Water samples from the 25 August 2022 survey were filtered through Whatman Nuclepore™ track-etched membranes with a 0.2 µm pore size and 47 mm diameter. The filters were frozen and stored at -80 ∘C. Each frozen filter was put into a separate 20 mL glass vial before a 5 mL aliquot of 1× phosphate-buffered saline with a pH = 7 was added to begin a pigment concentration analysis. Sample sonification was completed in an ultrasonic bath filled with crushed ice; this was carried out four times in 1 min pulses with 60 s pauses in between. After sonification, the glass vial were centrifuged and three 250 µL aliquots of the supernatant were transferred onto a 96-microwell plate. A BioTek Instruments Synergy H1 hybrid multi-mode microplate reader was used to determine the absorbance of the samples. Phycocyanin concentrations (µg L-1) were computed as recommended in the literature (Horváth et al., 2013). The remaining supernatant was decanted, and the filters were frozen again at -22 ∘C. Samples were freeze-dried for a day, and 5 mL of 99.5 % ethanol was then added to each glass vial. Alcoholic extracts were prepared as explained above, and the samples were centrifuged before the supernatant from each sample was measured in the microplate reader. Chlorophyll-a concentrations (mg L-1) were determined following the standard protocol (Ritchie, 2008).

Hyperspectral absorption coefficients in the visible spectrum (400–700 nm) were determined directly after sampling using a point-source integrating cavity absorption meter, PSICAM (Kirk, 1997; Röttgers and Doerffer, 2007). The PSICAM had an Illumination Technologies CF1000e halogen lamp as the light source and used an Avantes AvaSpec ULS 2048XL-RS-EVO spectrometer as the detector. The instrument was calibrated using a solution of nigrosin with a maximum absorption coefficient of ∼ 0.5 m-1. Sartorius Airum^® Pro ultrapure water was used as a reference for both the calibration and sample measurements. For the bloom in 2021, as explained above, the bloom water was very concentrated; therefore, we diluted the sample at a ratio of 1:80 (Fig. 1d), and only the total absorption of the water constituents (atot) was measured. The samples of the 2022 campaign were analysed for both atot and absorption of coloured dissolved organic material (acdom). For this, the samples were filtered through 0.2 µm pore size Whatman Nuclepore™ track-etched membrane filters. As the atot samples were less dense than in 2021, the samples were only diluted at a ratio of 1:10. The acdom samples remained undiluted. The absorption coefficients of the particulate fraction were obtained via the subtraction of acdom from atot and denoted as aph, assuming no relevant contribution of non-algae particles.

The absorption coefficient measurements were performed in triplicate, meaning the sample was put in and out of the PSICAM three times, and each time an absorption spectrum measurement was obtained. Each measurement was the average of 20 single readings by the spectrometer, to minimize noise in the recorded signal, which could also contribute to a smoother spectrum. Data processing of absorption spectra incorporates smoothing using loess local regression which is important in case of very low absorption coefficients. However, we compared unsmoothed and smoothed spectra, and no significant differences were found, which was expected due to the high absorption of the samples. After the measurement of the samples, the respective dilution factor used was considered in the final calculation of the absorption coefficient values.

Bloom water was pre-filtered through Whatman 0.7 µm mesh size GF/F filters and subsequently filtered through Whatman Nuclepore™ track-etched 0.2 µm pore size membranes to remove suspended solids within 72 h of sampling on 16 August 2021. A similar approach was implemented for the August 2022 campaign, but filtration was applied within 2 weeks of field sampling. Sample filtrate from the 2021 campaign was diluted with deionized water at a ratio of 1:4 to avoid the inner-filter effect and detector saturation. No dilution was applied for the 2022 samples. A HORIBA Aqualog benchtop fluorometer was used to obtain excitation–emission matrices (EEMs) and absorbance measurements in a 1 cm quartz cuvette containing the sample filtrate. The cuvette had been pre-rinsed three times with the sample filtrate. Spectra were recorded at a 10 nm resolution for both excitation and emission determinations. Excitation was done at 2 nm intervals from the ultraviolet to red spectrum (200–600 nm), whereas fluorescence was measured over a wavelength range of 200 to 620 nm at 1.617 nm steps. Although emission raw data can be obtained up to 620 nm, we restrict the analyses to 600 nm. Laboratory experiments have revealed that our HORIBA Aqualog benchtop fluorometer has a relatively low signal-to-noise ratio and sensitivity beyond 600 nm. A three-dimensional dataset is generated from the emission, excitation and sample measurement.

Within the scope of this dataset description, an additional analysis was performed to highlight the added value of the obtained EEM measurements to the scientific community. Using the 2021 dataset only, the derived EEMs were corrected using the DrEEM toolbox and further analysed using PARAllel FACtor analysis, PARAFAC (Murphy et al., 2013). The PARAFAC correction removed Raman and Rayleigh scattering components from the measurements followed by a normalization to Raman units. Multivariate analysis of the three-dimensional dataset obtained from the EEMs allows for the isolation of principal components associated with specific moieties in the dissolved organic material pool. A dense algal bloom can be characterized, with caveats, by the increment in the metabolic processes responsible for the release of dissolved organic compounds such as polymers, amino acids residues and decaying cells. Therefore, PARAFAC was utilized to characterize changes in the main composition of surface waters due to fluorescent dissolved organic matter produced in situ, thereby identifying and isolating fluorescent components related to microbiological activity. Principal components were validated by a split-half analysis in the DrEEM toolbox, and a model explaining over 99.5 % of the dataset variability was fitted for the analysed samples. PARAFAC was performed using MathWorks MATLAB 2017b, as detailed in a prior report (Miranda et al., 2020). Intercomparison of measured EEMs was completed in OpenFluor (Murphy et al., 2014).

Laboratory-based relative hyperspectral reflectance measurements of the undiluted sample were conducted on 12 August 2021 using a Spectral Evolution (SEV) SR-3501 spectroradiometer fitted with a lens having an 8∘ field of view and the collected data interpolated to a 1 nm resolution from the ultraviolet (UV, 280 nm) to short-wave infrared (SWIR, 2500 nm) spectrum. Observations with the SEV were performed in reflectance mode; this involved white referencing with a 20 cm × 20 cm SphereOptics Zenith Polymer^® SG3120 ≈ 99 % full-material PTFE diffuse standard to determine the relative reflectance of bloom water sample. The derived relative reflectance was automatically divided by the calibration values of the white diffuse standard supplied by the manufacturer. An ARRILITE Plus 575 W high-performance halogen lamp placed at a viewing angle of 45∘ and at a height of 30 cm from the sample was used as a light source in a dark calibration laboratory. No additional background corrections were done, as the sample was very dense and optically deep (Fig. 1). Furthermore, the sample had been placed in a dark container with negligible reflectance. Observations were done 5 cm above the sample at nadir, resulting in a target pixel with a diameter of 0.7 cm, and each measurement was an average of 30 scans. Additional analyses on these measurements were conducted in MathWorks MATLAB R2020b, including computing the derivative, spectral angle mapping and absorption feature identification, as proposed in previous work (Liutkus, 2015; Garaba and Dierssen, 2020; Garaba et al., 2021).

Radiance measurements that were calibrated in situ were also completed aboard a small electric-motor-powered boat on 25 August 2022 on Lake Bante. The SEV was operated in radiance mode with each observation set to be an average of 20 scans, and pseudo-replicate measurements were collected from stations 1 to 11 (Fig. 1a). Sampling at each station was performed in a series of three radiance measurements over the targets: (i) diffuse white panel = Ed, (ii) water surface =Lsfc and (iii) sky =Lsky at a ∼ 45∘ viewing angle from nadir. The diffuse white panel was the same 20 cm × 20 cm SphereOptics Zenith Polymer^® SG3120 ≈ 99 % full-material PTFE diffuse standard. Efforts were made to maintain a 90–135∘ azimuthal angle from the sensor heading to the Sun to mitigate specular reflection (Garaba and Zielinski, 2013). A set of at least four to six pseudo-replicate observations were achieved at the survey stations, as this was dependent on the drift and rotation of the boat. The rotation of the boat sometimes resulted in non-optimal viewing geometry that meant the presence of surface-reflected glint. Spectral reflectance (R) was derived by assuming a flat sea surface with a ρ=0.021, and the manufacturer calibration reflectance of the diffuse white panel was not applied for brevity. 1R=Lsfc-ρ⋅LskyEd

The surface-reflected glint as determined from visual inspection was believed to be minimal, as the lake was relatively calm during the field campaign. However, to allow future users of the radiance observation to apply a surface-reflected glint correction of choice, the quality-controlled, calibrated radiometric quantities required (Eq. 1) were made openly available (Garaba and Albinus, 2023).

The absorption coefficient spectra measured with the PSICAM for sampling done in August 2021 and 2022 (Fig. 3) showed shapes typical of water samples containing phytoplankton. In fact, it could be seen that the total absorption coefficient spectrum of the water constituents measured was dominated by the phytoplankton pigments, as their absorption peaks were clearly visible and the typical exponential increase towards the shorter wavelengths (Kirk, 2011) caused by coloured dissolved organic material and non-algal particles was less prominent. This is supported by the comparison of the atot and acdom measurements (Fig. 3) in 2022: it is noticeable that acdom is considerably smaller than atot in most parts of the spectrum. The values of acdom were nearly uniform in the lake; thus, spatial changes in atot were driven by particulate/phytoplankton absorption (aph). Generally, atot and aph spectra are dominated by two large peaks in the blue ∼ 440 nm and in the red ∼ 680 nm, attributable to the presence of chlorophyll a. Additional absorption peaks were revealed by fourth derivatives of the measured absorption coefficient spectra at ∼ 416, 464, 496, 630, 642 and 682 nm (Fig. 3). The peaks from the derivative analysis are related to chlorophyll a and the presence of other photosynthetic or photo-protective pigments. The spectral shapes of the collected water samples at the various sampling locations were quite similar, suggesting a general presence of N. spumigena in the various areas of Lake Bante. However, a striking difference that was observed was the presence of a peak at 574 nm in the samples considered to be non-bloom or those with a low presence of N. spumigena. Furthermore, the peaks around 630 and 642 nm were absent in the derivative analysis of some sample spectra (Fig. 3). One possible benefit of knowing about these absorption features would be the development of simplified algorithms that could be used to optically infer or detect a Nodularia bloom, for example, in Lake Bante using an automated continuous in-water absorption sensor.

In remote sensing, a common way of detecting cyanobacterial blooms involves the use of salient spectral features in the measured reflectance or absorption signal caused by the phycobilin pigments unique to the specific phytoplankton group. Four major groups of phycobilins that have different absorption maxima include phycoerythrin (490–575 nm), phycoerythrocyanin (570–595 nm), phycocyanin (615–640 nm) and allophycocyanin (620–655 nm), as proposed in literature (Seppälä et al., 2007; Sidler, 1994; Rowan, 1989). Furthermore, satellite remote sensing of cyanobacteria often utilizes the 620 nm waveband as a proxy for phycocyanin (Stefan et al., 2005; Wang et al., 2016). The Lake Bante bloom sample absorption coefficient signal did not reveal this feature, although peaks were observed in the derivative spectra at shifted neighbouring wavebands (630 and 642 nm). However, this is not necessarily contradictory to the presence of cyanobacteria, as the phycobilin content in these organisms can vary largely depending on the physiological status and environmental conditions, such as irradiation (Seppälä et al., 2007). It has been reported that the photobleaching of phycobilins tends to occur under strong-radiation conditions (Donkor and Häder, 1996). As the samples were taken from the surface, where ultraviolet and visible irradiation is strongest, we believe that this might be related to the absence of peaks in the region of phycocyanin and allophycocyanin. The possible differences in the phycobilin content of cells from various areas of the lake might also be explained by variations in the physiological status of the cells. Nodularia blooms usually develop in the upper 5 m and float with age towards the surface, where scum formation occurs (Gröndahl, 2009). Therefore, it is likely that the populations of some samples might have had some phycoerythrocyanin, whereas others did not, which could also explain the presence/absence of the observed peak at 574 nm. An alternative explanation, as the peak was absent especially in the very highly concentrated samples, could be strong masking by the proportionally higher absorption in the red part of the spectrum in the bloom sample or suggests the presence of other phytoplankton types in the non-bloom areas. The presence or absence of the peak at 574 nm could be used in a stepwise binary algorithm to eliminate or identify blooms in remote-sensing observations.

Diagnostic peaks were found in the fluorescence signals of the Lake Bante bloom sample, and these signals are characteristic and common for waters with terrestrial sources (Table 1). Three distinct excitation–emission (Ex/Em) regions were revealed in the raw fluorescence spectrum (Fig. 4).

T and B peaks located at Ex/Em = 275/325 nm indicated the possible presence of autochthonous tryptophan and methionine-like components, also associated with biological activity (Coble, 2007; Kwon et al., 2018). Anthropogenic activities around Lake Bante could contribute to the C peak found at Ex/Em =350/475 nm, also related to ultraviolet-A (UVA) humic-like compounds of allochthonous origin (Coble, 2007). The lake is surrounded by restaurants, gardens, and residential and industrial structures that might contribute to the production and release of humic-like compounds into the waterbody (Fig. 1). Pigment-like compounds with Ex/Em = 390/620 nm have been reported to be associated with protochlorophyllide species (Remelli and Santabarbara, 2018; Myśliwa-Kurdziel et al., 2003; Campbell et al., 1998). These highly fluorescent precursors are present in cyanobacteria exhibiting signatures that can be detected over an emission range of 620 to 650 nm (Böddi et al., 1998). It is also possible that cyanobacteria distributions can be assessed using phycocyanin-related peaks instead of chlorophyll-a in vivo fluorescence located in the non-fluorescing photosystem I. Furthermore, a correlation between the filamentous algal biomass and the intensity of the phycocyanin peak at Ex/Em = 620/650 nm has been proposed as an indicator in remote sensing of algal bloom events (Seppälä et al., 2007).

Example PARAFAC analyses also revealed the presence of four main components, identified as C1, C2, C3 and C4 (Fig. 5, Table 1). Interestingly, the peak located at 390/620 was missing in our analysis due to a limited sensor sensitivity in the red spectrum. In any case, PARAFAC components derived from our study correspond to well-known chemical groups, namely, tyrosine and tryptophane substances linked to C1 and C2 components as well as humic-A- and humic-C-like substances related to the C3 and C4 components (Hudson et al., 2007).

Fluorescence in our samples was dominated by peaks C1 and C2, commonly referred to as proxies for compounds derived in situ that are indicative of microbiological activity in the investigated waters. The overall composition of the fluorescence spectra shows that up to 30 % of the total fluorescence intensity can be assigned to humic-acid-like compounds ubiquitous in coastal waters (Coble, 1996; Repeta, 2015). Similarly, up to 70 % of the total fluorescence suggests materials derived in situ, namely, tyrosine- and tryptophane-like substances. Enhanced peaks have been associated with the biological reprocessing of DOM in surface waters (Khan et al., 2019). This is consistent with an algal bloom event in which photo-degraded residues of decaying organic matter, e.g. cells, exopolymers, or carbohydrates, are reutilized by microorganisms in surface waters, thereby producing high values for those indicators (Miranda et al., 2018; Lopes et al., 2020; Weiwei et al., 2019).

As expected in vegetation or algae, a red-edge feature was observed in the spectral reflectance measured from the bloom water and a peak matching the apparent dark-green (∼ 560 nm) colour of the bloom was also evident in the visible spectrum (Fig. 6). Derivative analysis showed major absorption features at 298, 437, 633, 676, 837, 986 and 1201 nm. Variability in the pseudo-replicate measurements observed could be related to the drift of the boat during field observations. For the 2 years, the laboratory-based 2021 data had the highest magnitude, reaching ∼ 0.6, whereas the in situ 2022 reflectance was ∼ 0.27 in the near-infrared wavebands. Indeed, differences in settings (e.g. density of algae, variable lighting conditions, changes in environment, and light source) in the laboratory and in situ spectral measurements could be sources of some uncertainty in the data.

We also put the measurement that we collected into context by including reported blooms caused by Ulva prolifera (Hu et al., 2017), Tricodesmium (McKinna et al., 2011) and N. spumigena (Soja-Woźniak et al., 2018). A comprehensive quantitative comparison of the spectra was not completed because of the missing metadata in the various studies; for example, some observations were conducted in a laboratory, such as Soja-Woźniak et al. (2018), whereas other observations, such as those undertaken by McKinna et al. (2011), were carried out on research vessels. However, it is evident that all of the spectra share similarities in shape, although they have differences in magnitude (Fig. 7). Therefore, any algorithm based on the measured spectral shape would be appropriate for the detection of such blooms, instead of the magnitude-dependent approach. A red edge is noticeable in all of the spectra, and there is diagnostic peak at around 550 nm. Ulva prolifera and N. spumigena had nearly the same spectral shape over the measured spectrum; however, the reflectance magnitude was alike up until ∼ 750 nm before differences were noted in the infrared band.