Validation and Comparison of Water Quality Products in Baltic Lakes Using Sentinel-2 MSI and Sentinel-3 OLCI Data

Inland waters, including lakes, are one of the key points of the carbon cycle. Using remote sensing data in lake monitoring has advantages in both temporal and spatial coverage over traditional in-situ methods that are time consuming and expensive. In this study, we compared two sensors on different Copernicus satellites: Multispectral Instrument (MSI) on Sentinel-2 and Ocean and Land Color Instrument (OLCI) on Sentinel-3 to validate several processors and methods to derive water quality products with best performing atmospheric correction processor applied. For validation we used in-situ data from 49 sampling points across four different lakes, collected during 2018. Level-2 optical water quality products, such as chlorophyll-a and the total suspended matter concentrations, water transparency, and the absorption coefficient of the colored dissolved organic matter were compared against in-situ data. Along with the water quality products, the optical water types were obtained, because in lakes one-method-to-all approach is not working well due to the optical complexity of the inland waters. The dynamics of the optical water types of the two sensors were generally in agreement. In most cases, the band ratio algorithms for both sensors with optical water type guidance gave the best results. The best algorithms to obtain the Level-2 water quality products were different for MSI and OLCI. MSI always outperformed OLCI, with R2 0.84–0.97 for different water quality products. Deriving the water quality parameters with optical water type classification should be the first step in estimating the ecological status of the lakes with remote sensing.


Introduction
The importance of inland water bodies has never been greater, because inland water, including lakes, are one of the key points of the carbon cycle and, therefore, local and regional climate [1,2]. Lakes are also economically important as a source of drinking water, and also provide fishing and recreation opportunities; therefore, the monitoring of the water quality has become a global concern. In Europe, for example, the Water Framework Directive [3] is aiming to achieve good water quality status in all lakes larger than 0.5 km 2 .
Lake water quality can be determined by its optical properties, as the concentrations of chlorophyll-a (Chl-a) and the total suspended matter (TSM), the absorption coefficient of colored dissolved organic matter (CDOM), and transparency of the water measured by Secchi disk (Secchi (1) Lake Razna is the second biggest lake in Latvia by area (57.6 km 2 ) and the largest by water volume (0.46 km 3 ) with a clear oligotrophic water [35]; where SD was measured at 3.2-6.6 m with median 5.5 m, the Chl-a was 1.4-10.9 mg m −3 with median 6.4 mg m −3 , during 2018. (2) Lake Lubans is a shallow, eutrophic, and macrophyte-dominated lake, where the water level is artificially regulated, in a way that the surface area of the lake can fluctuate from 25 to 100 km 2 [35]. Our measured SD varied from 0.4-1.2 m with median 0.9 m, and the Chl-a was 8.6-63.1 mg m −3 with median 16.8 mg m −3 . (3) Lake Burtnieks is a shallow, eutrophic, and CDOM dominated northern Latvian lake with surface area 40.2 km 2 [35], where the SD values are similar to Lake Lubans, 0.4-1.2 m with median 1 m, and the Chl-a was 6.7-117.1 mg m −3 with median 19.1 mg m −3 . (4) Lake Võrtsjärv is a second largest (270 km 2 ) lake in Estonia. It is a large, shallow, very turbid, and eutrophic lake [36]. The measured SD was 0.5-0.1 m with a median 0.6 m, and the Chl-a was 22.2-63.8 mg m −3 with median 33.1 mg m −3 in 2018.
Field sampling campaigns included: collecting water samples from the surface layer; and measuring depth, temperature of the water and oxygen from about 0.5 m depth; as well as measuring water transparency with a Secchi disc. In addition, the reflectance spectra were measured above the surface (about 10 cm) using hand-held spectrometer PSR-3500 (Spectral Evolution Inc., Haverhill, MA, USA). The spectral range was 348-1000 nm and the spectral sampling interval approximately 1.5 nm. Reflectance was calculated as the ratio of radiance from water to radiance from a white reference panel (99% Spectralon®, Labsphere Inc., North Sutton, NH, USA). In each sampling station, three measurements were made, and the average reflectance spectrum was calculated to represent measurement station reflectance; in this study, we use the term reflectance to mean the remote sensing reflectance. For MSI and OLCI analyses, in-situ measured reflectance spectra were calculated to the satellite bands using specific spectral response functions of the satellite sensor bands [29].
Laboratory analysis of water samples were carried out in Estonian Marine Institute, followed the methods of Lindell et al. [37] and they included: • The spectrophotometric determination of Chl-a, where water sample aliquots were passed through 47 mm Whatman (Whatman plc., Maidstone, UK) glass microfiber filters with pore size Reflectance was calculated as the ratio of radiance from water to radiance from a white reference panel (99% Spectralon®, Labsphere Inc., North Sutton, NH, USA). In each sampling station, three measurements were made, and the average reflectance spectrum was calculated to represent measurement station reflectance; in this study, we use the term reflectance to mean the remote sensing reflectance. For MSI and OLCI analyses, in-situ measured reflectance spectra were calculated to the satellite bands using specific spectral response functions of the satellite sensor bands [29].
Laboratory analysis of water samples were carried out in Estonian Marine Institute, followed the methods of Lindell et al. [37] and they included: The spectrophotometric determination of Chl-a, where water sample aliquots were passed through 47 mm Whatman (Whatman plc., Maidstone, UK) glass microfiber filters with pore size 0.7 µm (GF/F), ethanol extracted, and then measured spectrophotometrically at 665 nm with a U-3010 laboratory spectrophotometer (Hitachi High-Technologies Corporation, Tokyo, Japan).

•
Gravimetrically determined TSM, where pre-weighed and pre-combusted 47 mm Whatman GF/F glass microfiber filters were heated at 105 • C for 1 h.

•
Determining the spectral absorption coefficients of CDOM using a U-3010 laboratory spectrophotometer (Hitachi High-Technologies Corporation) on samples prepared by successive filtrations through 47 mm Whatman GF/F filters and Nucleopore 0.2 µm filter paper.

Satellite Data
The MSI and OLCI cloud-free match-up images (±1-day included: up to three match-ups possible) were used. The ±1-day images were used, because the optical properties should not have rapid (overnight) changes in given lakes. This might be more likely in early spring (on the ice-melt), during autumn storms, or throughout other rapid weather change (heavy rains or winds). We were avoiding such weather conditions prior or during our field campaigns. We found 41 MSI match-ups for 35 different sampling points, from witch 6 sampling points had 2 match-up images; and 79 OLCI match-ups for 42 different sampling points, from witch 28 sampling points had more than one match-up images.
The MSI Level-1 images were downloaded from Copernicus Open Access Hub [38] and pre-processed by resampling the image to a 20 m resolution and applying the multi-sensor pixel identification tool (IdePix); only the "clear inland water" pixels were used. The atmospheric correction processor has a strong impact to the derived OWTs [29], so based on the recommendations of the previous studies [29,32], MSI images were further processed with Case-2 Regional CoastColour (C2RCC) and Case-2 Extreme Cases (C2X) processors [39] using scientific image processing toolbox called the Sentinel Application Platform (SNAP v 6.0) developed by Brockmann Consult, Array Systems Computing and C-S [40]. The salinity and temperature were modified in the processors according to in-situ data. In validation process we used a 3 × 3 pixel average instead of a single pixel value.
The OLCI Level-1 full resolution (300 × 300 m) images were downloaded from Copernicus Online Data Access [41]. The pre-processing of OLCI data were similar to the pre-processing of the MSI images, except no resampling was needed. For the atmospheric correction, there was only the C2RCC option available and not the C2X. A single pixel value was used in validation process. 2.3. Algorithms to retrieve the water quality products.
To obtain different optical properties, we used either the ready products from different atmospheric correction processors or band ratios using C2RCC, C2X, or top-of-atmosphere (TOA) reflectances that have shown good results previously in lakes or marine coastal areas [42]. Here too, we use term reflectance to mean the remote sensing reflectance. Details of the used sources are shown in Table 1. The K d 490 was calculated using reflectances at 490, 560, and 705 nm (or 709.5 nm in the case of OLCI) [43]. Table 1. Used algorithms to retrieve the water quality products: chlorophyll-a (Chl-a), total suspended matter (TSM), absorption of colored dissolved organic matter (CDOM) at 400, 412, and 442 nm, Secchi depth (SD), and light attenuation coefficient over 400-700 nm (K d PAR). R in band ratio algorithms notes the reflectance at given wavelength. In total, we tested 21 different options to derive Chl-a values from MSI data. Those were: C2RCC and C2X conc_chl [39]; Maximum Chlorophyll Index (MCI) Chl-a [45]; and six different band ratios from TOA, C2RCC, and C2X reflectances (Table 1). For OLCI we had somewhat less options, mostly due to the lack of C2X reflectances: Level 2 Neural Network (NN) Chl-a [44]; C2RCC conc_chl [39]; Maximum Peak Height/Chlorophyll-a (MPH/CHL) [46]; MCI Chl-a [45]; and six different band ratios from TOA and C2RCC reflectances (Table 1).

Multispectral
We tested 14 different options to derive TSM values from MSI data. Those were: C2RCC and C2X conc_tsm [39]; and four different band ratios from TOA, C2RCC, and C2X reflectances (Table 1). For OLCI data we had nine different options: C2RCC conc_tsm [39]; and four different band ratios from TOA and C2RCC reflectances (Table 1).
We tested 10 different options to derive CDOM values from MSI data. Those were: C2RCC and C2X iop_agelb [39]; and four different band ratios of C2RCC and C2X reflectances (Table 1). For OLCI we had five different options: C2RCC iop_agelb [39]; and four different band ratios of C2RCC reflectances (Table 1).
We tested 21 different options to derive SD values from MSI data. Those were: C2RCC and C2X Kd_489 and Kd_z90max [39]; the K d 490 algorithm from TOA, C2RCC, and C2X reflectances [43]; and in addition, a K d PAR algorithm [43] based on the above named K d 490 products and a SD algorithm [60] based on K d PAR products. For OLCI, we had 12 different options: C2RCC Kd_489 and Kd_z90max [39]; K d 490 algorithm [43] from TOA and C2RCC reflectances; and in addition, a K d PAR algorithm [43] based on the K d 490 products and a SD algorithm [60] based on K d PAR products.
For validating each parameter, we had tested grouping our results in three ways: by our four lakes; by the in-situ measured values (usually by low, medium, and high concentrations); and by the derived OWTs ( Figure A1). We used always the OWTs from MSI C2X for the validation of the MSI water quality parameters, and OWTs derived from OLCI C2RCC for the validation of the OLCI water quality parameters. For the CDOM validation we compared the derived CDOM values with three different in-situ CDOM values: at 400 nm, 412 nm, and 442 nm, as the CDOM values from these three wavelengths are the most commonly used.

Classification of OWT
The method by Uudeberg et al. [29] was applied for the classification of OWTs. This method classifies inland and coastal waters of boreal region into five OWTs, each with specific bio-optical conditions: (1) Clear OWT is water with the highest water transparency and the lowest optically active substance concentrations; (2) Moderate OWT is water with relatively higher concentrations of the optically significant constituents, but none of them dominate; (3) Turbid OWT is water where the TSM dominates; (4) Very Turbid OWT is water where the Chl-a dominates and this type is associated with blooms; (5) and Brown OWT is water where the CDOM dominates, and waters are dark and reddish to brown.
This method is based on reflectance spectrum key features, such as the location of spectral maximum, slopes, and amplitude. The OWT for each reflectance spectra of the pixel of the satellite image was determined by the maximum likelihood of individual spectra to reference spectra using a combination of spectral correlation similarity and modified spectral angle similarity. The OWT classification was applied to each OLCI and MSI satellite image pixel reflectance spectrum.

Statistical and Performance Measures Used in the Study
In-situ measurement dataset was described using median values and range. Median is the middle value separating the greater and lesser halves of the data; and range shows the difference between the largest and smallest value.
The performance of OWT results from different processors (atmospheric correction approaches) was evaluated using accuracy as a measure. Accuracy represents correctly predicted OWT type cases (expressed as percentage) in comparison to the reference dataset. Error refers to a percentage the derived value was not the same as in-situ.
The performance of water quality products (Chl-a, TSM, CDOM and SD) from different processors and band ratios was evaluated using two measures: determination coefficient (R 2 ) and root mean squared error (RMSE). R 2 is used to analyze how well observed in-situ values are predicted by the model based on the proportion of total variation of outcomes explained by the model; it is calculated using Equation (1). RMSE is a frequently used measure of differences between values observed in-situ and predicted by a model, it is calculated using Equation (2).
where,ŷ is the predicted value, y is the observed value and y is the mean value of observed y values.

In-Situ Water Quality Parameters
The median and the range of measured optical properties from 49 sampling points across four lakes during 2018, are shown in Table 2. The water transparency decreased towards the summer and again increased in autumn, in all of the studied lakes. The average Secchi depth of the measurements was below 1 m, except in Lake Razna of TSM followed the seasonal patterns of Chl-a. CDOM showed reasonably low seasonal variability and decreased in all the lakes from spring to autumn. Table 2. The area and average depth (D), median and range in parenthesis of Secchi depth (SD), Chlorophyll-a (Chl-a), total suspended matter (TSM), and colored dissolved organic matter (CDOM) at 400 nm from four different lakes collected during April to November 2018. The n denotes the number of samples. Three out of four of the studied lakes are situated in Latvia (LV) and one in Estonia (EE).

Reflectance Spectra
The match-up in-situ and satellite's sensors reflectance spectra for each lake are shown in Figure 2. The in-situ reflectance spectra are calculated to the OLCI bands used OLCI spectral response function. The MSI and OLCI reflectances are atmospherically corrected (C2RCC and C2X).

Reflectance Spectra
The match-up in-situ and satellite's sensors reflectance spectra for each lake are shown in Figure  2. The in-situ reflectance spectra are calculated to the OLCI bands used OLCI spectral response function. The MSI and OLCI reflectances are atmospherically corrected (C2RCC and C2X).  Based on the comparison of the spectral bands ( Figures A2-A4) and the general shape of the spectra, in most cases, the best performer was MSI C2X, as it replicates the shape of the spectra well. OLCI C2RCC also caught the shape of the spectra, thanks to more available bands in the visible range than the MSI. In both cases, MSI and OLCI, the C2RCC underestimated the reflectance spectra, especially on the shorter and longer wavelengths (Figures 2 and A2, Figures A3 and A4).

Optical Water Types (OWTs)
Calculating the hyperspectral in-situ reflectance spectra to MSI bands did not change the obtained OWT. As MSI and OLCI have different bands for determining OWT, we first used in-situ data in the simulation to observe the possible differences on the MSI and OLCI bands to obtain the OWTs. All 38 match-up cases were determined correctly by using MSI bands. However, the calculation to the OLCI bands changed the obtained OWT in six cases. Half of them were from Turbid to Moderate, this and the other three possibilities are shown in Figure 3. Based on the comparison of the spectral bands ( Figures B1-3) and the general shape of the spectra, in most cases, the best performer was MSI C2X, as it replicates the shape of the spectra well. OLCI C2RCC also caught the shape of the spectra, thanks to more available bands in the visible range than the MSI. In both cases, MSI and OLCI, the C2RCC underestimated the reflectance spectra, especially on the shorter and longer wavelengths (Figures 2 and B1-3).

Optical Water Types (OWTs)
Calculating the hyperspectral in-situ reflectance spectra to MSI bands did not change the obtained OWT. As MSI and OLCI have different bands for determining OWT, we first used in-situ data in the simulation to observe the possible differences on the MSI and OLCI bands to obtain the OWTs. All 38 match-up cases were determined correctly by using MSI bands. However, the calculation to the OLCI bands changed the obtained OWT in six cases. Half of them were from Turbid to Moderate, this and the other three possibilities are shown in Figure 3. We obtained several sets of OWTs: OWTs from in-situ reflectances calculated to MSI bands (MSI in-situ) and to OLCI bands (OLCI in-situ), MSI processed with C2RCC processor (MSI C2RCC) and C2X processor (MSI C2X), OLCI processed with C2RCC processor (OLCI C2RCC). We compared the MSI in-situ with MSI C2RCC and MSI C2X OWTs; and OLCI in-situ with OLCI C2RCC OWTs. The MSI C2X OWTs showed the highest potential in obtaining correct OWT (accuracy 72%, R 2 = 0.81, and n = 32). MSI C2RCC accuracy was lower: 50%, R 2 = 0.5. C2RCC of OLCI performed somewhat better: accuracy 65%, R 2 = 0.49, n = 60. The detailed distribution of the obtained OWTs is shown in Figure 4. We obtained several sets of OWTs: OWTs from in-situ reflectances calculated to MSI bands (MSI in-situ) and to OLCI bands (OLCI in-situ), MSI processed with C2RCC processor (MSI C2RCC) and C2X processor (MSI C2X), OLCI processed with C2RCC processor (OLCI C2RCC). We compared the MSI in-situ with MSI C2RCC and MSI C2X OWTs; and OLCI in-situ with OLCI C2RCC OWTs. The MSI C2X OWTs showed the highest potential in obtaining correct OWT (accuracy 72%, R 2 = 0.81, and n = 32). MSI C2RCC accuracy was lower: 50%, R 2 = 0.5. C2RCC of OLCI performed somewhat better: accuracy 65%, R 2 = 0.49, n = 60. The detailed distribution of the obtained OWTs is shown in Figure 4.  To study the "magnitude of accuracy", we looked at how serious the errors in determining the OWT were. The cause of misclassification may be the fact that the reflectance spectra were in a border of two OWT classes. Therefore, it might happen that slightly different reflectance spectra are classified differently. If the derived OWT classes were far from each other: Clear-Turbid, Clear-Very Turbid, Clear-Brown, or Turbid-Brown, this was defined as a large OWT difference, and all the other options (for example, Turbid-Very Turbid, or Clear-Moderate, etc.) were defined as a little OWT difference ( Figure 5). The distribution of the derived optical water types (OWTs) accuracy: same OWT as in-situ defined as correct OWT (green); neighboring OWT from in-situ defined as little OWT difference (yellow); more than neighboring OWT than in-situ defined as large OWT difference (red); and not classified to any OWT (grey). To study the "magnitude of accuracy", we looked at how serious the errors in determining the OWT were. The cause of misclassification may be the fact that the reflectance spectra were in a border of two OWT classes. Therefore, it might happen that slightly different reflectance spectra are classified differently. If the derived OWT classes were far from each other: Clear-Turbid, Clear-Very Turbid, Clear-Brown, or Turbid-Brown, this was defined as a large OWT difference, and all the other options (for example, Turbid-Very Turbid, or Clear-Moderate, etc.) were defined as a little OWT difference ( Figure 5). To study the "magnitude of accuracy", we looked at how serious the errors in determining the OWT were. The cause of misclassification may be the fact that the reflectance spectra were in a border of two OWT classes. Therefore, it might happen that slightly different reflectance spectra are classified differently. If the derived OWT classes were far from each other: Clear-Turbid, Clear-Very Turbid, Clear-Brown, or Turbid-Brown, this was defined as a large OWT difference, and all the other options (for example, Turbid-Very Turbid, or Clear-Moderate, etc.) were defined as a little OWT difference ( Figure 5). Figure 5. The distribution of the derived optical water types (OWTs) accuracy: same OWT as in-situ defined as correct OWT (green); neighboring OWT from in-situ defined as little OWT difference (yellow); more than neighboring OWT than in-situ defined as large OWT difference (red); and not classified to any OWT (grey). Figure 5. The distribution of the derived optical water types (OWTs) accuracy: same OWT as in-situ defined as correct OWT (green); neighboring OWT from in-situ defined as little OWT difference (yellow); more than neighboring OWT than in-situ defined as large OWT difference (red); and not classified to any OWT (grey).
The errors were rather small in MSI C2X derived OWTs (only 6% OWT are classified as large OWT difference from in-situ). It is also seen that in the case of OLCI the large error is much higher (22%), from which 60% of this error came from the misclassifying spectrum to the Clear OWT (MSI data did not have any of this kind of cases). MSI C2RCC has the largest percentage of unclassified cases (9.5%).

Chlorophyll-a (Chl-a)
We compared 21 different methods to determine Chl-a values from MSI data with in-situ Chl-a values and 16 different methods for OLCI (Table 1). We searched for the best fits by sorting our data by lakes, in-situ Chl-a values, and by obtained OWTs. The best results for both, MSI and OLCI, gave different TOA band ratios for different OWTs using always either 705 nm or 709 nm in the band ratio. The overall results for both sensors, MSI and OLCI were quite similar: R 2 = 0.84 for MSI and 0.83 for OLCI, the root mean square error (RMSE) was 10.5 mg m −3 for MSI and 9.8 mg m −3 for OLCI. The correlation of best performed TOA band ratios together are shown in Figure 6 and the details of the band ratios and their R 2 are shown in Table 3. As seen from Figure 6 and Table 3, the low Chl-a values can be a challenge for sensors. The errors were rather small in MSI C2X derived OWTs (only 6% OWT are classified as large OWT difference from in-situ). It is also seen that in the case of OLCI the large error is much higher (22%), from which 60% of this error came from the misclassifying spectrum to the Clear OWT (MSI data did not have any of this kind of cases). MSI C2RCC has the largest percentage of unclassified cases (9.5%).

Chlorophyll-a (Chl-a)
We compared 21 different methods to determine Chl-a values from MSI data with in-situ Chl-a values and 16 different methods for OLCI (Table 1). We searched for the best fits by sorting our data by lakes, in-situ Chl-a values, and by obtained OWTs. The best results for both, MSI and OLCI, gave different TOA band ratios for different OWTs using always either 705 nm or 709 nm in the band ratio. The overall results for both sensors, MSI and OLCI were quite similar: R 2 = 0.84 for MSI and 0.83 for OLCI, the root mean square error (RMSE) was 10.5 mg m -3 for MSI and 9.8 mg m -3 for OLCI. The correlation of best performed TOA band ratios together are shown in Figure 6 and the details of the band ratios and their R 2 are shown in Table 3. As seen from Figure 6 and Table 3, the low Chl-a values can be a challenge for sensors.

Total Suspended Matter (TSM)
We compared 14 different methods to determine TSM values from MSI data with in-situ TSM values and nine different methods for OLCI (Table 1). Here, for MSI, the C2RCC Level 2 product conc_tsm performed well for Moderate and Very Turbid OWTs, and generally, the C2RCC reflectances outperformed C2X. For OLCI different band ratios were showing the best results for different OWTs. The correlation of the best results with OWT guiding is shown in Figure 7 and the details are shown in Table 4. There was one case where the difference between the derived TSM and in-situ TSM was very large: in-situ 53 mg L −1 , satellite derived 29 mg L −1 (Lake Lubans: in-situ Very Turbid OWT; OLCI Turbid OWT). This comes from different reflectance spectra, where the in-situ spectrum is 2.5 times higher than OLCI reflectance spectrum. MSI outperformed OLCI (R 2 = 0.89 for MSI and 0.81 for OLCI). Again, the low concentrations were a challenge for OLCI. The result was slightly better when the data was grouped by the in-situ TSM values (Table 4).

Total Suspended Matter (TSM)
We compared 14 different methods to determine TSM values from MSI data with in-situ TSM values and nine different methods for OLCI (Table 1). Here, for MSI, the C2RCC Level 2 product conc_tsm performed well for Moderate and Very Turbid OWTs, and generally, the C2RCC reflectances outperformed C2X. For OLCI different band ratios were showing the best results for different OWTs. The correlation of the best results with OWT guiding is shown in Figure 7 and the details are shown in Table 4. There was one case where the difference between the derived TSM and in-situ TSM was very large: in-situ 53 mg L -1 , satellite derived 29 mg L -1 (Lake Lubans: in-situ Very Turbid OWT; OLCI Turbid OWT). This comes from different reflectance spectra, where the in-situ spectrum is 2.5 times higher than OLCI reflectance spectrum. MSI outperformed OLCI (R 2 = 0.89 for MSI and 0.81 for OLCI). Again, the low concentrations were a challenge for OLCI. The result was slightly better when the data was grouped by the in-situ TSM values (Table 4).

Figure 7.
Correlation between in-situ and satellite derived total suspended matter (TSM, mg L -1 ) based on the optical water types (OWTs). Also, R 2 , RMSE, and the number of samples (n) are given.

Colored Dissolved Organic Matter (CDOM)
We compared 10 different methods to determine CDOM values from MSI data with in-situ CDOM values and five different methods for OLCI (Table 1). Yet again, we searched for the best fits

Colored Dissolved Organic Matter (CDOM)
We compared 10 different methods to determine CDOM values from MSI data with in-situ CDOM values and five different methods for OLCI (Table 1). Yet again, we searched for the best fits by sorting our data by lakes, in-situ CDOM values, and by OWTs. The best results with the OWT guided approach gave a comparison with in-situ CDOM at 400 nm. MSI CDOM results were better than for OLCI (R 2 = 0.91 for MSI and 0.76 for OLCI) (Figure 8). There is again one case standing out in the OLCI data: in-situ 5.8 m −1 and satellite derived 1.2 m −1 (Lake Lubans: in-situ Turbid OWT; OLCI Clear OWT). The in-situ and OLCI reflectance spectrum of this point was very different by its shape and magnitude (the in-situ spectrum was two times higher in the blue part of the spectrum, but 100 times higher in the red part). Although, the overall MSI result is very good then for the Clear OWT R 2 was only 0.15 (Table 5). Here, as for TSM, for different OWTs the most suitable results might come from different options: the TOA, C2RCC, or C2X band ratios. by sorting our data by lakes, in-situ CDOM values, and by OWTs. The best results with the OWT guided approach gave a comparison with in-situ CDOM at 400 nm. MSI CDOM results were better than for OLCI (R 2 = 0.91 for MSI and 0.76 for OLCI) (Figure 8). There is again one case standing out in the OLCI data: in-situ 5.8 m -1 and satellite derived 1.2 m -1 (Lake Lubans: in-situ Turbid OWT; OLCI Clear OWT). The in-situ and OLCI reflectance spectrum of this point was very different by its shape and magnitude (the in-situ spectrum was two times higher in the blue part of the spectrum, but 100 times higher in the red part). Although, the overall MSI result is very good then for the Clear OWT R 2 was only 0.15 (Table 5). Here, as for TSM, for different OWTs the most suitable results might come from different options: the TOA, C2RCC, or C2X band ratios.

Secchi Depth (SD)
We compared 21 different methods to determine SD values from MSI data with in-situ SD values and 12 different methods for OLCI ( Table 1). The OWT guided approach worked very well on MSI data (R 2 = 0.97) (Figure 9), but OLCI performance was weaker (R 2 = 0.69, RMSE = 1.1 m) (Figure 9). In all cases the Alikas et al. [43] algorithm to obtain the Kd490 was the best choice (Table 6), however, the source was different (either TOA or C2X for MSI, and C2RCC or TOA for OLCI).

Secchi Depth (SD)
We compared 21 different methods to determine SD values from MSI data with in-situ SD values and 12 different methods for OLCI ( Table 1). The OWT guided approach worked very well on MSI data (R 2 = 0.97) (Figure 9), but OLCI performance was weaker (R 2 = 0.69, RMSE = 1.1 m) (Figure 9). In all cases the Alikas et al. [43] algorithm to obtain the K d 490 was the best choice (Table 6), however, the source was different (either TOA or C2X for MSI, and C2RCC or TOA for OLCI).

Discussion
We used the C2RCC and C2X atmospheric correction processors for MSI, and C2RCC, for OLCI due to the results and recommendations of the previous studies [29,32,61]. The comparison of the insitu and satellite sensors reflectance spectra for each lake revealed MSI C2X as the best performer in most cases. It was surprising that even in the oligotrophic lake like Razna with only Clear OWTs, the obtained reflectance spectra magnitude and shape were quite similar to the in-situ spectra with the MSI C2X processor (Figure 2). Hence, we assumed that also the derived OWTs were the most accurate based on MSI C2X spectra (accuracy 72%, R 2 = 0.81). The OLCI and MSI C2RCC processed reflectance spectra showed large underestimation of the reflectance spectra. This also often missed the Chl-a absorption peak on or around 665 nm; the Chl-a peak in reflectance spectra of eutrophic waters can shift towards longer wavelengths with increasing biomass [53]. Lake Razna reflectance spectra demonstrate typical Clear OWT spectra; while Lake Burtnieks shows large peaks in the second Chla absorption maximum.

Discussion
We used the C2RCC and C2X atmospheric correction processors for MSI, and C2RCC, for OLCI due to the results and recommendations of the previous studies [29,32,61]. The comparison of the in-situ and satellite sensors reflectance spectra for each lake revealed MSI C2X as the best performer in most cases. It was surprising that even in the oligotrophic lake like Razna with only Clear OWTs, the obtained reflectance spectra magnitude and shape were quite similar to the in-situ spectra with the MSI C2X processor (Figure 2). Hence, we assumed that also the derived OWTs were the most accurate based on MSI C2X spectra (accuracy 72%, R 2 = 0.81). The OLCI and MSI C2RCC processed reflectance spectra showed large underestimation of the reflectance spectra. This also often missed the Chl-a absorption peak on or around 665 nm; the Chl-a peak in reflectance spectra of eutrophic waters can shift towards longer wavelengths with increasing biomass [53]. Lake Razna reflectance spectra demonstrate typical Clear OWT spectra; while Lake Burtnieks shows large peaks in the second Chl-a absorption maximum.
The study of the MSI and OLCI spectral resolution impact to the classification of the OWT showed that MSI bands are very good for the detection of the OWTs. From 38 hyperspectral in-situ reflectance spectra calculated to MSI and OLCI, the bands showed a 100% match in the case of MSI, but only 85% matches in the case of OLCI bands. Although authors in [29] found both instruments equally accurate for determining the OWTs, in a current study, the OLCI bands might end up with different OWTs. Figure 3 shows some of these examples of how different placements of the spectral bands can lead to deriving different OWTs. Usually the derived OWTs are neighboring (Turbid and Very Turbid), however, there might be bigger differences (Turbid and Brown). Similarly, Uudeberg et al. [29] found that for 8% of Very Turbid OWTs, OLCI misclassified them as Turbid and 5% of Turbid OWTs were misclassified as Moderate.
Due to of the slight differences in obtaining the OWTs for the MSI and OLCI bands, we used, in our OWT validation process, two different sets of in-situ OWTs: in-situ OWTs based on the MSI bands and based on the OLCI bands. It is shown that in the case of MSI the Clear OWT estimates are similar. But the other OWT classes might be more challenging (Figure 4). Our study of the accuracy of the OWT classification showed also that only 6% OWT are classified as "large OWT difference" from in-situ with MSI C2X ( Figure 5). OLCI C2RCC largest errors come with the false-estimation of the Clear OWT class (in 14% of the cases from all the cases). These spectra might have been affected by the in-water vegetation, however, due to of the insufficient quality flagging and IdePix tool performance on OLCI data, they are not dismissed and are used as clear water pixels [61]. Without this misclassification of Clear OWT, the "large OWT difference" would be only 8%, instead of 22% ( Figure 5). Based on these results, we decided that for the further water quality product validation the MSI C2X OWTs for MSI, and OLCI C2RCC OWTs for OLCI water quality products will be used.
Overall, the best results for different inland water quality products were obtained using OWT classification and specific for each OWT band ratio (Figures 6-9 and Tables 3-6). Similar to Lins et al. [62], we also used the principle of OWT guiding, rather than other methods of clustering the data. Although, fitting the algorithms to the specific lake might slightly improve the results, the methods would not be applicable for all the lakes in our region. Similarly, with in-situ concentrations clustering, prior knowledge of the given lake would be needed, while with OWT guiding any prior knowledge of the lake or water pixel is not necessary.
For the validation of the Chl-a we chose several algorithms proven to be successful in the given study area. For Chl-a, similarly to Toming et al. [31], the red and near-infrared (NIR) band ratios from TOA reflectance spectra were the most successful (Table 3). Overall MSI data were slightly better for obtaining the Chl-a in different OWTs (R 2 = 0.84). The red and NIR band ratio for Brown OWT worked especially well for both satellites. In comparison with different algorithm performance for MSI, the MCI was performing even better, but only when the Chl-a was obtained for each lake separately (R 2 = 0.91), so this approach demands the prior knowledge and adjusting the algorithm for each lake under the study. The results from OLCI were not as promising, with MCI grouped by lakes giving the overall average results, R 2 = 0.61; R 2 for MPH was even lower, 0.55. The only results slightly better than our selected OWT guided approach (R 2 = 0.83) were the TOA band ratios guided by the in-situ Chl-a values (R 2 = 0.85). The Clear OWT might be more challenging for sensors due to the typical Chl-a peaks that are not that pronounced at low Chl-a values, but surely the misclassified Clear OWT in the case of OLCI had part of the low correlation between the in-situ and OLCI derived Clear OWT Chl-a values (R 2 = 0.34; Table 3).
The OWT-guided approach was the most successful in determining accurate TSM concentrations. In the case of MSI, the C2RCC Level 2 TSM product worked well for Moderate and Very Turbid OWTs. For the rest of the OWTs the 705 nm band value from different reflectance spectra was the best (Table 4). If considering that, previously, the validation of TSM has been very challenging due to its variable properties [8], the overall results of current study can be considered as very good (R 2 = 0.89, RMSE = 3.4 mg L −1 ). In the case of OLCI, the results were somewhat weaker, overall R 2 = 0.81, this was because, similarly to Chl-a, they failed to derive Clear OWT TSM concentrations (R 2 = 0.21). For TSM, our results somewhat match with the results found in [42], where good performing TSM algorithms had bands from 700-710 nm.
Although the overall CDOM retrieval was very good for MSI (R 2 = 0.93) with OWT guiding, the Clear OWT CDOM seemingly showed poor results (R 2 = 0.15), while, in fact, the error to derive correct CDOM was less than 5% (RMSE only 0.07 m −1 ) ( Table 5). This Clear OWT poor R 2 value is caused by the small data range (in-situ 1.54-1.79 m −1 ; and MSI derived 1.57-1.65 m −1 ). For MSI, the best option was to use a two-band model, using red and green or blue bands, as also shown in some previous studies [9,31,42]. For OLCI the two-band model did not always work the best, and the three-band model offered the best fit, but still using red and green bands (Table 5).
For water transparency, the use of Alikas et al.'s [43] method on MSI TOA (for Clear, Turbid, and Very Turbid) or C2X reflectances (for Brown) to derive K d 490, gave the best results. Only for Moderate OWT, the best was to use C2X K d 489 product. The overall R 2 was very high (0.97) and RMSE was 0.36 m. The OLCI data was more challenging, due to some failings in turbid areas (in-situ SD < 1 m). However, it can also be observed that in the case of OLCI data for SD > 2 m the uncertainty of the derived SD increases; this could be related with the spatial resolution of OLCI sensor, signal strength at such depths, or turbidity. The overall RMSE was large for OLCI, 1.1 m (R 2 = 0.69). Here the K d 490 calculations are based on TOA reflectance bands for Very Turbid and Brown OWTs. For the other OWTs the atmospherically corrected reflectances worked the best. In many cases, the similar K d 490 algorithm has been successful [16,63] even if they have not considered using the TOA reflectances.
Overall the MSI seems to work better in lake water quality remote sensing than OLCI. The advantage for MSI might come from the C2X option, which shows good results in atmospheric correction and therefore the classification of the OWTs, and is, unfortunately, currently not available for OLCI. It is very important to retrieve accurate reflectance spectrum in order to obtain the correct OWT, when using the OWT guided approach for water quality retrieval. Therefore, the atmospheric correction should be chosen carefully. The MSI and especially OLCI C2RCC atmospheric correction over the lakes should be improved, because even a small error can cause significant changes in obtained Chl-a, TSM, CDOM, or SD [11]. The most challenging part of this study proved to be obtaining the Chl-a, TSM, and especially, SD for the Clear OWT from OLCI data. The challenges with data have been described before, mostly in coastal areas [11,42,64]. The issues with the retrieval of the water quality products for Clear OWTs might be due to several reasons. Firstly, in the case of low concentrations, the specific spectral shape for different optical properties might not be obvious and the spectra might miss the maximum absorption peaks necessary for the further calculations. Secondly, the errors in the classification of the OWTs, as a substantial proportion of the cases were falsely classified to Clear OWT, which comes either from the atmospheric correction issues over lakes with C2RCC [29,61] or from the poor clear water pixel masking of the processors and IdePix tool (especially for the OLCI data) [61].
Despite the relatively good overall accuracy of the remote sensing products, several issues still remain to be addressed in order to incorporate remote sensing into the national monitoring programs: the above-mentioned atmospheric correction of OLCI, spatial resolution of the OLCI that limits the use of it in smaller waterbodies; and lastly, the optical depth of the waters. The latter means that it is difficult to estimate water properties of the water column when the remote sensing signal comes primarily from the lake bottom or when water properties vary significantly in the water column [65][66][67]. The latter is not an issue in large lakes or seas where top tens of meters are uniformly mixed and the optical depth (from which the satellite signal originates) is smaller than the mixed layer depth. Smaller lakes, however, may have a quite sharp stratification of water properties that cannot be resolved with remote sensing.
Future work should include studies on the spatial, vertical, and temporal variability of the water quality in the studied lakes; obtaining the ecological status of the lakes based on the Chl-a and SD products. Future work should also extend this method by offering the water quality products and ecological state estimation for all lakes in Latvia or even larger areas.

Conclusions
Using the OWT guided approach proved to be a very good method for estimating water quality parameters in lakes using both MSI and OLCI data. We used in-situ data from 49 sampling points across four different lakes collected during 2018 to validate MSI and OLCI Level 2 water quality products, such as Chl-a, TSM, CDOM, and SD. For processing the MSI and OLCI match-ups, we used C2RCC and C2X processors. MSI and OLCI C2RCC reflectance spectra mostly underestimated the in-situ reflectances, but MSI C2X were able to retrieve similar spectral shapes and often also the magnitude, even in clear waters. The MSI C2X reflectance spectra gave the most accurate OWTs (R 2 = 0.8) and we used this for the water quality validation process. Our study shows that the used reflectance spectra have a large impact on the OWT guided approach to estimate the water quality products, as the obtained OWTs are directly dependent on the shape of the reflectance spectra. For the water quality products, the algorithms from different band ratios based (often TOA reflectance) on OWTs gave the best results for both satellites. The MSI outperformed OLCI every time. As there are more issues with OLCI (atmospheric correction, spatial resolution, and masking), we recommend using MSI for national monitoring with OWT guiding prior application of the algorithms for the estimation of the water quality or the ecological state of the lakes.  Using the OWT guided approach proved to be a very good method for estimating water quality parameters in lakes using both MSI and OLCI data. We used in-situ data from 49 sampling points across four different lakes collected during 2018 to validate MSI and OLCI Level 2 water quality products, such as Chl-a, TSM, CDOM, and SD. For processing the MSI and OLCI match-ups, we used C2RCC and C2X processors. MSI and OLCI C2RCC reflectance spectra mostly underestimated the in-situ reflectances, but MSI C2X were able to retrieve similar spectral shapes and often also the magnitude, even in clear waters. The MSI C2X reflectance spectra gave the most accurate OWTs (R 2 = 0.8) and we used this for the water quality validation process. Our study shows that the used reflectance spectra have a large impact on the OWT guided approach to estimate the water quality products, as the obtained OWTs are directly dependent on the shape of the reflectance spectra. For the water quality products, the algorithms from different band ratios based (often TOA reflectance) on OWTs gave the best results for both satellites. The MSI outperformed OLCI every time. As there are more issues with OLCI (atmospheric correction, spatial resolution, and masking), we recommend using MSI for national monitoring with OWT guiding prior application of the algorithms for the estimation of the water quality or the ecological state of the lakes.      Figure A3. The comparison of the in-situ and MSI C2X spectral bands. Figure A4. The comparison of the in-situ and OLCI C2RCC spectral bands.