Water samples for C_Chl-a analyses were gathered from surface water (e.g., 0.5 m depth) in Saadjärv and from various depth integrated water (surface, then after every 0.5 m) in Võrtsjärv. Water samples were kept in the dark and cooled container and filtered during the same day of the fieldwork.

Duplicate samples for C_Chl-a were filtered onto 25 mm ø GF/F filters (0.7 μm pore size). Filters were stored at −20°C until being extracted with 5 ml 96% ethanol for 24 h, centrifuged for 10 min (4,000 rpm), measured spectrophotometrically (Hitachi, 2020) and C_Chl-a was calculated for mixed phytoplankton assemblage according to Jeffrey and Humphrey (1975).

For HPLC analysis, 100–700 ml of sampled lake water was vacuum filtered through 47-mm Whatman GF/F, triplicate filters were stored in 5 ml plastic vials, frozen immediately and kept at −70°C before analysis. Phytoplankton pigments were extracted in 100% acetone (2 ml) containing internal standard and sonicated (Branson 1210) for 5 min. Samples were stored at −20°C for 24 h. After that, the extracts were filtered through 0.45 μm syringe filters (Millex LCR, Millipore) and stored in dark refrigerator until HPLC analysis (for details, see Tamm et al., 2015). C_Chl-a and Chlorophyllide a values were summed up for total C_Chl-a.

Data from two AHFM buoy stations measuring fluorescence were used (Figure 1). Võrtsjärv AHFM buoy (58.211798 N, 26.103163 E) was equipped with a Yellow Springs Instruments (YSI) model 6600 V2-4 multiparameter sonde in 1-m depth. The sonde has been fitted with a chlorophyll fluorescence probe (model 6025) and was recording after every 10 min frequency. Saadjärv AHFM buoy station (58.536963 N, 26.647558 E) was equipped with a YSI EXO-2 multiparameter sonde and worked as a vertical profiler within 2–20 m water column. This sonde was fitted with an EXO Total Algae-Phycocyanin sensor. The buoy was set to make profiles after every 30 min in 2018 and 1-h frequency in 2019, from surface to bottom and the data was recorded every 4–5 cm. The automated sensor-based measurements of chlorophyll fluorescence (ChlF) was converted into C_Chl-a using standard manufactory coefficient and local conversion factors, derived via linear interpolation from monthly in vitro spectrophotometrically measured C_Chl-a. All underwater sensors in both AHFM systems were calibrated at least once per month according to the manufacturer instructions.

Both AHFM systems were also equipped with the multiparameter weather stations (Vaisala Weather Transmitter WXT520 in Võrtsjärv; Airmar 200WX Weather Station Instrument in Saadjärv) and solar irradiance sensors for above-water measurements. Photosynthetically active radiation (PAR) for Saadjärv was recorded with a Li-Cor quantum sensor (model LI-190SZ), while in Võrtsjärv, the buoy was equipped with a Li-Cor pyranometer (model LI-200SA), where PAR was calculated as 0.436 x Q (Q—incident global radiation) (Reinart and Pedusaar, 2008).

The non-photochemical quenching (NPQ) correction was performed according to Moiseeva et al. (2020): P A R z = P A R 0 × e − K d × z (1) d o p = e − 0.0019 × P A R z (2) F r e a l = F t × 2 − d o p (3) where PAR_z is photosynthetically active radiation, which penetrates to depth z, PAR₀ is PAR falling to the lake surface, K _d is a diffuse attenuation coefficient, d_op is a portion of the open reaction centres (photosystem 2), F_t is a quasi-stationary level of fluorescence in an object adapted to light and F_real is a corrected chlorophyll fluorescence. In Võrtsjärv, K _d was obtained from the WISPStation radiometric data (Alikas et al., 2015). In Saadjärv, in situ measured Secchi depth was used to derive the euphotic depth (Z_eu) as a ratio between coefficient 2.69 and Secchi depth (Luhtala and Tolvanen, 2013), which was then converted to K _d (Koenings and Edmundson, 1991). The corresponding Z₉₀ depth (depth at which 90% of the surface downwelling irradiance is attenuated) and Z_eu (reflects the depth where PAR is 1% of its surface value) were derived.

Fixed WISPStation was located in the pier of Võrtsjärv (Figure 1E, 58.211186 N, 26.107979 E). The station contains three radiometers that measure radiance and irradiance under fixed angles (Peters et al., 2018) with 15-min frequency. For a detailed description of the measurement setup, data processing and calibration of WISPStation, see Peters et al. (2018). Processed WISPStation data was downloaded from the WISPweb (https://wispweb.waterinsight.nl), where C_Chl-a has been calculated from derived reflectance according to Gons (1999). Data was filtered based on the solar zenith angle (>70°), and exceptionally high values of C_Chl-a (>200 μg/L), not consistent with the known natural background, were removed.

Satellite images from S2 MSI and S3 OLCI were used. Data was downloaded from Estonian National Satellite Data Centre ESTHub (ESTHub, 2022) with a pixel size of 60 m for S2 MSI and 300 m for S3 OLCI. First, S2 and S3 L1 data were processed with IDEPIX in SeNtinel Application Platform (SNAP) and pixels marked with cloud, cloud ambiguous, cloud sure, cloud buffer, cloud shadow, snow_ice and Sun glint risk flags were removed. Next, lake specific C_Chl-a algorithms were applied (Table 2).

Previous studies (Mograne et al., 2019; Pereira-Sandoval et al., 2019; Warren et al., 2019; Alikas et al., 2020) have shown that C2RCC and POLYMER (Steinmetz et al., 2011) tend to work relatively well compared to other available atmospheric correction methods on MSI and OLCI data over optically different waters. The atmospherically corrected data, standard C_Chl-a products from these processors together with previously developed approaches, based on L1 data (Alikas et al., 2015; Ansper and Alikas, 2018; Alikas et al., 2020), were tested over both lakes in terms of their accuracy and data availability.

In eutrophic Võrtsjärv (mean C_Chl-a 36.3 μg/L, TSM 19.9 mg/L, a_CDOM(440) 3.0 m⁻¹), L1 data based C_Chl-a retrieval showed to be more robust and resulted in more retrievals than atmospherically corrected L2 or any standard product for deriving C_Chl-a. Therefore, the Maximum Chlorophyll Index (MCI) (Gower et al., 2008) was applied to L1 data and C_Chl-a was derived by using empirical algorithms from S2 and S3 data in Võrtsjärv (Table 2).

In mesotrophic Saadjärv (mean C_Chl-a 4.8 μg/L, TSM 1.5 mg/L, a_CDOM(440) 1.0 m⁻¹), for S2 data POLYMER products resulted only in two quality controlled points in 2018 and four points in 2019, therefore C2RCC was chosen. C2RCC processor’s standard C_Chl-a product (chl_conc) with regional conversion factors was applied to S2 data. Also various empirical approaches were tested but due to high uncertainties in the shape and in the magnitude of the water-leaving reflectance from C2RCC, it did not result in more accurate C_Chl-a retrievals. For S3 data, POLYMER atmospheric correction was applied to derive remote sensing reflectance (ρ) and a ratio of 709 and 665 after Gilerson et al. (2010) was applied with lake-specific coefficients (Table 2).

For S2 and S3 images, 3 × 3 pixel area centered at the coordinates (ROI—region of interest) of the in situ stations were extracted for further analyses (Figure 1). The mean (µ) and standard deviation (σ) were calculated within the ROI. Each ROI was checked for outliers following the OLCI validation guidelines (EUMETSAT, 2019). Single pixel outliers were removed if C_Chl-a < (µ—1.5σ) or C_Chl-a > (µ + 1.5σ). Entire ROI was excluded when the ratio between standard deviation and mean e.g., coefficient of variation (CV), was greater than 0.2 (e.g. 20%).

Comparison of in vitro methods showed generally higher C_Chl-a by spectrophotometric approach compared to HPLC (Figure 3D). HPLC readings were, on average, 31% lower than spectrophotometrically measured C_Chl-a in Võrtsjärv (Table 4). In Saadjärv, the discrepancy was even more considerable.

The difference between the in vitro methods reflected also in the comparison with other methods. Comparison with WISPstation data showed underestimation of spectrophotometric C_Chl-a (24% bias, 29% MAE) and overestimation of HPLC C_Chl-a (11% bias, 13% MAE).

Compared to all methods, the smallest bias and MAE were derived between spectrophotometric and S3 (e.g., 2% bias in Võrtsjärv) and fluorescence (e.g., 6% bias in Võrtsjärv) based estimates in both lakes (Table 4).

In both lakes, the AHFM on ChlF delivered more than 100 days of data per year to study the seasonal dynamics of phytoplankton. As seen on Figure 2, the changes can be with high magnitude and rapid (e.g., daily changes in C_Chl-a ∼10 μg/L in Saadjärv and ∼30 μg/L in Võrtsjärv). This seasonal dynamics is well captured by all methods with varying measurement frequency in Võrtsjärv (Figure 2A, B) with a bias from 6%–13% and MAE from 7%–38% in respective to fluorescence measurements (Table 4). In Saadjärv, there is a clear difference between the S2 and S3 derived seasonal dynamics (Figures 2C, D), with S3 tends to follow more similar pattern with fluorescence measurements than S2. It resulted in statistically significant different retrievals with 55% bias and 70% MAE.

In terms of the fluorescence measurements, in both lakes, the difference between the midday and night-time ChlF increased with increasing phytoplankton amount. Night-time ChlF tends to be higher during more abundant phytoplankton e.g. during the spring bloom in Saadjärv (up to 5.9 RFU) and late summer bloom in Võrtsjärv (up to 1.7 RFU). With this in mind, statistics between all methods in respective to ChlF night measurements were derived, which showed that the daytime ChlF measurements resulted in better consistency in eutrophic Võrtsjärv with all methods. In Saadjärv, the differences in the derived statistics were small and more data would be needed to study the impact of choosing between night or daytime ChlF as a reference data.

The comparison of in-water fluorescence measurements showed that the short-term temporal variability was 60% higher on average in the clear water Saadjärv (MAPD 11%) than in turbid Võrtsjärv (MAPD 4.5%) within the ±30 min time interval (Supplementary Figure S1). While in Saadjärv the short-term variability in recorded ChlF measurements was higher during spring bloom (in both day and night measurements), no seasonal dependence respective to the phytoplankton quantity was observed in Võrtsjärv. In comparison, spectral data (i.e., WISPStation) showed higher standard deviation around the midday measurements towards autumn—during low light conditions. The comparison of in-water fluorescence and above-water radiometric methods in Võrtsjärv showed the in-water measurements tend to be more stable while the above-water measurements are more prone to outliers (Supplementary Figure S1).

Despite the methodological similarities in deriving C_Chl-a from WISPStation, S2 and S3 data, the comparison showed statistically significant differences, high scatter (Figures 3A1,2) and error up to 45% (Table 4) between WISPStation and EO data. Consistency was better between EO approaches in Võrtsjärv (Figure 3C; Table 4). In Saadjärv, although S2 and fluorescence measurements resulted in smallest bias (5%) and error (51%) from all methods in Saadjärv (Table 4), the C2RCC derived C_Chl-a estimates from S2 data resulted in fairly stable phytoplankton seasonal dynamics (Figures 2C, D) which was not supported by S3 and fluorescence based data.

In terms of spatial variability within the ROI, it was higher in Saadjärv during periods with more abundant phytoplankton (i.e. spring), but there were no systematic seasonal differences in S2 and S3 data over Võrtsjärv despite of the distinctive periods with higher C_Chl-a (Figures 2A, B).

The data from two AHFM systems (WISPstation and fluorescence buoy) in Võrtsjärv, resulted in 140 simultaneous measurements over 2 year period. Despite their moderate agreement (R ² = 0.5), C_Chl-a from WISPStation was statistically significantly lower (on average 26%) than from fluorescence measurements (Table 4), larger values (>80 μg/L) were especially underestimated (Figure 3). Based on the statistics (Table 4), the fluorescence derived C_Chl-a tends to have better consistency with other methods than radiometric WISPStation measurements.

There are various methods available to estimate the C_Chl-a from the ChlF measurements (Ferreira et al., 2012; Zeng et al., 2017). The fluorescence yield per chlorophyll unit is very variable and depends on phytoplankton community composition, cell size, packaging effect and NPQ (Carberry et al., 2019 and references therein) and is difficult to account for regular basis. This is especially a challenge in the waters where phytoplankton community consists of many different species and various life cycle phases are present.

High-frequency measurements allow obtaining information from ChlF in sufficient temporal scale relevant to natural dynamics of the phytoplankton community. Photoprotection against high light induced by the xanthophyll cycle will lead to a non-photochemical quenching. The effect of NPQ correction clearly increased with increased PAR (Kromkamp et al., 2008; Ruban, 2016) and also depended on the level of OAS (optically active substances), leading up to 15% change in ChlF readings in Võrtsjärv compared to 81% in Saadjärv (Supplementary Figure S2). The amount of PAR of the total solar radiation depends on the wavelength, solar zenith angle, the aerosol amount in the atmosphere and clouds (Ross & Sulev, 2000). In Estonian geographic location the monthly total PAR is highest in June and decreases towards spring and autumn (Russak & Kallis, 2003). Here we showed the consistency between C_Chl-a derived from above water radiometry (S2, S3, WISPStation) and fluorometers tended to decrease during high flux intensities in summer, especially pronounced in clear water Saadjärv (Figure 5). On the contrary, during autumn, when the illumination conditions were poorer, the consistency between the same methods was better during high flux intensities and decreased during low flux intensities (Figure 4). While in Võrtsjärv day-time ChlF was better reference based on the derived statistics (results not shown here), there was no clear pattern in Saadjärv. As the night-time ChlF tends to be higher during the bloom period and the difference was substantially higher in Saadjärv (up to 230%) compared to Võrtsjärv (up to 40%), it should be studied further in conjunction with inter-comparison of different methods to account for the NPQ.

It has been shown that CDOM and non-algal particles impede the accurate estimation of Sun-induced ChlF from the total reflectance spectra (McKee et al., 2007; Gilerson et al., 2008). Despite, the results from eutrophic Võrtsjärv show a strong correlation between C_Chl-a derived from below water fluorometry and above-water radiometry (Figure 6), it was also shown that both methods depend on the background turbidity (Figure 4). Proctor & Roesler (2010) and Kuha et al. (2020) outlined that organic matter may lead to an underestimation of C_Chl-a by absorbing excitation or emission wavelengths or, on the other hand, cause seemingly intensified Chl emission by contributing to the signal detected by Chl fluorometers. For example, a significant overestimation of C_Chl-a with increased organic matter concentrations in an estuary was shown by Goldman et al. (2013). Results by Cremella et al. (2018) showed a linear response between ChlF and a_CDOM(440) up to 20 m⁻¹ and a non-linear response between ChlF and CDOM at a_CDOM(440) > 20 m⁻¹, also noting the negligible effect in CDOM ranges (a_CDOM(440) < 2 m⁻¹) and pointing out the lack of interaction between turbidity and CDOM effects. In Saadjärv, the effect of CDOM and non-algal particles can be considered negligible. In Võrtsjärv, both the mean value and seasonal variation of a_CDOM(440) and TSM were higher (Table 1), which requires the adaption of algorithms to different levels of OAS and more frequent measurements to calibrate ChlF readings.

The fact that spectrophotometric measurements give higher values in comparison with HPLC, is not a new finding (Meyns et al., 1994; Sørensen et al., 2007). A strong positive correlation has been demonstrated between HPLC and spectrophotometrically measured C_Chl-a, with C_Chl-a being 15%–20% higher via spectrophotometry than via HPLC (Sørensen et al., 2007; Tamm et al., 2015). Meyns et al. (1994) associated the differences in the measurements by HPLC and spectrophotometric methods with the degradation products of C_Chl-a in the samples. Spectrophotometric measurements resulted in higher C_Chl-a values, especially due to Chlorophyllide a. In this study Chlorophyllide a was included in HPLC measurements. This discrepancy could be attributed to the presence of other C_Chl-a derivatives (allomers and epimers) and accessory pigments with overlapping spectra (Picazo et al., 2013; Tamm et al., 2015).

In case of above-water radiometry (S2, S3, WISPStation), C_Chl-a is evaluated via indirect methods by the absorption and scattering features. In Lake Võrtsjärv, same type of approach was applied on both S2 and S3 data, which resulted in good agreement (9% bias and 22% MAE) even in the changing environmental and background conditions. The discrepancies were larger between WISPStation and EO-based approaches (bias ≥30%, MAE ≥39%) (Table 4). This can be due to sensor (i.e. different spectral response function, spatial resolution, sensitivity of the sensor) and also algorithm specific differences. This was especially evident during periods with elevated turbidity, indicating the need for optical water type specific algorithms. Similarly in Saadjärv, different approaches, the empirical (S3) and neural network (S2) derived C_Chl-a showed clearly poorer agreement and stronger water type dependence. The study on optically different lakes indicates, despite the magnitude of seasonal dynamics of phytoplankton i.e. C_Chl-a and other optically active substances, the change in the optical water type requires the adaption of algorithms to have confidence in the derived C_Chl-a product throughout the season and over spatial scale.

Lake-specific approaches and previously developed regional conversion factors tuned with spectrophotometric C_Chl-a (Alikas et al., 2010; Ansper and Alikas, 2018) were used. The tuning of the algorithm is sensitive to the calibration dataset, e.g., good agreement between spectrophotometric, S2, S3 derived C_Chl-a in case of Võrtsjärv. The systematic underestimation of WISPStation C_Chl-a (∼20%) in Võrtsjärv compared to other methods (except HPLC) could be potentially corrected by further tuning or development of lake specific algorithm in order to minimize the differences between the methods. As shown also in previous studies, the agreement even between spectrophotometrically measured C_Chl-a depends largely on the solvent but also on the calculation method. For example, the calculation method according to Lorenzen (1967) yielded on average 16% smaller C_Chl-a values compared to Jeffrey and Humphrey (1975). Therefore, the inherent differences in the calibration dataset have to be considered and uncertainties evaluated, which will be then reflected in the higher order products (e.g., conversion factors, training dataset for neural network, satellite-based products, spatio-temporal analyses).

It was also observed, in case of both lakes and both S2 and S3 data, that the amount of quality-controlled data decreased towards autumn, which can be partly explained by clouds. However, this issue was stronger for narrower and smaller Saadjärv (width 1.8 km, length 6 km), where C_Chl-a and TSM gradually decreased towards autumn, therefore the level of signal from the lake decreased, but the constant strong signal from the surrounding area continued. The land adjacency effect correction is known issue in the use of EO data over water surfaces (Kiselev et al., 2014; Bulgarelli & Zibordi, 2018) and might limit the use of data obtained over smaller water bodies or from coastal sites. In eutrophic Võrtsjärv, the propagated errors due to adjacency effect and atmospheric correction in the final C_Chl-a measurement resulted in the use of L1 as the basis of the processing which showed more reliable results. In clear Saadjärv, in case of S2 data, only few POLYMER processed pixels passed the quality control during 1 year, therefore C2RCC neural network C_Chl-a product was used. Despite providing continuous seasonal time series, it had low sensitivity to C_Chl-a patterns detected by fluorescence and S3 data. Due to the inaccuracies in the shape and the magnitude of the C2RCC derived Rrs, the application of various empirical algorithms did not improve the result.

The environmental effects had lower impact on S2 and S3 data compared to WISPStation measurements. High wind speed, increase in wave height and poor illumination conditions resulted in high uncertainties in the measured radiometric data (Alikas et al., 2020), which propagated errors to C_Chl-a retrievals (Figure 4) and could explain occasional outliers and seasonal patterns (e.g., increased variations in the recorded signal).