyanobacteria compared with other phytoplankton groups. A negative value for CI3 is expected in the presence of cyanobacteria whereas a positive value is expected in the absence of cyanobacteria.

$$CI3={\rho }_w\left(681\right)-{\rho }_w\left(665\right)-[{\rho }_w\left(709\right)-{\rho }_w\left(665\right)]\left[\frac{681-665}{709-665}\right]$$

(11)

3. Results

3.1. Water-Leaving Reflectance

The dataset consists of 8116 spectra measured in the period 2021-01-27 to 2022-08-03 (Figure 2). A first quality check is performed to filter out the suspect data (see filter 1 in Section 2.1.2), reducing the number of valid data to 7933. Data with $ϵ(720,780)$ greater than 0.005 (filter 2) are then removed, leaving 6535 spectra. The threshold of 0.005 is large enough to retain the optically complex waters and severe enough to remove non-water or glint-affected spectra. Finally, all spectra showing (non-physical) negative values between 400 and 900 nm are removed, leaving 2988 individual measurements in the dataset, ranging from 2021-02-03 to 2022-08-03. More data are removed during the winter (high wind, rain, low illumination, changing weather, etc.) with the number of sequences retained in December, January and February being 38, 74 and 89, respectively. In the summer (June, July and August) the number of sequences varies between 622 and 307 sequences per month.

Figure 3 shows the daily mean reflectance of all the retained sequences over the different seasons. As expected, all spectra show typical spectral shapes for inland waters, i.e., low reflectances in the blue compared with reflectances in the green and red spectral range [49]. The magnitude of the reflectance value is higher in the winter season compared with the rest of the year. This is especially visible for the 500–600 nm range and results from higher particle concentrations in the winter months or the resuspension of sediments in the reservoir due to high winds. Several distinguishing finer spectral features are also observed. First, the so called red-edge reflectance (i.e., ∼680–750 nm) is visible, essentially in spring and summer. This increase in reflectance can be associated with increasing $Ch{l}_a$ concentrations [50]. Likewise, a clear U-shape is present for many spectra in the 650–710 nm range in spring and summer, but less in winter and autumn. This U-shaped feature is a typical indication of high particle scattering bounded by high pigment absorption ([49], and references therein). These spectra are expected to present high $Ch{l}_a$ concentrations and the possible presence of cyanobacteria [48,51].

3.2. $Ch{l}_a$ Retrieval

Out of the 43 water sampling points taken between 2021-01-27 and 2022-08-03, 27 $Ch{l}_a^W$ concentrations were taken during a HYPSTAR reflectance spectrum (i.e., passing the quality check mentioned above) (see Figure 4). Concentrations range from 1.5 $\mathsf{\mu }$g/L to 39.4 $\mathsf{\mu }$g/L. As mentioned, a U-shaped spectra around 650–670 nm is clearly visible for the higher $Ch{l}_a$ concentrations, typical for high phytoplankton absorption. Note, however, that absorption at 670–675 nm is not only related to $Ch{l}_a$ but may also be related to suspended particulate absorption [34]. Similarly, higher $Ch{l}_a$ concentrations tend to show an increase in the reflectance peak around 710–720 nm and 555 nm.

Figure 5 shows the accuracy of the retrieved $Ch{l}_a$ concentrations using the CRAT and SIMIS algorithms, respectively. Note that the CRAT algorithm using the wavelength range 704–740 nm is not expected to perform well for $Ch{l}_a$ values outside the ∼10–100 $\mathsf{\mu }$g/L range, as reported by Ruddick et al. [34]. Therefore, statistics are limited here for $Ch{l}_a$ concentration within this range, reducing the number of match-ups between water sampling and HYPSTAR data to 11.

An $a^{*}ph\left(672\right)$ of 0.0177 ± 0.00151 m${}^2$mg${}^{-1}$ is found to be the most appropriate for the CRAT algorithm. An $a^{*}ph\left(665\right)$ of 0.01995 ± 0.00165 m${}^2$mg${}^{-1}$ performs better for the SIMIS algorithm. These values are in line with what is found in the literature (i.e., [34] found an $a^{*}ph\left(672\right)$ of 0.0205 m${}^2$mg${}^{-1}$ for the IJssel lagoon with the CRAT algorithm, and [14] an $a^{*}ph\left(665\right)$ between 0.0153 and 0.0203 m${}^2$mg${}^{-1}$ for different lakes with the SIMIS algorithm).

The RMSEs are 8.20 $\mathsf{\mu }$g/L and 5.61 $\mathsf{\mu }$g/L for the CRAT and SIMIS algorithm, respectively. For the CRAT algorithm, we observe a MAPE of 28.5 % and a BIAS of 1.27 $\mathsf{\mu }$g/L. For the SIMIS algorithm, MAPE and BIAS are 75.6% and 0.16 $\mathsf{\mu }$g/L, respectively. Both algorithms tend to overestimate the lower $Ch{l}_a$ concentrations and underestimate the higher concentrations. The minimum $Ch{l}_a$ concentration retrieved with the CRAT algorithm is 12.14 $\mathsf{\mu }$g/L (Figure 5). Using an $a^{*}ph\left(672\right)$ of 0.0177 m${}^2$mg${}^{-1}$ and with ${\lambda }_1$ equal to 672 nm and ${\lambda }_2$ in the 704–740 nm wavelength range, the retrieval range of the CRAT algorithm is between 12.14 $\mathsf{\mu }$g/L and 119.56 $\mathsf{\mu }$g/L $Ch{l}_a$ (i.e., between [$a_w\left(704\right)-a_w\left(672\right)$]/0.0177 and [$a_w\left(740\right)-a_w\left(672\right)$]/0.0177 with $a_w\left(\lambda \right)$ equals 0.415, 0.6303 and 2.5319 m${}^{-1}$, at 672, 704 and 740 nm, respectively).

The value of $a^{*}ph\left(\lambda \right)$ decreases from 0.08 to 0.006 m${}^2$/mg at the red $Ch{l}_a$ absorption peak (i.e., 665 nm, Figure 6a). At 672 nm, $a^{*}ph\left(\lambda \right)$ ranges from approximately 0.04 to 0.006 m${}^2$/mg. Note that the high $a^{*}ph\left(\lambda \right)$ values should not be considered as these correspond to $Ch{l}_a$ concentrations below the retrieval range (i.e., see points before gray dashed line in Figure 6b). Table 1 provides the values for A and B in Equation (5) when retrieved with the SIMIS algorithm. For comparison, A and B values reported by Bricaud et al. [41] and Paavel et al. [44] are also added. Since the number of match-ups with $Ch{l}_a$ concentrations above 8 $\mathsf{\mu }$g/L is limited (i.e., 11 out of 27), we used the A and B values from Paavel et al. [44] for the CRAT algorithm. These constant values for A and B were estimated with 155 water samples taken in different lakes (including cyanobacteria-dominated lakes) for $Ch{l}_a$ concentrations ranging from 3 to 315 $\mathsf{\mu }$g/L [44]. Overall, this relationship fits well with our data for $Ch{l}_a$ concentrations above 10 $\mathsf{\mu }$g/L (see Figure 6b).

For the SIMIS algorithm, the performance of the algorithm increases when using a variable $a^{*}ph\left(\lambda \right)$ (Figure 5). MAPE and BIAS decrease to 55.51 % and 1.04 $\mathsf{\mu }$g/L, respectively. For the CRAT algorithm, the MAPE and RMSE increase, but the regression line is closer to the 1:1 line (slope and intercept approaches 1 and 0, respectively).

3.3. $Ch{l}_a$ and $SPM$ Time Series

The HYPSTAR-retrieved $Ch{l}_a^S$ and $Ch{l}_a^C$ follow closely the water sampling $Ch{l}_a^W$ concentrations (Figure 7), although $Ch{l}_a^S$ and $Ch{l}_a^C$ show higher peaks in $Ch{l}_a$ concentrations relative to the water sampling data. These peaks may be explained by (1) the high temporal resolution of the HYPSTAR data depicting $Ch{l}_a$ variations that are omitted by the water sampling data, (2) failure of the quality check of the reflectance data to remove some non-valid or suspect reflectance spectra, and/or (3) the erroneous estimation of $a^{*}ph\left(\lambda \right)$ at high $Ch{l}_a$ concentrations. As mentioned in Section 3.1 and illustrated in Figure 3, we indeed observe overall lower $Ch{l}_a$ concentrations in autumn (although some high $Ch{l}_a$ values are seen early September) and relatively higher values in the spring and summer. Note that during the spring and summer months, measures are taken to prevent algal blooms (i.e., the injection of algaecide) within the reservoir. This could explain why the $Ch{l}_a$ concentrations are limited, in particular in the summer of 2021, and why in some cases sudden drops in $Ch{l}_a$ concentrations are observed (e.g., in July 2022).

Figure 8 shows the $Ch{l}_a$ values over a shorter time period and the corresponding reflectance spectra for the given $Ch{l}_a$ values. A peak in $Ch{l}_a$ concentration was observed around 2021-04-30 (Figure 8a). Afterwards, the $Ch{l}_a$ concentration slowly decreased until 2021-05-05 and remained low until the end of May. In early June, the $Ch{l}_a$ concentration increased again up to ∼40 $\mathsf{\mu }$g/L. Spectra associated with higher $Ch{l}_a$ values (i.e., 2021-05-03 and 2021-06-02) can clearly be distinguished by their pronounced reflectance peaks at 700 nm and deep U-shaped features between 640 and 700 nm (Figure 8b). Differences in spectral shapes between 2021-05-03 and 2021-06-02 may be related to the higher abundance of diatoms on 2021-05-03 (i.e., 29.8 mm${}^3$/L compared with 7.8 mm${}^3$/L on 2021-06-02). On 2021-05-03, we also observe two absorption features around 490 nm and 620 nm typical for $Ch{l}_c$ and the possible presence of PC. However, according to the water sampling data, there was no presence of cyanobacteria at that time.

$SPM$ values range from 0 to 30 g/m${}^3$ with some peaks reaching 60 g/m${}^3$ (Figure 9). There were no coincident data with the water sampling points for $SPM$. However, for both peaks in $SPM$, i.e., around 2022-01-09 and 2022-02-07, a Landsat-9 OLI and Sentinel 2B-MSI image confirmed the relatively high plume of sediments along the southern side of the wall (see Figure 10).The timing of these sediment plumes also corresponded to storm events (see high gustwind in Figure 9). Storm events may induce high sediment concentrations in the water that is pumped into the reservoir, and/or the resuspension of the sediments present in the bottom of the water reservoir. The latter explanation may be more valid since during heavy storms, and especially during flooding, the intake of water into the reservoir is stopped.

Relatively high values of reflectance, in particular in the red and NIR spectral range, are observed for those two dates (Figure 10). $SPM$ values retrieved from the reflectance spectra taken the same day (i.e., 2022-02-07) or the day before (i.e., 2022-01-10) are also relatively high, with concentrations of 27.59 and 33.07 g/m${}^3$. For comparison, two spectra corresponding to high $Ch{l}_a$ values (see Figure 7) and low $SPM$ values (see Figure 9), i.e., 2021-03-06 and 2022-07-17, are also shown (Figure 10).

Table 2. Mean and standard deviation of $SPM$ and $Ch{l}_a$ concentrations retrieved with the CRAT and SIMIS algorithm on different dates (see orange bars in Figure 7 and Figure 9, and associated reflectance spectra in Figure 10).

Date	Numb. Seq.	SPM (g/m3)	${Chl}_a^C$ (μg/L)	${Chl}_a^S$ ($\mathsf{\mu }$g/L)
2021-03-06	5	8.78 ± 0.07	86.88 ± 3.47	60.60 ± 1.85
2022-02-07	18	27.59 ± 2.92	–	0.14 ± 0.12
2022-01-10	2	33.07 ± 4.74	–	2.52 ± 0.13
2022-07-17	13	7.75 ± 0.57	207.25 ± 5.5	142.94 ± 11.71

reports, for these four dates, the number of sequences per day (after quality check), the estimated $SPM$ values, and the retrieved $Ch{l}_a^S$ and $Ch{l}_a^C$ concentrations.

3.4. Cyanobacteria Detection: A Feasibility Study

To evaluate the different PC indexes, extra hyperspectral remote sensing data, acquired by a PANTHYR sensor [16], were also analyzed as they provided some concomitant water reflectance spectra during large cyanobacteria bloom events. Indeed, in 2021–2022, Blankaart reservoir experienced only small cyanobacteria bloom events with a maximum of 5 $\mathsf{\mu }$g/L observed from water samples on 2021-09-06 and 2021-09-21. Figure 11 displays water reflectance spectra and the second derivative from a diatom bloom (2021-05-02) and from cyanobacteria blooms (2019-08-01, 2020-06-15 and 2021-09-06, see also

Table 3. Phytoplankton parameters measured in the Blankaart reservoir at the same date as water reflectance spectra displayed in Figure 9 and cyanobacteria indexes ($CI$) calculated from the water reflectance. See text for details about the different $CI$.

Date	Sensor	Diatoms ($\mathsf{\mu }$g/L)	Cyano (μg/L)	Chla (μg/L)	CI1	CI2	CI3
2019-08-01	PANTHYR	0	41.58	328.6	426.41	2.073	−0.020
2020-06-15	PANTHYR	0	11.80	32.5	82.57	1.046	−0.007
2021-05-02	HYPSTAR	29.18	0	39.4	22.15	0.777	−0.001
2021-09-05	HYPSTAR	0	4.7	24.2	37.95	0.86	0.00

). From these spectra, the different indexes are calculated.

$CI$1, as suggested by [14], shows increasing values when the concentration of cyanobacteria increases. In particular, when cyanobaceria concentration increases from 0 to 4.70 $\mathsf{\mu }$g/L, $CI$ increases from 22.15 to 37.95, suggesting that $CI$1 is able to detect small increases in cyanobacteria concentration. However, in the absence of cyanobacteria, it is higher than 0, which suggests an offset problem. This erroneous offset may be explained by the need for accurate local parameterization [52].

The second index, $CI$2, uses the band ratio between water reflectance at 600 nm and 700 nm [15]. Figure 11 confirms the interest of the 600 nm spectral band as the second derivative is positive in the presence of cyanobacteria and negative otherwise. In fact, a positive second derivative means that the curve is convex, which is consistent with the PC pigment absorption. As for $CI$1, one can notice that $CI$1 increases with the concentration of cyanobacteria.

The third index, $CI$3, is expected to have a negative value in the presence of cyanobacteria and a positive value otherwise [48]. Our results show that $CI$3 is not able to detect cyanobacteria blooms with a concentration of 4.7 mg m${}^{-3}$, as $CI$3 is null in these cases, whereas it is slightly negative when no cyanobacteria are observed. However, for higher concentrations of cyanobacteria, $CI$3 is significantly lower, suggesting that this index can only be used to detect important changes in cyanobacteria concentrations. As the spectral region (i.e., 665–709 nm) is also highly impacted by $Ch{l}_a$ absorption, it is possible that variability in $Ch{l}_a$ concentration also affects $CI$3.

4. Discussion

4.1. Further Improving HYPSTAR Derived Water Quality Products

The results above confirm that the HYPSTAR system provides valuable data at very high temporal resolution with low manpower requirements. Currently, the data are processed automatically and shared with the water managers for monitoring trends of $Ch{l}_a$ in the water reservoir and to investigate the effectiveness of the strategies to control algal blooms. Hence, the HYPSTAR can be used as a quick and cost-effective method to monitor water quality and detect algal blooms, provided that an initial phase of quality check is performed followed by a tuning and validation of the water quality algorithms. However, further improvement in the data processing may aid in reducing the inaccuracies in water quality products, and increase the number of data. For instance, improving the correction for the air–water interface reflectance will reduce the number of spectra removed by the quality checks (e.g., negative reflectance). The present study relies on external wind speed reflectance data to correct the upwelling radiance for air–water interface reflectance and retrieves the air–water interface reflectance factor, ${\rho }_F$ in Equation (1), based on the look-up tables of [26]. However, radiative transfer simulations were made by [26] with simulated sea surfaces, where the wave slope statistics match those of the Cox–Munk model [53] based on an input wind speed. These wave models assume a fully developed wind wave sea, where the wave field is in equilibrium with the local wind. This assumption is often not valid, especially for fetch-limited conditions typical of inland waters. In addition, the 400 m long wall within the reservoir significantly affects the wind at the water surface. Hence, the VLINDER wind speed may differ from the wind speed that should be considered to obtain the proper wave slope statistics at the surface. To increase the accuracy of the reflectance data, the HYPSTAR should be placed further away from the wall (i.e., increase the distance between the wall and the field-of-view of the sensor), and the processing could be improved by testing other methods to correct for the air–water interface reflectance (e.g., see [6] and references therein).

The results show that the retrieved $Ch{l}_a$ concentrations closely follow the water sampling data. However, more coincident HYPSTAR and water sampling data are required to further validate the algorithms and cover a larger range of $Ch{l}_a$ concentrations. Indeed most values for the validation of the algorithms are below 15 $\mathsf{\mu }$g/L, which is not considered to be harmful for the drinking water production. In addition, as mentioned earlier, the CRAT algorithm with ${\lambda }_2$ retrieved in the 704–740 nm spectral range is not expect to perform well outside the ∼12–120 $\mathsf{\mu }$g/L. Hence, a proper validation of the CRAT algorithm requires additional data in this concentration range. For higher concentrations of $Ch{l}_a$ (i.e., >100 $\mathsf{\mu }$g/L), we could also test the same algorithm but using ${\lambda }_2$ >= 825 nm, as suggested by Ruddick et al. [34].

The rapid testing of different cyanobacteria indexes confirms that the data from autonomous hyperspectral radiometers have the potential to monitor the real-time presence/ absence of cyanobacteria. However, a larger dataset of coincident water sampling and radiometric data is required to further validate these algorithms and possibly quantify the cyanobacteria biomass. Indeed, in our dataset only very moderated cyanobacteria blooms were recorded during the HYPSTAR deployment, while strong cyanobacteria bloom events were recorded during the PANTHYR deployment. However, the number of valid matchups between in situ samples and radiometric data was low (i.e., 4). The results only allow us to analyze certain spectra and test some algorithms without being able to provide a proper calibration. Nevertheless, the results tend to confirm the feasibility of monitoring cyanobacteria bloom with both HYPSTAR and PANTHYR. Since water sampling and HYPSTAR are still ongoing, the number of match-ups is expected to increase and new algorithms may be developed in the future using the very high spectral resolution of the HYPSTAR. Indeed, by focusing on very small regions of the spectra, we could limit the impact of non-specific pigments (e.g., using the spectral position of the second-derivative minimum or maximum [54]).

4.2. Early Warnings and Spatial Monitoring

With its very high temporal resolution, the autonomous hyperspectral radiometers may be used to obtain first warnings and, when required, to take actions before bloom. In addition to the water quality products mentioned here, i.e., $Ch{l}_a$ and $SPM$, radiance and irradiance data may therefore also be used to approximate environmental conditions known to influence the development of algal blooms. For instance, downwelling irradiance measurements can be used as a proxy for air temperature and sun illumination. Upwelling radiance measurements taken within the principal plane may also be used to estimate the wind speed at the air–water surface [55], and, subsequently, serve to investigate the degree of water column stratification.

While the HYPSTAR depicts very small water quality variations over time, it is limited by the first optical depth and does not provide a synoptic view of the entire reservoir. Therefore, $Ch{l}_a$ concentrations may be underestimated when blooms occur deeper in the water column, and complementary online in-water monitoring systems may be useful. Monitoring the spatial variability of the water quality products is also essential for effective water management, e.g., to ensure the better efficiency of strategies and control algal growth by acting over given sites within the reservoir. The Landsat-9 and Sentinel-2B images shown in Figure 10 confirm this need by showing how significantly the water properties may vary spatially over the reservoir. To exploit satellite images over the reservoir, common atmospheric correction algorithms for high spatial resolution images, such as ACOLITE [23], need to be improved. Indeed, due to the relatively small extent of the water reservoir, the satellite also records the light that has been reflected or emitted from adjacent vegetation pixels and scattered by the atmosphere within the field of view of the sensor [56]. Hence, atmospheric correction should also correct for the adjacency effect. The HYPSTAR measurements therefore represent very valuable validation data.

5. Conclusions

This study shows the feasibility of using hyperspectral remote sensing techniques as a rapid assessment tool for monitoring water quality. The results illustrate the benefits of hyperspectral data with very high temporal resolution. Compared with the water sampling points, the autonomous measurements were able to depict small temporal variations in $Ch{l}_a$ concentrations and $SPM$ over long periods of time and required very limited manpower. Hyperspectral sensors also aid in retrieving more advanced water quality parameters relying on very fine spectral features (e.g., for the detection of cyanobacteria). The HYPSTAR data should be further complemented with satellite images to provide a synoptic view of the reservoir, both spatially and temporally. Meanwhile, these data will also be valuable to validate the performance of the atmospheric correction algorithms over inland waters (in particular, the algorithms for the correction of adjacency effects). While timely and accurate water quality estimates are required to catch the heterogeneous and ephemeral nature of algae blooms, our results show that hyperspectral remote sensing sensors such as HYPSTAR may be confidently used to obtain first warnings and, when required, to take actions before bloom—ultimately, helping the management of drinking and recreational reservoirs with prime water quality data.