Spectral Absorption Characteristics and Phytoplankton Dynamics Across Optical Water Types: Evaluating Sentinel-2 and Sentinel-3 Phytoplankton Absorption Retrieval Accuracy in Boreal Lakes

Highlights

What are the main findings?

• Based on in situ data, phytoplankton absorption correlates well with Chl-a; R² was stronger in the red spectral region (670 nm).
• Accuracy of the satellite retrieval of a_pig, using C2RCC, is low for both Sentinel-2 and Sentinel-3.

What are the implications of the main findings?

• Low R² and high error metrics of a_pig indicate the need for enhancements in the performance of C2RCC.
• OWT-specific approach to derive Chl-a from a_pig data may be beneficial, but only if satellite detection of a_pig improves significantly.

Abstract

Accurate detection of chlorophyll-a (Chl-a) is critical for monitoring water quality in inland waters, where high concentrations of coloured dissolved organic matter (CDOM) complicate retrieval process. Reliable Chl-a estimation depends on the precise determination of the phytoplankton absorption coefficient (a_ph). This study evaluates Chl-a detection from in situ a_ph measurements and assesses the accuracy of phytoplankton absorption retrieval from Sentinel-2/MSI (S2) and Sentinel-3/OLCI (S3) using the Case-2-Regional-Coast-Colour (C2RCC) processor across diverse optical water types (OWTs) in boreal lakes. OWTs were classified based on remote sensing reflectance features, representing Clear, Moderate, Turbid, Very Turbid, and Brown conditions. CDOM absorption strongly influenced the underwater light field, particularly in Brown and Turbid waters. Linear relationships between a_ph and Chl-a were generally strong across OWTs, with improved relationships in the red spectral region (670 nm). Satellite-derived a_pig estimates showed a weak relationship with in situ data (R² = 0.26–0.45). Both sensors overestimated small a_ph values, while S3 underestimated larger ones. S2 underestimated a_ph in Clear and Brown OWTs, with median absolute percentage differences near 100% for all OWTs. These findings emphasize the challenges posed by bio-optical complexity in boreal lakes and highlight the need for OWT-specific algorithms to improve satellite-based absorption and Chl-a retrieval accuracy.

1. Introduction

Phytoplankton are a key element in aquatic food chains and are often characterized by the amount of the main photosynthetic pigment, chlorophyll-a (Chl-a). Phytoplankton pigment absorption (a_ph) is an important parameter in models for remote sensing retrieval of Chl-a, whose determination using remote sensing methods remains challenging in turbid inland and coastal waters [1,2]. In these optically complex waters, light is absorbed not only by phytoplankton, as is typical in the open ocean, but also by various organic and mineral substances that do not correlate with each other, making accurate detection of Chl-a and phytoplankton-specific grouping difficult. The magnitude of a_ph varies with pigment concentration, with additional variability introduced by the pigment packaging effect [3]. Understanding the spectral, temporal, and spatial variability of phytoplankton absorption is essential to enhancing knowledge on phytoplankton ecological variability and to improving satellite-based detection in optically complex inland waters.

The Water Framework Directive (WFD, Directive 2000/60/EC) requires the monitoring of and reporting on the ecological status of all lakes larger than 0.5 km²—an initiative and obligation that have led to a stronger focus on lakes, including setting the classification of lake types and characterizing their ecological status on a European scale Phytoplankton are the central biological element; assessments in various European countries include Chl-a concentration, phytoplankton biomass, biomass of specific groups, taxonomic composition, focus on species preferring specific conditions, and indices based on taxonomic composition [4,5]. Currently, satellites can support monitoring of various parameters; e.g., Chl-a concentration, phytoplankton biomass, transparency, and bloom information [6,7,8]. More detailed phytoplankton characterization is, however, more closely linked to hyperspectral sensing capabilities [9,10].

Five multispectral satellites in the European Space Agency’s (ESA’s) Copernicus program support ocean colour mapping [11]: Sentinel-2A, 2B, and 2C/Multispectral Instrument (S2/MSI), which have lower spectral resolutions but higher spatial resolutions (10 × 10 m) [12], and Sentinel-3A and 3B/Ocean and the Land Colour Instrument (S3/OLCI), which provide daily revisit times over Estonia but with coarser spatial resolution (300 × 300 m), suitable mainly for mapping larger water bodies [13]. These satellites can be used to characterize changes in water bodies. Finding the best atmospheric correction remains an active field of research [14,15,16]. Only a few available atmospheric correction algorithms for S2 and S3 provide inherent optical properties as output (e.g., Case-2 Regional Coast Color (C2RCC) [17] and Atmospheric Correction for Optical Water Types (A4O) [18]), while others provide concentrations of optically active substances (e.g., the POLYnomial-based algorithm applied to MERIS (POLYMER) by Steinmetz [19] and ACOLITE by Vanhellemont & Ruddick [20]). Recent research on Chl-a retrieval from the study area [21] focused to in-water algorithm (ACOLITE, POLYMER, C2RCC, and A4O) validation for retrieving Chl-a, TSM, total absorption, and absorption components in optically complex waters. Total absorption was retrieved more reliably than its individual components, including phytoplankton absorption.

The division of waters according to optical water types is not a new approach—Jerlov started with ocean optical classification as early as 1951 [22,23]. The earliest and simplest optical classification distinguished between Case 1 and Case 2 waters [24], followed by more detailed schemes [25,26,27,28], with up to 25 types identified by Spyrakos [29]. Here, five optical water types, based on spectral features of remote sensing data, are derived from optically complex boreal lake data [30,31]. The OWT approach helps disentangle the overlapping optical signatures of multiple constituents in turbid waters [28,32] and supports selection of the best algorithms for describing spatial and temporal changes to in-water particles across various optically complex water bodies.

The aim of this article is to characterize absorption properties measured in boreal lakes according to their optical water types, and to determine the strongest relationship for remote sensing applications to derive Chl-a from a_ph measurements. A validation example of a_ph from S2 and S3 satellite data using the atmospheric correction processor C2RCC and in situ data is provided. An overview of the seasonal variation in phytoplankton community composition across different OWTs is presented, with a link to OWTs in various WFD lake types.

2. Materials and Methods

2.1. Laboratory Measurements

Water samples for laboratory analyses were gathered three times over the vegetation period (April–May in spring, July August in summer, and September–October in autumn) of three years (2018–2020) from various small lakes in Estonia (30 lakes in total) (Figure 1). Samples were collected from a depth of 0.5 m and stored in the dark until filtration and further analyses. For the large lakes Lake Võrtsjärv and Lake Peipsi, integrated water samples—collected from 1 m above the bottom at 0.5 m intervals and pooled together—were analysed as part of the national monitoring programme (May–October 2016–2020). Absorption parameters (phytoplankton absorption a_ph, absorption of non-algal particles a_dp) were measured according to Tassan & Ferrari [33], using NaClO as a solvent. Samples for Chl-a were extracted with 96% ethanol, measured spectrophotometrically, and calculated according to Jeffrey & Humphrey [34]. Samples for coloured dissolved organic matter (CDOM) were filtered through 0.2 µm cellulose acetate filters and measured spectrophotometrically with a Hitachi U-3010 (Hitachi Ltd., Tokyo, Japan) spectrophotometer [35]. The phytoplankton community composition, phytoplankton and cyanobacterial biomass were assessed using light microscopy [36] (Zeiss Axio Vert.A1 (Carl Zeiss, Jena, Germany)). At least 400 units were counted and converted to biovolume according to the measurements of cells, trichomes, or colonies, approximating them to simple geometric shapes [37]; biomass is reported as wet weight [38].

Figure 1. An overview of the study area and locations of the studied lakes (red dots).

2.2. In Situ Measurements of Remote Sensing Reflectance

The water-leaving reflectance spectra were calculated from:

(1) Measurements with three above-water TriOS-RAMSES hyperspectral radiometers, following the protocol of REVAMP [39]. The calculations included the following steps: first, all measured radiance and irradiance spectra were corrected for stray light [40]; second, water-leaving reflectance ρw(λ) was calculated using remote sensing reflectance Rrs(λ) [41].

$$\mathsf{\rho }\mathrm{w}(\mathsf{\lambda })=\mathsf{\pi }\mathrm{R}\mathrm{r}\mathrm{s}(\mathsf{\lambda })=\mathsf{\pi }{\displaystyle \frac{\mathrm{L}\mathrm{u}(\mathsf{\lambda })-\mathsf{\rho }(\mathrm{W})\mathrm{L}\mathrm{d}(\mathsf{\lambda })}{\mathrm{E}\mathrm{d}(\mathsf{\lambda })}}$$

(1)

where $\mathrm{L}\mathrm{u}\left(\mathsf{\lambda }\right)$ is the upwelling radiance from the water, $\mathrm{L}\mathrm{d}\left(\mathsf{\lambda }\right)$ is the downwelling radiance from the sky, $\mathrm{E}\mathrm{d}\left(\mathsf{\lambda }\right)$ is the downwelling irradiance, and $\mathsf{\rho }$(W) is the water surface reflectance as function of wind speed W (m·s⁻¹), calculated according to Equation (2).

$$\mathsf{\rho }\left(\mathrm{W}\right)=0.0256+0.00039\mathrm{W}+0.000034{\mathrm{W}}^2$$

(2)

Third, the Near-Infrared (NIR) similarity correction with λ₁ = 720 nm, λ₂ = 780 nm, and α = 2.35 was applied to the water-leaving reflectance [42,43].

(2) Measurements with two TriOS-RAMSES hyperspectral radiometers for in-water measurements. Five measurements were recorded at each depth, then Rrs(λ) was retrieved:

$$\mathsf{\rho }w(\mathsf{\lambda })=\mathsf{\pi }\mathrm{R}\mathrm{r}\mathrm{s}(\lambda )=\mathsf{\pi }{\displaystyle \frac{\mathrm{L}\mathrm{u}(\mathsf{\lambda })}{\mathrm{E}\mathrm{d}(\mathsf{\lambda })}}$$

(3)

where Rrs(λ) is the remote sensing reflectance, Lu(λ) is the upwelling radiance from the water and Ed(λ) is the downwelling irradiance. The NIR similarity correction was again applied to the water-leaving reflectance [43].

The reason for using two methods is related to the availability of reflectance data, as for the majority of small-lake samples, reflectance data measured with a two-sensor TriOS-RAMSES system were available. We used only measurements where one sensor was placed above the water surface and the other below the surface to reduce sun-glint effects.

Based on in situ radiometric measurements, optical water types were calculated based on remote sensing reflectance spectra according to Uudeberg [44].

2.3. Satellite Data

We wanted to test an atmospheric correction processor, which is available for both S2 and S3, is applicable to all our lake types, and gives phytoplankton absorption as an output. This was the reason behind choosing C2RCC.

S2/MSI Level 1 (L1) and S3/OLCI L1 and Level 2 (L2) full-resolution (FR) non-time critical data were downloaded for the period 2016–2020:

• S2 A/B MSI data were downloaded from the Copernicus Open Access Hub (https://scihub.copernicus.eu, accessed on 10 October 2023), with processing baselines of 02.04–04.00, data now available for download from Copernicus Data Space Browser (https://browser.dataspace.copernicus.eu/, accessed on 10 October 2023).
• S3 A/B OLCI L1 data were downloaded from the EUMETSAT Data Store (https://data.eumetsat.int, accessed on 1 October 2023) for 2016–2019 (reprocessed), or from the EUMETSAT Data Centre (https://archive.eumetsat.int/, accessed on 12 October 2023) for 2020 with a processing baseline of 002, data now available for download from Copernicus Data Space Browser (https://browser.dataspace.copernicus.eu/, accessed on 12 October 2023).

L1 data for both S2 MSI and S3 OLCI were processed using C2RCC (v.2.1) [17,45] atmospheric correction. Its water part uses a neural network that estimates five inherent optical properties at 443 nm from the water-leaving reflectance spectrum [17]; here, we focused on a_pig. The following C2RCC flags were used for S3: Rhow_OOS, Cloud_risk, Rhow_OOR, Rtosa_OOR, quality_flags_invalid, Rtosa_OOS, and quality_flags_sun_glint_risk.

For S2 the first approach consisted of the application of commonly used C2RCC flags: Rtosa_OOS, Rtosa_OOR, Rhow_OOS, Rhow_OOR Cloud_risk, Valid_PE and IdePix (9.0.2) [46] flags: IDEPIX_VEG_RISK, IDEPIX_INVALID, IDEPIX_CLOUD, IDEPIX_CLOUD_AMBIGOUS, IDEPIX_CLOUD_SURE, IDEPIX_CLOUD_BUFFER, IDEPIX_CIRRUS_SURE, IDEPIX_CLOUD_SHADOW, IDEPIX_CIRRUS_AMBIGUOUS, IDEPIX_BRIGHT, IDEPIX_POTENTIAL_SHADOW, IDEPIX_LAND, and IDEPIX_VEG_RISK.

For the second approach, the following flags were applied in addition to those used in the first approach to ensure that values were within the neural network training range (Apig, Adet, Agelb, Bpart, Bwit_at_max; Apig, Adet, Agelb, Bpart, and Bwit at_min) and to remove datapoints outside this range. Additionally, filtering was performed based on satellite-derived uncertainty estimates for a_pig, removing values with uncertainty >5 m⁻¹.

The salinity was set to 0.001 PSU, which differs from the default settings, as we were focusing on freshwater conditions. Regions of interest (ROIs) of 1 × 1 pixels, matching in situ sampling locations, were exported using the Pixel Extraction tool in Sentinel Application Platform (SNAP v9.0) software. For match-up analysis, the S3 validation protocol was followed [47] for both S2 MSI and S3 OLCI. An in situ dataset, used for validation, was collected on the same day as the satellite overpass and consisted of 109 water samples for S2 and 118 water samples for S3, including 20 water samples for S2 and 66 water samples for S3 without radiometric data or OWT classification. Satellite observations were taken on 30 different dates; however, several data points originate from the same day because fieldwork was intentionally planned during clear-sky conditions and S2 overpasses. The maximum number of small lakes visited in one day was six.

2.4. Error Metrics

Statistical analysis was performed using R software (v2025.05.1). Standard deviations for error estimation of in situ data were calculated when multiple measurements were available for one water sample or one measurement station. MDE (mean difference error) measures the average signed difference between predicted and observed values as follows:

$$\mathrm{M}\mathrm{D}\mathrm{E}={\sum }_{\mathrm{i}=1}^{\mathrm{N}}(\mathrm{X}\mathrm{s}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{l}\mathrm{l}\mathrm{i}\mathrm{t}\mathrm{e},\mathrm{i}-\mathrm{X}\mathrm{i}\mathrm{n} \mathrm{s}\mathrm{i}\mathrm{t}\mathrm{u},\mathrm{i})$$

(4)

The median absolute percentage error (MdAPE) was calculated to estimate the dispersion of the data around the 1:1 line.

$$\mathrm{M}\mathrm{d}\mathrm{A}\mathrm{P}\mathrm{E}=\underset{1\le \mathrm{i}\le \mathrm{N}}{\mathrm{m}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{n}}\left\{|{\displaystyle \frac{\mathrm{X}\mathrm{s}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{l}\mathrm{l}\mathrm{i}\mathrm{t}\mathrm{e},\mathrm{i}-\mathrm{X}\mathrm{i}\mathrm{n} \mathrm{s}\mathrm{i}\mathrm{t}\mathrm{u},\mathrm{i}}{\mathrm{X}\mathrm{i}\mathrm{n} \mathrm{s}\mathrm{i}\mathrm{t}\mathrm{u},\mathrm{i}}}|\right\}\times 100$$

(5)

R² value (Equation (6)) indicates the fit of the model to the data; higher values represent a better fit.

$${\mathrm{R}}^2=1-{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{(\mathrm{X}\mathrm{i}\mathrm{n} \mathrm{s}\mathrm{i}\mathrm{t}\mathrm{u},\mathrm{i}-\mathrm{X}\mathrm{s}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{l}\mathrm{l}\mathrm{i}\mathrm{t}\mathrm{e})}^2}{{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\left(\mathrm{X}\mathrm{i}\mathrm{n} \mathrm{s}\mathrm{i}\mathrm{t}\mathrm{u},\mathrm{i}-\mathrm{X}\mathrm{i}\mathrm{n} \mathrm{s}\mathrm{i}\mathrm{t}\mathrm{u},\mathrm{a}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{e}\right)}^2}}$$

(6)

Root mean square error (RMSE) was used to estimate the model error.

Normality was assessed using the Shapiro–Wilk test [48], which evaluates the null hypothesis that the data are drawn from a normally distributed population. Following best-practice guidelines, p > 0.05 was interpreted as a failure to reject normality.

2.5. Water Framework Directive Lake Types

Estonian lakes are classified according to five key parameters: water area, stratification, water colour, chloride content, and water hardness (Table 1) [49].

Table 1. Relevant parameters for lake classification.

Parameter	Classes	Definition/Threshold
Water area	Small	<10 km2
	Large	100–300 km2
	Very large	>1000 km2
Stratification	Stratified	Thermal stratification present
	Unstratified	Entire water column is mixed
Water colour	Light	Absorption coefficient at 400 nm (a400) < 4 m−1; Pt-Co < 100°
	Dark	a400 ≥ 4 m−1; Pt-Co > 100°
Chloride content	High	>25 mg/L
	Low	<25 mg/L
Water hardness	Soft	HCO−3 < 80 mg/L
	Medium	80–240 mg/L
	Hard	>240 mg/L

Based on these criteria, eight Water Framework Directive (WFD) lake types are defined (Table 2).

Table 2. Classification of WFD lake types, based on parameter values from Table 1.

WFD Lake Type	Water Area	Stratification	Water Colour	Chloride Content	Water Hardness	Notes
I	Small	Unstratified	Light	Low	Hard
II	Small	Unstratified	Light	Low	Medium
III	Small	Stratified	Light	Low	Medium
IV	Small	Unstratified	Dark	Low	Low
V	Small	Unstratified	Light	Low	Low
VI	Large	Unstratified	-	Low	Medium	Võrtsjärv type
VII	Very large	Unstratified	-	Low	Medium	Peipsi type
VIII	-	-	-	High	-	Coastal lakes, within 5 km of sea

In this article, WFD lake types I and VIII were excluded, as these lakes are generally transparent with a visible bottom, and were therefore not visited during remote sensing-focused fieldwork.

3. Results

3.1. Overview of Optically Active Substances in Different Optical Water Types

General characteristics of the dataset: Chl-a ranged from 0.01 to 109.6 mg m⁻³ (average, 23.2 mg m⁻³); Secchi depth was in the range 0.1–7 m (average, 1.6 m); absorption by CDOM (a_CDOM(443)) was in the range 0.7–47.3 m⁻¹ (average, 4.1 m⁻¹); and the amount of total suspended matter (TSM) ranged from 0.8 to 143.3 g/m³ (average, 8.4 g/m³).

An example of various optical water types is shown in Figure 2. Clear and Moderate OWTs have maximum reflectances between 540 and 580 nm. For the Clear OWT, the reflectance at 500 nm exceeds that at 650 nm and is also higher than the reflectance at 500 nm for the Moderate OWT. The Turbid OWT has a reflectance maximum between 580 and 605 nm. The Very Turbid OWT has a reflectance maximum at wavelengths between 685 and 715 nm. The Brown OWT has a maximum in the red part of the spectrum, whereas reflectance tends to be below 0.006.

Figure 2. An example of reflectance spectra for five different optical water types.

Absorption by CDOM largely determined the underwater light field (Figure 3). This effect was most pronounced in the Brown water type, where CDOM absorption accounted for nearly 100%. In the Moderate and Clear types, absorption by all components ranged from 10% to 50%, with both phytoplankton (a_ph) and CDOM dominating. Turbid waters exhibited high CDOM absorption, whereas Very Turbid waters showed a stronger contribution from phytoplankton (up to 55%).

Figure 3. The share (%) of different absorbers (absorption by phytoplankton, a_ph; absorption by CDOM, a_CDOM; and absorption by non-algal particles, a_dp) at 443 nm. The colours represent various OWTs (C—Clear, M—Moderate, B—Brown, T—Turbid, and VT—Very Turbid).

3.2. OWTs in Different WFD Lake Types

WFD classification for Estonian lakes partly relies on water colour, distinguishing highly absorbing waters (Type IV). This type corresponded well to the Brown OWT with a small share (4%) of the Moderate OWT (Figure 4). The main difference between WFD lake types II and III is the presence of summer stratification in type III. Both types had similar shares (58–59%) of the Moderate and Turbid OWTs, but Type II included a larger proportion of Very Turbid waters (27%), whereas Type III had more Clear waters (28%). Type V consists mainly of Moderate waters (56%), with Brown (35%) and a minor share of Very Turbid (9%) waters. Type VI (Lake Võrtsjärv) is predominantly Turbid (77%), occasionally Very Turbid (15%), and includes small proportions (4%) of Moderate and Brown OWTs. Type VII (Lake Peipsi) is mostly Moderate (47%), with nearly equal shares of Brown, Turbid, and Very Turbid OWTs.

Figure 4. The share of different optical water types (C—Clear, M—Moderate, B—Brown, T—Turbid, and VT—Very Turbid) in the six lake types used for WFD reporting in Estonia.

3.3. Phytoplankton Absorption in Relation to Chlorophyll-a Concentration

A moderate-to-strong linear relationship was observed between phytoplankton pigment absorption and Chl-a concentration (Figure 5). In the Clear OWT, Chl-a remained below 15 mg/m³, corresponding to low pigment absorption (up to 0.4 m⁻¹). The Moderate OWT exhibited Chl-a up to 60 mg/m³, with a_ph(443) reaching 1.7 m⁻¹. The Brown OWT generally had Chl-a below 80 mg/m³, except for one extreme value near 100 mg/m³ from Lake Valguta Mustjärv; most a_ph(443) values were below 2 m⁻¹. The Turbid OWT reached Chl-a up to 70 mg/m³ and a_ph(443) up to 2.7 m⁻¹, while the Very Turbid OWT exceeded 100 mg/m³, with a_ph(443) peaking at 3.8 m⁻¹.

Figure 5. In situ-measured phytoplankton absorption (a_ph) in relation to in situ-measured Chl-a amount for 5 different optical water types, (a) at 443 nm and (b) at 670 nm. Note the difference in x and y scales.

The effect of season was visible only in Turbid and Very Turbid OWTs, where the majority of spring values were lower in comparison with summer and autumn values. At the same time, in the Moderate OWT, the highest values for a_ph and Chl-a were found in spring. In the Clear OWT, higher values were from spring and autumn.

The linear relationship between measured a_ph and Chl-a was strongest in the case of the Turbid OWT (R² = 0.96), and the weakest in the case of Clear (R² = 0.53) and Brown OWTs (R² = 0.56). The use of the red spectrum (670 nm) resulted in a stronger relationship between a_ph and Chl-a for all OWTs. RMSE for Chl-a retrieval from a_ph(443) was the highest in the case of Very Turbid (14.4 mg/m³) and Brown OWTs (12.7 mg/m³), and the lowest in the case of the Clear OWT (1.72 mg/m³). The RMSE was 3.89 mg/m³ for the Turbid OWT and 5.91 mg/m³ for the Moderate OWT. For all OWTs, Chl-a retrieval from a_ph(670) resulted in a lower RMSE.

Because the Chl-a-a_ph relationship differs among OWTs, model transformation was applied on OWT-specific bases. Log-scale residuals for Clear, Moderate, and Brown OWTs showed Shapiro–Wilk p-values exceeding 0.05, indicating no significant deviation from normality, supporting the use of log–log models for these optical water types. The resulting log-scale models for these three OWTs were used to model Chl-a, whereas a linear model was retained for the Turbid and Very Turbid OWTs.

Model performance is shown in Figure A1; the coefficient of determination (R²) was higher for the Turbid OWT, despite the use of log-scale models for the Clear, Brown and Moderate OWTs. However, Chl-a retrievals achieved lower MDAPEs (%) for the Clear, Brown, and Moderate OWTs—for both predictors (a_ph(443) and a_ph(670)) under the log-scale transformation—and a lower MDE when modeling from a_ph(670) (Appendix A, Table A1 and Table A2).

3.4. Phytoplankton Biomass and Communities in Different OWTs

Having established how phytoplankton absorption varies with Chl-a across OWTs and seasons, the next section focuses on broader ecosystem characteristics by examining phytoplankton biomass and community composition within the same OWT framework.

3.4.1. Phytoplankton and Cyanobacterial Biomass

Phytoplankton total biomass and cyanobacterial biomass were the lowest in Clear OWT (<10 g/m³), with minimal variability and the lowest values in spring (Figure 6). In the Moderate OWT, biomass peaked in summer, reaching 20 g/m³. The Brown OWT generally showed lower biomass in spring and higher in summer, with occasional extreme values driven by cyanobacteria—an atypical feature caused by one lake with exceptionally high pH (>9). In the Turbid OWT, biomass was lowest in spring and increased steadily toward autumn (>30 g/m³), with cyanobacteria dominating in summer and autumn. The Very Turbid OWT exhibited high biomass already in spring due to cyanobacteria, reaching 50 g/m³ in summer, with the greatest variability during summer and autumn.

Figure 6. Phytoplankton biomass (TBM) and cyanobacterial biomass (CyBM) based on in situ data for different OWTs. (a) Clear, (b) Moderate, (c) Brown, (d) Turbid, and (e) Very Turbid (VT).

3.4.2. Seasonal Changes in Phytoplankton Community Composition in Selected Lakes

Here, an example of phytoplankton community structure across various OWTs is given. In the presented lakes, OWTs remained consistent throughout the vegetation period. Seasonal changes were present in all types of lakes. Diatoms were present in all samples except in the Clear OWT during spring (Figure 7). In the Clear OWT, phytoplankton biomass was the lowest (Figure 6), and the phytoplankton community was characterized throughout the vegetation period by chlorophytes, while cyanobacteria and Desmidiales dominated during summer and autumn (Figure 7). In all selected lakes, phytoplankton biomass was lowest in springtime, except for the Brown OWT, which had the lowest biomass in summer. The Brown OWT exhibited a peak in biomass in autumn, driven by the raphidophyte Gonyostomum semen. Phytoplankton biomass increased in selected lakes from spring to autumn in the Clear, Moderate, and especially in the Very Turbid OWT. The biomass of cyanobacteria increased from summer to autumn; their biomass increased from Clear to Moderate to Turbid OWTs, being the highest in the Very Turbid OWT. In the Turbid OWT, phytoplankton biomass was highest in summer, with clear cyanobacterial dominance.

Figure 7. An example of seasonal dynamics of phytoplankton biomass (TBM) of various groups (Bac—diatoms, Chloro—chlorophytes, Chryso—chrysophtes, Crypto—cryptophytes, Cy—cyanobacteria, Desmi—desmidiales, Dino—dinoflagellates, Eugleno—euglenophytes, Raphi—raphidophytes, and Xantho—xanthophytes) in selected lakes belonging to different OWTs. Note the differences in the y-axes.

3.5. Satellite Retrieval of Phytoplankton Absorption and Validation

S2 initially retrieved much larger a_pig values (up to 16 m⁻¹) than S3, which obtained values below 3 m⁻¹ (Figure 8). All unrealistically large values retrieved by S2 belonged to the Very Turbid OWT and were associated with high uncertainties. Further analysis of S2 data was carried out after removing values with uncertainties exceeding 5 m⁻¹.

Figure 8. Phytoplankton absorption (a_pig) from different satellites, (a) Sentinel 2 (S2) and (b) Sentinel 3 (S3), in comparison with in situ-measured phytoplankton absorption (a_ph) values. Y-scale error bars denote satellite uncertainties, x error bars denote in situ variability. Colours represent various OWTs; grey is for untyped water samples lacking radiometric measurements. The dashed line is a 1:1 line. Black solid line represents a linear relationship between parameters.

The coefficient of determination between in situ-measured a_ph and satellite-estimated a_pig was low (R² = 0.11–0.26). After removing values with high uncertainties, this relationship increased slightly (0.29) for S2 (Figure 9), accompanied by a decrease in RMSE and MDAPE% (Table 3).

Figure 9. Phytoplankton absorption (a_pig) from S2, using C2RCC. (a) Commonly used flags (approach 1); (b) commonly used flags + additional flags pertaining to training ranges in comparison with in situ-measured phytoplankton absorption values (a_ph). x error bars denote variability between in situ samples. Colours represent various OWTs; grey is for unclassified water samples lacking radiometric measurements. The dashed line is a 1:1 line. Black solid line represents a linear relationship between parameters.

Table 3. Error metrics for C2RCC performance (S2 and S3). * denotes S2 data after removing values with high uncertainties (>5 m⁻¹).

Sat	N	MDE	MdAPE%	RMSE	R2
S2	109	0.27	99.9	2.31	0.11
S2 *	105	0.25	99.82	0.81	0.29
S3	118	0.14	46.18	0.59	0.26

S3 tended to overestimate smaller a_ph values and underestimate higher ones (Figure 8b). As the dataset for S3 originated mainly from the large lakes Peipsi and Võrtsjärv, which are generally more turbid, the dataset included only one datapoint from the Brown OWT and three from the Clear OWT, specifically meaning there were no low a_ph values represented.

For S2, several values close to zero were observed for Clear and Brown OWTs (Figure 9). Most values from Turbid and Very Turbid OWTs were overestimated, while values from Brown and Clear OWTs were underestimated (Table 4, MDE).

Table 4. Error metrics for a_pig retrieval from S2 (with commonly used flags; approach 1) and S3 data in different OWTs.

Sat	OWT	N	MDE	MdAPE%	RMSE	R2
S2	Brown	9	−0.53	99.82	0.97	0.01
S2	Clear	19	0.02	99.86	0.42	0.58
S2	Moderate	20	0.22	99.86	0.47	0.35
S2	Turbid	29	0.52	122.38	1.06	0.15
S2	Very Turbid	8	0.83	99.71	1.42	0.17
S2	Undetermined	20	0.43	47.57	0.46	0.86
S3	Brown	1	2.32	641.07	2.32	0
S3	Clear	3	0.11	56.48	0.32	0.81
S3	Moderate	11	0.15	65.18	0.25	0.06
S3	Turbid	23	0.14	53.09	0.62	0
S3	Very Turbid	14	0.25	29.81	0.71	0.35
S3	Undetermined	66	0.13	45.71	0.52	0.25

When analyzing results by OWT, MDAPE was the lowest for S3 in the Very Turbid OWT (29.8%) and higher for the Turbid (53%), Clear (56%) and Moderate (65.2%) OWTs. For S2, the Turbid OWT showed the largest dispersion in a_pig (>100%), while for other OWTs MDAPE% remained slightly below 100%.

Including S2 flags related to the neural network training range reduced the number of data points across all OWTs (Table 5) and increased the strength of the relationship, with R² improving to 0.45 for the entire dataset (Figure 9b). This adjustment improved detection in the Turbid and Very Turbid OWTs (MDE decreased from 0.56 to 0.45 and from 0.83 to 0.70, respectively), along with lower RMSE. R² for the Clear OWT increased to 0.6.

Table 5. Error metrics for a_pig retrieval from S2 with all flags (approach 2) applied in different OWTs.

Sat	OWT	N	MDE	MdAPE%	RMSE	R2
S2	All	75	0.27	91.62	0.72	0.45
S2	Brown	5	−0.53	99.81	0.86	0.04
S2	Clear	9	−0.06	99.95	0.54	0.60
S2	Moderate	11	0.25	99.90	0.52	0.44
S2	Turbid	24	0.45	102.05	0.87	0.35
S2	Very Turbid	6	0.70	80.80	1.10	0.38
S2	Undetermined	20	0.43	47.57	0.46	0.86

4. Discussion

Space-based observations of ocean colour remain the primary approach for estimating chlorophyll-a concentration in the surface layer of both oceanic and inland waters at regional-to-global scales [50,51]. Applying OWTs to reflectance spectra enables the development of OWT-specific methods for detecting in-water substances [31].

The performance of any optical model depends on the interaction between phytoplankton pigments, community composition, and background water constituents. As shown in Figure 2, the lakes examined here exhibit a high proportion of CDOM absorption, which strongly shapes the underwater light climate and strongly reduces Chl-a detection in the blue region. CDOM darkens the water and decreases the depth of the euphotic zone [52,53]. However, dark water does not necessarily indicate low Chl-a concentrations, as inputs of allochthonous organic material may increase nutrient availability and promote phytoplankton growth [53].

Retrieving accurate satellite signals from such dark waters is challenging (as also shown in Section 3.3), because most light is absorbed and only a small fraction reaches the sensor, making the instrument’s signal-to-noise ratio particularly important. This is evident in S2 retrievals (Figure 8 and Figure 9), where most Brown OWT values were retrieved as near-zero. In this OWT, strong CDOM absorption overlaps with phytoplankton pigment absorption (especially <500 nm), which makes a_ph(443) a less reliable predictor of Chl-a than a_ph(670).

In the Clear OWT, CDOM influence is low, but pigment concentrations are also low, still limiting satellite detectability. Turbid and Very Turbid OWTs have higher Chl-a concentrations and phytoplankton biomass than other OWTs, which should in principle improve retrieval performance. However, S2 retrievals for the Very Turbid OWT frequently exhibited high uncertainties, demonstrating that high biomass values alone do not guarantee reliable retrievals when factors such as high turbidity, backscattering variability and adjacency effects also affect reflectance.

We used C2RCC because it has demonstrated strong statistical performance in inland waters, alongside POLYMER, when compared to ACOLITE’s DSF mode, iCOR, L2gen, and Sen2Cor [54,55]. C2X is another processor, known to perform well in turbid and very turbid waters [56,57], but because it rejects many oligotrophic–mesotrophic pixels [56], it is not suitable for the full range of lake types considered here. C2RCC is available for both S2 and S3 and is specifically designed for Case II waters [58], making its training range suitable for boreal and temperate lakes. Yet, recent work shows that no atmospheric correction algorithm consistently outperforms others across diverse atmospheric and water conditions [14,59], highlighting the need for continual refinement of atmospheric correction schemes. Accurate atmospheric correction must also address adjacency effects, which influence all small lakes, and the RadCor processor has shown promising results [60].

A4O-ONNS represents a further development of C2RCC for S3 by incorporating OWTs [18,61]. Its performance has been tested for optically complex waters [21], and A4O-ONNS (v1.0) retrieved a_pig(440) with slight improvement (R² = 0.43), but with systematic overestimation (MdPE = 72%).

The relationship between the Chl-a concentration and the phytoplankton absorption coefficient varies across optically complex inland waters [62]. The modified quasi-analytical algorithm (MQAA) applied to optically complex lake data produced a linear model for Chl-a (up to 45 mg/m³) and inversed a_ph(674) with R² = 0.87, resulting in a similar R² as linear relationship between measured Chl-a concentration and a_ph(674) [63]. The Chl-a retrieval uncertainties in most semi-analytical algorithms from satellite data are primarily determined by phytoplankton absorption and composition, while machine leaning algorithms show high sensitivity to CDOM and detritus absorption [64]. This again highlights the importance of in situ absorption measurements and phytoplankton community composition information.

The key limitation is evident—accurate Chl-a retrieval is only possible if a_pig is reliably retrieved. The C2RCC optical model for Case II waters is based on the MERIS Case II bio-optical model, which was trained using the NOMAD dataset, containing measurements from the North and Baltic Seas, and the Coastcolour dataset, representing mainly open ocean, sea, or coastal waters [65], suggesting the need for the inclusion of data from various inland waters. Because C2RCC computes Chl-a directly from a_pig using CHL_NN = 21 × a_pig¹·⁰⁴, even a small 0.1 increase in a_pig translates into a ~2 mg m⁻³ change in Chl-a. Our relatively weak relationships between measured and satellite-derived a_pig (Figure 8 and Figure 9, Table 3, Table 4 and Table 5) suggest the need for improved a_pig detection models. An OWT-specific approach (Figure 5) may enhance Chl-a detection, but only if a_pig retrieval is reliable. There is strong potential for this, as in situ data-based analysis showed an R² up to 0.97 for Chl-a detection from a_ph, particularly in the Turbid OWT. A logarithmic model was more appropriate for Clear, Brown, and Moderate OWTs, resulting in lower MdAPE%, while the Turbid OWT achieved the highest accuracy (relatively low RMSE and high R²).

The selection of quality and range flags is important. Adding additional flags to S2, related to the training range of IOP parameters, reduced the number of data points and increased R². However, S2 still produced extremely low and high a_pig values, which were not eliminated by flagging. Extremely large a_pig values originated from two lakes, Kaiu and Joemoisa, both classified into the Very Turbid OWT during summer and autumn, where high CDOM and phytoplankton biomass coexist. Figure 7 illustrates cyanobacterial dominance and high biomass in Joemoisa Lake. As these lakes are connected, phytoplankton composition is similar in both, yet measured a_ph values were much lower than satellite-derived values.

Lake Valguta Mustjärv (Brown OWT) is another exception among dark lakes—despite its dark water, it has a high pH and annual cyanobacterial blooms, attributed to historical nutrient additions for lawn growth in the bog during preparations for the 1980 Moscow Olympic Games [66]. Additional raphidophyte blooms were observed in several dark-water lakes, such as Lake Meelva (Figure 7) and L. Valguta Mustjärv, typically occurring during summer and autumn. In L. Meelva, these blooms were a consistent feature throughout the study period, peaking in autumn, whereas in Valguta Mustjärv a Gonyostomum bloom occurred only in 2018. In non-bloom dark lakes, raphidophytes were absent (e.g., Nohipalu Mustjärv), total biomass was much lower (around 3 g/m³), and the community was dominated by cryptophytes and desmidiales. From a satellite imaging perspective, non-bloom dark lakes are particularly challenging because CDOM masks the blue spectral region and reflectance is low [67].

Chlorophyll-a content varies across phytoplankton taxonomic groups. The carbon:chlorophyll-a (C:Chl-a) ratio differs among taxa and responds to environmental conditions [68]. Cyanobacteria typically contain more accessory pigments and less chlorophyll-a per cell or per unit biomass compared to many eukaryotic algae (e.g., diatoms, chlorophytes, or picoeukaryotes) [69,70]. As eutrophication promotes cyanobacterial dominance, and because their pigment composition enhances their success in turbid environments [71,72], this group becomes more abundant in Turbid and Very Turbid OWTs (Figure 6 and Figure 7). Diatoms on the other hand have a high amount of Chl-a per cell, but the ratios are variable, depending on the genera [70]. Large diatoms are usually dominant in moderate nutrient conditions during spring and autumn, when mixing keeps them in the water column [73].

External factors—including light availability, nutrient concentrations, and temperature—also influence chlorophyll-a content. Low-light conditions tend to increase Chl-a content, whereas nutrient limitation generally reduces it [68,74]. Consequently, this variability makes using Chl-a as a proxy for phytoplankton biomass challenging, especially across diverse waterbodies where phytoplankton communities respond differently to variations in nutrient and light regimes.

Accurate detection of Chl-a using satellite remote sensing is important for WFD reporting, as EO data may be increasingly incorporated into future reporting frameworks [75]. Our analysis revealed that only one lake type (IV) was almost entirely characterized by a single OWT (Brown), which remained seasonally stable. Other WFD types exhibited substantial variability in OWTs, with shallower types II, VI, and VII showing more turbid OWTs than type III (Figure 4). Stratification tends to result in clearer water during summer, as surface nutrients are depleted, whereas in shallow lakes, wind-induced sediment resuspension increases available phosphorus, promoting phytoplankton growth [76,77]. Here, there were lower Chl-a and a_ph values in the Clear OWT during summertime, and, in Turbid and Very Turbid OWTs phytoplankton biomass, Chl-a amount, and a_ph increased from spring to autumn.

Generally, small lakes are characterized by one dominant OWT, while large lakes may exhibit multiple OWTs simultaneously. In the selected small lakes (Figure 7), where seasonal changes in phytoplankton community composition were examined, the OWT remained stable during the vegetation period. However, this is not universal, and seasonal variability of OWTs in small lakes must also be considered.

Overall, our findings highlight that the performance of current atmospheric correction and bio-optical models is strongly affected by CDOM content, low signal-to-noise conditions, and the diverse composition of phytoplankton communities. These uncertainties are compounded by seasonal and taxonomic shifts that alter phytoplankton absorption and the relationship between biomass and Chl-a. While OWT-specific approaches offer improvements, their success depends on accurate retrieval of underlying absorption parameters.

Future advances include hyperspectral sensors such as NASA’s PACE Ocean Colour Instrument (OCI), which promise to reduce uncertainties in pigment retrievals [10,78], improve detection of phytoplankton composition, and support development of new algorithms for optically complex waters [79,80,81]. Additional targeted calibration datasets will be essential to refine bio-optical models to better represent variability within lakes.

5. Conclusions

In the lakes we studied, CDOM was the main light absorber, especially in the Brown OWT, where it accounted for up to 100% of absorption at 443 nm. Despite this, in situ Chl-a was strongly correlated with in situ a_ph, especially in the Turbid OWT. Linear models performed well for most OWTs and were suitable for Turbid and Very Turbid OWTs, whereas logarithmic models were more appropriate for Clear, Moderate, and Brown OWTs. This improved Chl-a modelling from in situ data and resulted in lower MdAPE%. Across all OWTs, R² was higher and RMSE lower when using the red spectral region (670 nm). At the same time, a_pig detection accuracy was low when using C2RCC. The relationship between a_pig and a_ph was 0.26 for S3, based on a dataset with mostly Moderate, Turbid and Very Turbid OWTs. After removing the values retrieved with large uncertainty, the determination coefficient increased from 0.11 to 0.29 for S2, and with all flags applied, R² improved to 0.45. S2 tends to overestimate a_pig in Turbid and Very Turbid OWTs and underestimate it in Clear and Brown OWTs, often producing near-zero values. OWT distribution within WFD lake types revealed more Turbid and Very Turbid OWTs in shallow, non-stratified lakes and the Clear OWT in stratified lakes. Although the bio-optical complexity of boreal lakes still poses significant challenges, improved Chl-a retrievals provide more reliable inputs for using satellite data for lake management purposes and support implementation of environmental directives.