Hyperspectral Differentiation of Phytoplankton Taxonomic Groups: A Comparison between Using Remote Sensing Reflectance and Absorption Spectra

Abstract

The emergence of hyperspectral optical satellite sensors for ocean observation provides potential for more detailed information from aquatic ecosystems. The German hyperspectral satellite mission EnMAP (enmap.org) currently in the production phase is supported by a project to explore the capability of using EnMAP data and other future hyperspectral data from space. One task is to identify phytoplankton taxonomic groups. To fulfill this objective, on the basis of laboratory-measured absorption coefficients of phytoplankton cultures (a_ph(λ)) and corresponding simulated remote sensing reflectance spectra (R_rs(λ)), we examined the performance of spectral fourth-derivative analysis and clustering techniques to differentiate six taxonomic groups. We compared different sources of input data, namely a_ph(λ), R_rs(λ), and the absorption of water compounds obtained from inversion of the R_rs(λ)) spectra using a quasi-analytical algorithm (QAA). R_rs(λ) was tested as it can be directly obtained from hyperspectral sensors. The last one was tested as expected influences of the spectral features of pure water absorption on R_rs(λ) could be avoided after subtracting it from the inverted total absorption. Results showed that derivative analysis of measured a_ph(λ) spectra performed best with only a few misclassified cultures. Based on R_rs(λ) spectra, the accuracy of this differentiation decreased but the performance was partly restored if wavelengths of strong water absorption were excluded and chlorophyll concentrations were higher than 1 mg∙m⁻³. When based on QAA-inverted absorption spectra, the differentiation was less precise due to loss of information at longer wavelengths. This analysis showed that, compared to inverted absorption spectra from restricted inversion models, hyperspectral R_rs(λ) is potentially suitable input data for the differentiation of phytoplankton taxonomic groups in prospective EnMAP applications, though still a challenge at low algal concentrations.

1. Introduction

Optical observations of various water types based on in situ measurements and remote sensing data have provided comprehensive information on optical properties and concentrations of optically-significant constituents in aquatic systems. Most notably there have been extensive studies focusing on bio-optical algorithms for estimating the concentration of chlorophyll-a as a general proxy for phytoplankton biomass and primary production from water surface reflectance, e.g., [1,2,3]. Recently, different bio-optical and ecological algorithms have been developed for identifying and differentiating between phytoplankton functional types (PFTs) or size class (PSCs), and taxonomic composition of phytoplankton at the ocean surface, including remote sensing algorithms for monitoring and detecting harmful algal blooms, and for identifying specific phytoplankton species [4,5,6,7,8]. These methods can be summarized into four main types: (1) methods using information on chlorophyll or light absorption to distinguish between PFTs or PSCs [9,10,11]; (2) spectral response methods based on reflectance anomalies for different PFTs/PSCs (e.g., the PHYSAT approach by Alvain et al. [12,13,14]); (3) absorption-based spectral approaches by deriving a phytoplankton size factor [15,16], through look-up tables [17], by a phytoplankton size discrimination model [18], by the partial least squares regression method [19], or by Differential Optical Absorption Spectroscopy (PhytoDOAS) [20,21]; and (4) a backscatter-based method to infer particle size distribution (PSD) and PSCs [22]. Most approaches mentioned above have been tested globally and applications for using these satellite products have been started. However, validations and adaptions of these approaches to new sensors need to be carried out prior to becoming operational.

With recent advances in optical measurements, comprehensive understanding of the light field within the water, and improvements in satellite sensors, the possibility of taxonomic discrimination of phytoplankton groups has been investigated [23,24,25,26,27]. As satellite sensors expanded from multispectral to hyperspectral detection (e.g., Hyperion, HICO, and the future missions EnMAP [28], PACE, and HyspIRI), the consequently higher number of wavebands, narrower spectral bandwidths and fully-covered range of the visible light spectrum provide more comprehensive remote sensing data on spectral properties of the water reflectance. For instance, as one of the advanced hyperspectral satellite missions, EnMAP is currently in its production phase. One of the major scientific tasks to which EnMAP will contribute is the aquatic ecosystems, not only focusing on oceans but also coastal and inland waters, regarding various water applications such as improvements on quantification of water constituents and taxonomically identification of algal and phytoplankton groups [28]. The future availability of hyperspectral sensors from space provides a high potential for distinguishing phytoplankton groups by their spectral pigment absorption alone and, thereby, provide a better mapping of the phytoplankton community composition, both for global oceans and regional waters.

Several algal groups have distinct optical properties that are related to their taxonomy and size. The shape of phytoplankton light absorption spectra results from absorption of individual pigments contained in the cell, where the pigment composition is genetically fixed and the concentrations in the cell are influenced by photoacclimation [29]. Due to the genetically fixed pigment composition, taxonomically different algal groups of phytoplankton can be differentiated by analyzing spectral absorption properties. Fourth derivative transformation is typically applied to absorption spectra to enhance spectral features, and then the similarity between the fourth derivative spectra of the targeted phytoplankton and a reference spectrum of a known algal species or taxonomic group is analyzed [30]. Using this method Millie et al. successfully detected the potential harmful algae Gymnodinium breve (now named Karenia brevis) [30] and expanded this application to natural waters [31]. Based on spectral derivative analysis methods, such as principal component analysis (PCA), cluster analysis, and Discriminant Analysis, have been tested for this application [32,33,34]. Instead of a single parameter, measurements of biomass, pigment composition, and fluorescence excitation were combined with the absorption spectra. These supplementary data have also played an important role in precisely distinguishing phytoplankton groups [25,26,35,36,37,38].

Optical remote sensing typically provides water-leaving reflectance, R_rs(λ) as one measured parameter; this has been used to discriminate phytoplankton communities and to identify some single algae species. Craig et al. [4] applied two numerical methods to in situ hyperspectral measurements of R_rs(λ) to assess the feasibility of remote detection of the toxic dinoflagellate, Karenia brevis. A quasi-analytical algorithm (QAA) was used to invert R_rs(λ) to derive phytoplankton absorption a_ph(λ) [39], then the fourth derivatives of derived a_ph(λ) were compared to the fourth derivative of a reference K. brevis absorption using a similarity index analysis. Similar studies to Craig et al. [4] have been carried out subsequently to distinguish phytoplankton types or monitor algal blooms by using multispectral and hyperspectral approaches, band ratios, or empirical relationships between a R_rs ratio and typical pigment concentration of specific algae [6,23,40,41,42]. An ocean reflectance inversion model was also developed for inverting marine inherent optical properties for the use of phytoplankton community structure in order to discriminate Noctiluca miliaris and diatoms [43]. Most of the studies on algae detection were proposed to identify only a single species or to differentiate a single phytoplankton group in natural waters.

Due to the diversity of phytoplankton in the global oceanic and coastal waters, it is necessary to spectrally differentiate most common phytoplankton groups using hyperspectral data from advanced optical sensors, such as EnMAP. The aims of the current study are to assess the feasibility of using R_rs(λ) spectra to differentiate several phytoplankton taxonomic groups, and to compare the performance when using R_rs(λ) directly and absorption obtained from inversion of the reflectance as input data, so that the suitable input data can be determined. This study is proposed to be a preparation research for an application of phytoplankton group differentiation using EnMAP hyperspectral data.

2. Data and Methods

2.1. Algal Cultures

125 cultures of various algal species from six major phytoplankton taxonomic groups were prepared. These cultures included 19 diatom species (heterokontophyta (bacillariophyceae)), 13 species of dinophytes (dinophyta (dinophyceae)), four species of prymnesiophytes (haptophyta (prymnesiophyceae)), three species of cryptophytes (cryptophyta (cryptophyceae)), 23 species of chlorophytes (chlorophyta), and six species of cyanobacteria (cyanophyceae)). Additional cultures of three different taxonomic groups that were represented by just a single species were not included in this study, as genetic variability inside a single taxonomic group shall be represented by results of several species inside a group. Diatom species were isolated from water samples taken in the North Sea. Species of the other groups were provided by the Alfred-Wegener-Institute, Helmholtz-Center for Polar and Marine Research, and the Leibnitz-Institute of Freshwater Ecology and Inland Fisheries. The cultures were grown in f/2 medium [44] prepared from filtered North Sea water, in the case of marine species, and modified Waris solution [45] in case of the fresh water species. The algae were grown from single isolated cells in light culturing chambers (Rumed, Germany) at 20 °C under 24 h artificial light (day-light fluorescence tubes) of 50 and 100 µmol photons m⁻²∙s⁻¹ photosynthetically available radiation. The different light conditions were chosen to take variations in a_ph(λ) due to photoacclimation in one species into account. Therefore the original isolate was grown for a few days under 50 µmol photons m⁻²∙s⁻¹ until a sufficient cell concentration was reached to divide the culture into two 1L-flasks and then grown for another 5−10 days under the two different light intensities. After this photoacclimation period the cell concentrations were still low and the algae still in the exponential growth phase when sampled. During acclimation and until sampling a good physiological status of the cells was controlled daily by measuring the maximum quantum efficiency of photochemistry with a PhytoPAM (Walz, Germany). Cultures were used only when this efficiency was high and cells can, hence, be considered to be in a healthy state; this is typically the case when cultures are in an exponential growth phase under nutrient replete situation. In case the efficiency was too low for a specific culture, a single cell of that culture was isolated and a new culture established.

2.2. Absorption Measurements and Normalization

The absorption coefficient spectra of phytoplankton, a_ph(λ) (m⁻¹), were measured with a Point-Source Integration-Cavity Absorption Meter (PSICAM) following the procedures outlined by Röttgers et al. [46,47]. Determination of a_ph(λ) are performed in the spectral range of 350–725 nm with a 2 nm resolution, by measuring the absorption coefficient of the culture sample and subtracting the absorption coefficient of the 0.2 µm-filtrate of the same culture sample. All measurements were done at least in triplicate against pure water as the reference. The PSICAM was calibrated daily against a spectrophotometer (Lambda 800, Perkin-Elmer) using solutions of the colored dye Nigrosine. The PSICAM offers accurate and very sensitive determinations of the absorption coefficient without errors induced by light scattered on the algal cells.

The measured a_ph(λ) spectra were normalized for further utilization. According to Roesler et al. [48], absorption spectra exhibit two kinds of variance: variance in magnitude and variance in spectral shape. Magnitude variances are due to changes in the spectrally averaged absorption coefficient, A_ph (m⁻¹), which is the area of the spectral curve over the pre-defined spectral range and can be expressed as:

$$A_{ph}=\frac{{{\displaystyle \int }}_{{\lambda }_{min}}^{{\lambda }_{max}}{a}_{ph}\left(\lambda \right) \left(m^{-1}\right) d\lambda \left(nm\right)}{{\lambda }_{max}-{\lambda }_{min}\left(nm\right)}$$

(1)

where λ_max and λ_min are the integration upper and lower spectral limits. Therefore, for each phytoplankton culture, the absorption spectrum was normalized to the underlying area in the range of 400–700 nm:

$$a_{ph}^n\left(\lambda \right)=\frac{a_{ph}\left(\lambda \right)}{A_{ph}}$$

(2)

where $a_{ph}^n\left(\lambda \right)$ (dimensionless) is the area-normalized absorption curve.

2.3. HydroLight Simulations of Hyperspectral Remote Sensing Reflectance

As one of the most important apparent optical properties, remote sensing reflectance, R_rs(λ), is commonly used in bio-optical models and ocean color remote sensing for water component retrieval and biomass estimation. In this study the absorption spectra of cultures were measured in the laboratory and based on these absorption spectra, and a radiative transfer model was used to compute radiance distributions through the water column and, finally, R_rs(λ). The simulations were carried out with HydroLight 5.2 (Sequoia Scientific, Inc., Bellevue, WA, USA) [49,50]. HydroLight allows the user to provide input files that define the inherent optical properties (IOPs) used in a simulation in controlled environments and other environmental parameters,, such as ocean surface wind speed, sun and sky irradiance, sun zenith angle, and so forth.

In the present study, hyperspectral R_rs(λ) were simulated with the following setups and assumptions taken into account:

• Phytoplankton absorption: eight chlorophyll concentrations (Chl) were set for each measured $a_{ph}^n\left(\lambda \right)$varying from 0.1 to 100 mg∙m⁻³ (0.1, 0.5, 1, 3, 5, 10, 50, and 100 mg∙m⁻³). To determine a_ph(λ) for these different Chl concentrations, we used the empirical relationship by Bricaud et al. [51] to calculate a_ph(440): a_ph(440) = 0.0654 [Chl]^0.728. The modeled absorption spectra were obtained by multiplying the normalized absorption spectra with a_ph(440), i.e., $a_{ph}^{mod}\left(\lambda \right)=a_{ph}\left(440\right) a_{ph}^n\left(\lambda \right)$. In the end, eight sets of $a_{ph}^{mod}\left(\lambda \right)$ were obtained for these different Chl concentrations.
• Chromophoric dissolved organic matter (CDOM) absorption: the HydroLight default exponential model for CDOM absorption with fixed values at 440 nm, a_CDOM(440) (m⁻¹), and a single spectral slope of 0.014 nm⁻¹ was used [52]: $a_{CDOM}\left(\lambda \right)=a_{CDOM}\left(440\right)\text{exp}\left(-0.014\left(\lambda -440\right)\right)$. In the simulations two CDOM concentrations were used with a_CDOM(440) = 0.0 and 0.1 m⁻¹ to check how the varying CDOM concentrations do influence the performance of the differentiation.
• Absorption by non-algal particles, a_NAP(λ) (m⁻¹), was determined using a mass-specific absorption coefficients due to non-algal particles, a_NAP^*(λ) (m²∙mg⁻¹), and a particle mass concentration, which is often referred to as total suspended matter concentrations (TSM), i.e., $a_{NAP}\left(\lambda \right)=TSM{a}_{NAP}^{*}\left(\lambda \right)$. The spectral shape of the a_NAP^* used is very similar to the HydroLight standard (average coefficient), but is based on spectrally exponentially-shaped in situ measurements from the Baltic Sea and Elbe River (unpublished data). Two TSM values were set for different simulation scenarios (TSM = 0 and 1 g∙m⁻³).

We used standard settings for these HydroLight simulations [50], with the exception of pure seawater absorption and scattering coefficients, a_w(λ) and b_w(λ), which were calculated with the Water Optical Properties Processor (WOPP) [53], assuming a temperature of 10°C and salinity of 30 PSU. Plankton and non-algal particles are assumed to scatter like Petzold’s average particle scattering (each with a backscatter fraction of 0.0183) [50,54]. Furthermore, the ocean is assumed to be infinitely deep and optically homogeneous. Raman scattering, as well as chlorophyll and CDOM fluorescence are taken into account. Other assumptions regarding the atmosphere included: the sun is in zenith, wind speed is 5 m∙s⁻¹, a standard atmosphere with marine aerosols, and a clear sky; this results in an aerosol optical thickness at 550 nm of 0.261. R_rs(λ) spectra were simulated from 400 nm to 700 nm with 2.5 nm spectral resolution, therefore 4000 R_rs(λ) spectra were finally obtained for the above scenarios including the different Chl and TSM concentrations and CDOM absorption coefficients. As done above for a_ph(λ), area-normalization was applied to each simulated R_rs(λ) spectrum in the range of 400–700 nm to obtain as set of $R_{rs}^n\left(\lambda \right)$ for each $a_{ph}^n\left(\lambda \right)$spectrum.

2.4. Inversion of Absorption Spectra from Simulated R_rs(λ)

As R_rs(λ) can be obtained directly from satellite sensors, potential applications in phytoplankton groups differentiation using satellite data will rely, firstly, on reflectance data. However there can be two ways to utilize R_rs(λ) data. The simple one is to apply the differentiation approaches directly to the normalized R_rs(λ); the other one is to invert absorption spectra from R_rs(λ) data using bio-optical models and then utilize the inverted absorption spectra for differentiation. In the present study, absorption spectra were inverted from the simulated hyperspectral R_rs(λ) using the quasi-analytical algorithm (QAA) version 5 as described in Lee et al. [39,55,56] for optically-deep waters. Prior to choosing the QAA, other reflectance inversion models had been tested (e.g., semi-analytical algorithms in GIOP model [57]), but showed significant discrepancies between the inverted absorption spectra and measured ones, as in these algorithms the a_ph(λ) are modeled by derived specific absorption coefficients of phytoplankton or by empirical equations using absorption coefficients at a reference wavelength, which deteriorates the spectra features of pigment composition in the full spectral region. The QAA is simple and quick to apply, as its calculation efficiency is similar to that of empirical models, but its accuracy has been shown to be similar to that of optimization methods [39]. Both absorption coefficient spectra, NAP and CDOM, are characterized by an absorption exponentially decreasing with wavelength without pronounced maxima or minima, thus their absorption will have very little influence on the spectral shape of the fourth derivative spectrum [19]. Compared to that, the inversions using bio-optical models for a_ph(λ), a_CDOM(λ) and a_NAP(λ) usually include uncertainties and errors due to a series of assumptions and empirical relationships between the absorption coefficient and wavelength. The total absorption coefficient is typically a precisely retrieved parameter from QAA inversion [39] and as pure water absorption is relatively accurately known, subtraction of the pure water absorption from total absorption might reduce its deteriorating effect on the spectral absorption features, though uncertainties introduced by the inversion will somehow influence the quality of inverted absorption data. Therefore, we only inverted the non-water absorption, a_pg(λ) = a_ph(λ) + a_CDOM(λ) + a_NAP(λ), instead of a_ph(λ), when using QAA. A detailed description of the mathematical steps involved in the QAA inversion process can be found in [56]. The QAA-inverted a_pg(λ) spectra were also area-normalized to obtain $a_{pg}^n\left(\lambda \right)$, as done for the lab-measured a_ph(λ) and HydroLight-simulated R_rs(λ).

2.5. Derivative Analysis

Derivative spectroscopy has been widely used in the analysis of hyperspectral data using various computation algorithms [58,59,60]. It can be applied to hyperspectral measurements of both inherent and apparent optical properties (e.g., absorption spectra, remote sensing reflectance). Derivative analysis enhances spectral features and, thus, better distinguishes subtle features in the spectra. In the present study, a finite divided difference algorithm was used to estimate the derivative spectra by taking the difference of a given spectrum over a sampling interval (Δλ), defined as Δλ = λ_j − λ_i, where λ_j > λ_i. The first and the nth derivative are obtained using equations of ${\frac{ds}{d\lambda }|}_i\approx \frac{s\left({\lambda }_j\right)-s\left({\lambda }_i\right)}{\text{Δ}\lambda }$ and $\frac{d^{n}s}{d{\lambda }^n}=\frac{d}{d\lambda }(\frac{d^{(n-1)}s}{\text{d}{\lambda }^{(n-1)}})$, respectively, where s is the spectrum used for the derivative transformation [59]. It is noteworthy that the use of the second derivative and fourth derivative transformation yield different meanings: the second derivative provides qualitative identification of pigments only, whereas the fourth derivative provides quantitative identification [61]. Therefore, the fourth derivative spectra of the absorption are often computed to resolve the positions of the absorption maxima attributable to photosynthetic pigments [30]. To be consistent, the fourth derivative transformation was applied to $a_{ph}^n\left(\lambda \right)$, $R_{rs}^n\left(\lambda \right)$ and QAA-inverted $a_{pg}^n\left(\lambda \right)$ spectra in this study. However, as the derivative computation increases noise in the spectrum, smoothing has to be applied to the data [36,58]. The Savitzky-Golay filter was used for smoothing the original data in which a polynomial order of four and a frame size of 21 were appropriately selected after multiple attempts to determine the best compromise between the noise removal and the ability to resolve the fine spectral information. The Savitzky-Golay filter was selected based on the advantage that the filter exhibits excellent properties of the distribution such as relative maxima and minima. Other smoothing filters by conventional methods, such as a moving average usually distort some spectral features by flattening or shifting [62].

2.6. Similarity Index (SI) and Clustering Analysis

All derivative spectra, those of $a_{ph}^n\left(\lambda \right)$, simulated $R_{rs}^n\left(\lambda \right)$, and of QAA-inverted $a_{pg}^n\left(\lambda \right)$, were compared between the 125 input spectra using a similarity index (SI) analysis as described by Millie et al. [30]. In the present study the cosine distance was considered as SI and was computed from the angle between two vectors such that $\text{SI}=\frac{A_1·A_2}{\left|A_1\right|\times \left|A_2\right|}$, where A₁ and A₂ are vectors that comprise the two derivative spectra. The SI calculation yields a number from 0 to 1, where 0 indicates no similarity, while 1 indicates absolute similarity between the two spectra.

SI analysis is adequate to differentiate two spectra that may represent, or not, two optically-different spectra/groups, but a single SI number alone is not sufficient for the differentiation of several phytoplankton groups. Therefore, hierarchical cluster analysis (HCA) was used to create a hierarchical cluster tree and to partition the dataset into clusters using a single linkage (nearest neighborhood) algorithm. The linkage algorithm is based on the cosine distance (i.e., SI) between derivative spectra. HCA traditionally displays a dendrogram to represent the hierarchical tree, with individual observations at one end and the clusters to which the data belong at the other. According to the SI defined here, the closer to one the SI is, the more similar are the features of the two compared spectra. Therefore, spectra with a similar phytoplankton composition are expected to appear closer with larger SI in the cluster tree than those having a very different composition [25]. Cluster trees in the current study were generated by using a free software for scientific data analysis PAST version 2.17 [63].

3. Results

3.1. Derivative Analysis and Clustering of Algal Absorption Spectra

Figure 1a shows some representative $a_{ph}^n\left(\lambda \right)$spectra of cultures from the six investigated taxonomic algal groups. These $a_{ph}^n\left(\lambda \right)$ spectra showed considerable variability in spectral shape from one group to another, indicating significant differences in absorption spectral features. Figure 1b shows fourth derivative transformed spectra of $a_{ph}^n\left(\lambda \right)$, indicating positions of the absorption maxima attributable to single photosynthetic pigments [35]. The HCA cluster analysis was performed on these fourth derivative spectra of $a_{ph}^n\left(\lambda \right)$. Note that the result of this clustering is sensitive to the selection of the spectral range, as specific absorption imprints caused by accessory pigments usually occur in a narrower range than 400–700 nm. A sensitivity analysis of spectral regions was performed in detail by Torrecilla et al. [25] and here we used a similar procedure to determine the optimal spectral range before performing cluster analysis. The optimal range was determined by choosing low SI between each pair of groups based on the SI calculated for all possible varying ranges from 400 to 700 nm (Figure 2 shows the SI variation between heterokontophyta and cryptophyta as an example). There were, in total, 15 optimal ranges generated for all pairs of groups from the six phytoplankton groups. Those spectral ranges that also show low SI values, but are too narrow (i.e., close to the 1:1 line as shown in Figure 2), were not considered because they present too little information. As we try to differentiate the six groups simultaneously, due to the fact that for each pair of groups the optimal range may vary, in the present study we combined the 15 spectral ranges to include all the important pigment information. This analysis showed that the spectral range of 430–660 nm gave the best information of all accessory pigments and, thus, was used as the optimal spectral range. Clustering results showed that the fourth derivative spectra of measured $a_{ph}^n\left(\lambda \right)$can be used to differentiate the six phytoplankton groups quite precisely. All species of haptophyta, chlorophyta, cryptophyta, and cyanobacteria were well grouped together. For the heterokontophyta, 27 out of 30 cultures, and for the dinophyta, 20 of 21 cultures, were classified together (Figure 3 and Table 1). Using a SI of 0.90, cryptophyta could be differentiated; while the SI of 0.95 and 0.96 were the thresholds for cyanobacteria and chlorophyta, respectively. The highest SI was found between heterokontophyta and dinophyta (0.98) with a few misclassified cultures in between the two groups.

Figure 1. Examples of (a) normalized absorption spectra for different phytoplankton groups with (b) corresponding fourth-derivative spectra.

Figure 1. Examples of (a) normalized absorption spectra for different phytoplankton groups with (b) corresponding fourth-derivative spectra.

Figure 2. SI variation between heterokontophyta and cryptophyta for all possible varying ranges from 400 to 700 nm. The optimal spectral range containing most pigment information with low SI is between λ_min ≈ 430 nm and λ_max ≈ 660 nm.

Figure 2. SI variation between heterokontophyta and cryptophyta for all possible varying ranges from 400 to 700 nm. The optimal spectral range containing most pigment information with low SI is between λ_min ≈ 430 nm and λ_max ≈ 660 nm.

Figure 3. Cluster tree of the six phytoplankton groups generated by using fourth-derivative spectra of measured phytoplankton absorption a_ph(λ).

Figure 3. Cluster tree of the six phytoplankton groups generated by using fourth-derivative spectra of measured phytoplankton absorption a_ph(λ).

Table 1. Clustering accuracy using measured a_ph(λ), HydroLight-simulated R_rs(λ), and QAA-inverted a_pg(λ) (a_{pg_QAA}(λ)) spectra for different water types.

	Heterokontophyta	Dinophyta	Haptophyta	Cryptophyta	Chlorophyta	Cyanobacteria
Measured aph(λ)	27/30	20/21	7/7	5/5	51/51	11/11
Simulated Rrs(λ) (Chl = 0.1 mg∙m−3)	Mixed			4/5	Mixed	8/11
Simulated Rrs(λ) (Chl = 0.5 mg∙m−3)	Mixed			4/5	Mixed	8/11
Simulated Rrs(λ) (Chl = 1 mg∙m−3)	Mixed			5/5	51/51	11/11
Simulated Rrs(λ) (Chl = 5 mg∙m−3)	Mixed			5/5	51/51	11/11
Simulated Rrs(λ) (Chl = 10 mg∙m−3)	Mixed			5/5	51/51	11/11
Simulated Rrs(λ) (Chl = 50 mg∙m−3)	Mixed			5/5	51/51	11/11
Simulated Rrs(λ) (Chl = 1 mg∙m−3, CDOM = 0.1 m−1, TSM = 1 g∙m−3)	Mixed			5/5	51/51	11/11
apg_QAA(λ) (Chl = 1 mg∙m−3)	Mixed			5/5	49/51	11/11
apg_QAA(λ) (Chl = 1 mg∙m−3, CDOM = 0.1 m−1, TSM = 1 g∙m−3)	Mixed			5/5	49/51	11/11
apg_QAA(λ) (Chl = 50 mg∙m−3)	29/30	18/21	Undistinguishable	5/5	51/51	11/11

Numbers “m/n” indicates that m spectra are grouped together from total n spectra.

Table 1. Clustering accuracy using measured a_ph(λ), HydroLight-simulated R_rs(λ), and QAA-inverted a_pg(λ) (a_{pg_QAA}(λ)) spectra for different water types.

	Heterokontophyta	Dinophyta	Haptophyta	Cryptophyta	Chlorophyta	Cyanobacteria
Measured aph(λ)	27/30	20/21	7/7	5/5	51/51	11/11
Simulated Rrs(λ) (Chl = 0.1 mg∙m−3)	Mixed			4/5	Mixed	8/11
Simulated Rrs(λ) (Chl = 0.5 mg∙m−3)	Mixed			4/5	Mixed	8/11
Simulated Rrs(λ) (Chl = 1 mg∙m−3)	Mixed			5/5	51/51	11/11
Simulated Rrs(λ) (Chl = 5 mg∙m−3)	Mixed			5/5	51/51	11/11
Simulated Rrs(λ) (Chl = 10 mg∙m−3)	Mixed			5/5	51/51	11/11
Simulated Rrs(λ) (Chl = 50 mg∙m−3)	Mixed			5/5	51/51	11/11
Simulated Rrs(λ) (Chl = 1 mg∙m−3, CDOM = 0.1 m−1, TSM = 1 g∙m−3)	Mixed			5/5	51/51	11/11
apg_QAA(λ) (Chl = 1 mg∙m−3)	Mixed			5/5	49/51	11/11
apg_QAA(λ) (Chl = 1 mg∙m−3, CDOM = 0.1 m−1, TSM = 1 g∙m−3)	Mixed			5/5	49/51	11/11
apg_QAA(λ) (Chl = 50 mg∙m−3)	29/30	18/21	Undistinguishable	5/5	51/51	11/11

Numbers “m/n” indicates that m spectra are grouped together from total n spectra.

3.2. Derivative Analysis and Clustering on HydroLight-Simulated R_rs(λ)

Figure 4 shows examples of HydroLight-simulated $R_{rs}^n\left(\lambda \right)$ that are based on individual absorption spectra of the cultures for four Chl concentrations varying from 0.1 to 50 mg∙m⁻³ and their corresponding fourth-derivative spectra. With increasing Chl concentration, the spectral shape of the $R_{rs}^n\left(\lambda \right)$ and derivative spectra were more distinct between the different taxonomic groups. It is obvious that at low Chl concentrations, the R_rs(λ) were dominated by water absorption features and little variations could be found between $R_{rs}^n\left(\lambda \right)$ based on absorption spectra of the different cultures. Figure 4 supports this statement and shows that when Chl concentration is low (0.1 mg∙m⁻³), little difference between cultures of different taxonomic groups can be seen either in R_rs(λ) or their derivative spectra, especially in the red to NIR region where absorption by water dominates. The differences between taxonomic groups became more distinct with increasing Chl concentration. In order to understand to what extent the Chl concentration did influence the differentiation, SI values between different groups were calculated using derivative spectra in the range of 430–620 nm only (to reduce the direct influence of water absorption features). Figure 5 shows an example for the SI between a single spectrum of each group varying with Chl concentration. The SI decreases dramatically with increasing Chl concentration between most of the groups, except for the SI between heterokontophyta, dinophyta, and haptophyta, which stayed relatively high compared to the SI between other groups. Figure 4 and Figure 5 indicated that phytoplankton groups differentiation using hyperspectral R_rs(λ) might not be feasible if Chl concentrations are lower than 1 mg∙m⁻³ due to high similarity in the derivative spectra.

Figure 4. Examples of HydroLight-simulated $R_{rs}^n\left(\lambda \right)$ with different Chl values (left panel) and the corresponding fourth-derivative spectra (right panel). Note that the y-axis of the derivative spectra is reversely displayed.

Figure 4. Examples of HydroLight-simulated $R_{rs}^n\left(\lambda \right)$ with different Chl values (left panel) and the corresponding fourth-derivative spectra (right panel). Note that the y-axis of the derivative spectra is reversely displayed.

Figure 5. Similarity index between the six taxonomic groups varying with Chl concentrations. Note that SI was calculated using the fourth derivative spectra of $R_{rs}^n\left(\lambda \right)$ within 430–620 nm and between single representative spectra of each group.

Figure 5. Similarity index between the six taxonomic groups varying with Chl concentrations. Note that SI was calculated using the fourth derivative spectra of $R_{rs}^n\left(\lambda \right)$ within 430–620 nm and between single representative spectra of each group.

To further assess the performance of remote sensing reflectance in differentiating phytoplankton groups, HydroLight-simulated $R_{rs}^n\left(\lambda \right)$ with a Chl concentration of 0.1, 0.5, 1, 5, 10, and 50 mg∙m⁻³ were used and the influence by water absorption features at longer wavelengths was reduced by using the fourth derivative spectra in the range of 430–620 nm only. The simulated R_rs(λ) dataset with Chl = 1 mg∙m⁻³, CDOM = 0.1 m⁻¹, and TSM = 1 g∙m⁻³ were also tested to assess whether CDOM and TSM concentrations influence the differentiation. Cluster trees for different water types are displayed in Figure 6 (cluster trees for water types with Chl of 0.5 and 10 mg∙m⁻³ were not shown but the accuracy data were listed in Table 1). Results showed that R_rs(λ) spectra for Chl of 0.1 and 0.5 mg∙m⁻³ cannot to be used to efficiently differentiate the different taxonomic groups; only cyanobacteria and cryptophyta were distinct (Table 1 and Figure 6a). When Chl = 1 mg∙m⁻³ (Figure 6b), the most distinct phytoplankton group are the cryptophytes with the SI of 0.80; the cyanobacteria cultures were grouped together at SI = 0.90, but showed larger variation in SI between species in the group due to the above-described variation in spectral absorption. Chlorophytes were well grouped (SI = 0.94), again indicating similar pigment composition among species as presented in the performance of measured a_ph(λ). When Chl is higher than 1 mg∙m⁻³, similar clusters were found but with lower SI thresholds (Figure 6c,d). More subtle differences were shown for cyanobacteria cultures, which were separated into two subgroups. When including CDOM and TSM in the derivation of R_rs(λ), the differentiation performance was not visibly deteriorated (Table 1 and Figure 6e). The overall results showed that four main groups can be differentiated using simulated R_rs(λ) when Chl ≥ 1 mg∙m⁻³, as all species of chlorophyta, cryptophyta, and cyanobacteria were classified in a single cluster, whereas heterokontophyta, dinophyta, and haptophyta were hardly distinguishable (Table 1 and Figure 6).

Figure 6. Cluster trees of the six phytoplankton groups generated by using fourth-derivative spectra of HydroLight-simulated R_rs(λ) for water types with (a) Chl = 0.1 mg∙m⁻³, (b) Chl = 1 mg∙m⁻³, (c) Chl = 5 mg∙m⁻³, (d) Chl = 50 mg∙m⁻³, and (e) Chl = 1 mg∙m⁻³, CDOM = 0.1 m⁻¹, and TSM = 1 g∙m⁻³. Color-code legend denotes different phytoplankton taxonomic groups, as shown in (b).

Figure 6. Cluster trees of the six phytoplankton groups generated by using fourth-derivative spectra of HydroLight-simulated R_rs(λ) for water types with (a) Chl = 0.1 mg∙m⁻³, (b) Chl = 1 mg∙m⁻³, (c) Chl = 5 mg∙m⁻³, (d) Chl = 50 mg∙m⁻³, and (e) Chl = 1 mg∙m⁻³, CDOM = 0.1 m⁻¹, and TSM = 1 g∙m⁻³. Color-code legend denotes different phytoplankton taxonomic groups, as shown in (b).

3.3. Phytoplankton Group Differentiation Using Absorption Inverted from R_rs(λ)

R_rs(λ) showed expected weaker differentiation performance compared to that using original $a_{ph}^n\left(\lambda \right)$spectra. This is due to the introduction of additional uncertainties during the HydroLight simulation to R_rs(λ) and due to absorption features by water itself. In this section, the QAA-inverted non-water absorption spectra a_pg(λ) were analyzed and used to differentiate phytoplankton groups. Its performance was compared with that using R_rs(λ) data in Section 3.2 to determine the suitable input data for phytoplankton groups differentiation.

Figure 7. Examples of QAA-inverted absorption spectra of $a_{pg}^n\left(\lambda \right)$and the corresponding fourth derivative spectra of the six taxonomic groups for (a,b) water type I with only Chl = 1 mg∙m⁻³, (c,d) water type II with Chl = 1 mg∙m⁻³, CDOM = 0.1 m⁻¹, and TSM = 1 g∙m⁻³ for 400–560 nm, and (e,f) water type III with Chl = 50 mg∙m⁻³ for 400–700 nm.

Figure 7. Examples of QAA-inverted absorption spectra of $a_{pg}^n\left(\lambda \right)$and the corresponding fourth derivative spectra of the six taxonomic groups for (a,b) water type I with only Chl = 1 mg∙m⁻³, (c,d) water type II with Chl = 1 mg∙m⁻³, CDOM = 0.1 m⁻¹, and TSM = 1 g∙m⁻³ for 400–560 nm, and (e,f) water type III with Chl = 50 mg∙m⁻³ for 400–700 nm.

In order to verify if the absorption by non-algal particles and CDOM has an influence on the derivative analysis, two different water types were considered to retrieve a_pg(λ): (I) water with Chl = 1 mg∙m⁻³, CDOM and TSM were set to zero, and (II) water with Chl = 1 mg∙m⁻³, CDOM = 0.1 m⁻¹, and TSM = 1 g∙m⁻³. Note that for water type I and II only data for 400–560 nm were selected because, beyond this range, there is insufficient information to reliably invert a_pg(λ) from R_rs(λ) due to errors induced by the strong absorption by water [55]. A water type III was also considered with extremely high Chl = 50 mg∙m⁻³, based on the assumption that high absorption by phytoplankton might reduce the QAA inversion errors at longer wavelengths. Results showed that the inverted a_pg(λ) at longer wavelengths were successfully retrieved for extremely high Chl waters (Figure 7e). To assess the performance of the QAA, the measured a_pg(λ) were compared with the QAA-inverted a_pg(λ) for the three water types at some discrete wavelengths [55]. Despite of the slight underestimation of the a_pg(λ) values (slope < 1 in Table 2), the QAA algorithm could give satisfactory retrievals of absorption coefficients from R_rs(λ) in the considered spectral range (Table 2).

Table 2. Summary of the QAA performance for different water types. A linear regression without the interception term was used between inverted and measured a_pg(λ), i.e., QAA-inverted a_pg(λ) = Slope * (measured a_pg(λ)), at four selected bands (410, 440, 490, and 510 nm). The slope, determination coefficient (R²), root-mean-square error (RMSE), and number of points are shown.

	Slope	R2	RMSE (m−1)	N
Water type I (Chl = 1 mg∙m−3)	0.942	0.961	0.0086	500
Water type II (Chl = 1 mg∙m−3, CDOM = 0.1 m−1, TSM = 1 g∙m−3)	0.938	0.948	0.0245	500
Water type III (Chl = 50 mg∙m−3, an extreme case)	0.953	0.975	0.117	500

Table 2. Summary of the QAA performance for different water types. A linear regression without the interception term was used between inverted and measured a_pg(λ), i.e., QAA-inverted a_pg(λ) = Slope * (measured a_pg(λ)), at four selected bands (410, 440, 490, and 510 nm). The slope, determination coefficient (R²), root-mean-square error (RMSE), and number of points are shown.

	Slope	R2	RMSE (m−1)	N
Water type I (Chl = 1 mg∙m−3)	0.942	0.961	0.0086	500
Water type II (Chl = 1 mg∙m−3, CDOM = 0.1 m−1, TSM = 1 g∙m−3)	0.938	0.948	0.0245	500
Water type III (Chl = 50 mg∙m−3, an extreme case)	0.953	0.975	0.117	500

Similar to the measured algal absorption spectra, the inverted a_pg(λ) were normalized and then fourth-derivative-transformed for the cluster analysis. Figure 7a,b shows the normalized a_pg(λ) ($a_{pg}^n\left(\lambda \right)$) and the corresponding fourth derivative spectra for water type I and Figure 7c,d shows the same for water type II. In fact, the inverted non-water absorption spectra a_pg(λ) for water type I correspond to a_ph(λ) as no absorption by CDOM and non-algal particles was included in the simulation, Nevertheless, the fourth derivative spectra of $a_{pg}^n\left(\lambda \right)$for both water types showed little differences in the spectral shape (Figure 7b,d), though their magnitude was different because the inclusion of CDOM and TSM for water type II would change the derivative values due to the exponential slopes of CDOM and non-algal particle absorption spectra. As a_pg(λ) was reasonably inverted at longer wavelengths only for water type III, the normalization and fourth derivative transformation were done within the entire spectral range as shown in Figure 7e,f. Positions of maxima and minima in Figure 7f were in good agreement with that in Figure 7b–d at 420–550 nm and were comparable to that in Figure 1b at 560–660 nm. Cluster analysis was applied to these QAA-inverted a_pg(λ) data set for water type I and II and a similar clustering was obtained (Figure 8a). The clustering results showed that four taxonomic groups can be distinguished. All cultures of the cyanobacteria and all of the cryptophyta were clustered together; 49 of 51 chlorophytes were correctly grouped, however the three other groups (heterokontophyta, dinophyta, and haptophyta,) were mixed within each others’ clusters (Table 1 and Figure 8a). Nevertheless, a better clustering was shown for water type III with extremely high Chl using the same range of 430–620 nm as done on simulated R_rs(λ) spectra (Figure 8b). In this case, all cultures of the chlorophyta, cryptophyta, and cyanobacteria were clustered together; only very few cultures of heterokontophyta and dinophyta were misclassified, but haptophyta were undistinguishable.

Figure 8. Cluster trees of the six phytoplankton groups generated by using (a) fourth derivative spectra of QAA-inverted non-water absorption a_pg(λ) from simulated R_rs(λ) for water type I and II, and (b) fourth derivative spectra of QAA-inverted non-water absorption a_pg(λ) from simulated R_rs(λ) for water type III. Note that the spectral range used in (a) is 430–550 nm, and in (b) is 430–620 nm.

Figure 8. Cluster trees of the six phytoplankton groups generated by using (a) fourth derivative spectra of QAA-inverted non-water absorption a_pg(λ) from simulated R_rs(λ) for water type I and II, and (b) fourth derivative spectra of QAA-inverted non-water absorption a_pg(λ) from simulated R_rs(λ) for water type III. Note that the spectral range used in (a) is 430–550 nm, and in (b) is 430–620 nm.

4. Discussion

4.1. HydroLight Simulations

Part of the scope of this work is to assess how phytoplankton groups are reflected from the hyperspectral remote sensing perspective. As R_rs(λ) were simulated by HydroLight in this study, two points regarding the HydroLight simulations are discussed in the following. Firstly, as for the scattering properties, besides the fact that scattering properties of hydrosols underlie particular variability and uncertainties [64], some assumptions have been made in the simulations to simplify and unify scattering properties, thus, to strengthen the sensitivity of absorption effects. The volume scattering function, VSF, exhibits spectral variations and shape changes for different algal species [65]. Depending on the nature and concentrations of planktonic particles in sea water, particle backscatter fractions in the ocean can vary from a fraction of a percent to several percent. If this would be implemented more accurately into the Hydrolight simulations, this would affect the R_rs(λ) spectra, e.g., [66,67]. However, VSF, particle backscattering, and total scattering of the culture samples have not been measured. Thus, we have no information on the actual particle backscatter fraction b_bp/b_p, which is necessary to define the appropriate Fournier-Forand phase function. Various formulas for the backscatter fraction as function of Chl can be found in the literature, e.g., [68]. However, it must be remembered that the backscatter fraction correlates poorly with Chl, and there can be order-of-magnitude variability in the measured value of b_bp/b_p for a given Chl. Mobley et al. [66] show effects of phase functions on simulation data. They state that the use of a phase function with the correct backscatter fraction could reduce RMS percentage errors in the predicted upwelling irradiance and R_rs(λ) by roughly an order of magnitude. However, the exact shape of the phase function in backscatter directions does not greatly affect the light field, so long as the overall shape of the phase function does not deviate greatly from the correct shape, and this is provided by using the Petzold average-particle phase function with b_bp/b_p = 0.0183.

Secondly, with regards to CDOM and non-algal particles absorption, two simple cases were considered, namely, one case without any CDOM and non-algal particles absorption and one with a_CDOM(440) = 0.1 m⁻¹ and TSM = 1 g∙m⁻³. The CDOM concentration would correspond to roughly 10 mg∙m⁻³ chlorophyll concentration in the Case-1 parameterization by Morel [69]. According to our routinely-obtained in situ dataset mainly from the North Sea, a very high variability of CDOM versus Chl was found with a_CDOM(440) = 0.1 m⁻¹ corresponds to Chl between 0.1 and 10 mg∙m⁻³ (two orders of magnitude). In the open sea, CDOM is mainly a product of phytoplankton degradation. In an extreme and fresh algal bloom event with concentrations of more than 10 mg∙m⁻³, the used CDOM absorption is probably not unrealistic. Furthermore, CDOM and TSM were included in this study to test how they would influence the differentiation performance of the simulated R_rs(λ), and our results showed that no significant influence was found in results of the cluster analysis (Table 1 and Figure 6).

4.2. Phytoplankton Groups Differentiation Using Absorption and R_rs(λ) Data—Performance Comparison

The fourth derivative analysis uncovers more distinct the absorption of pigments maxima within the overall absorption spectra [30]. By using the similarity index (SI) with the hierarchical cluster analysis (HCA) it was possible to effectively characterize all absorption spectra and, thus, to allow the detection of differences among phytoplankton taxonomic groups. In the present study, cluster analysis on the fourth derivative of a_ph(λ) spectra efficiently separated the 125 algal absorption spectra into distinct groups. As expected, the fourth derivative spectra from species of heterokontophyta (diatoms) and dinophyta were highly similar with SI > 0.975 and few cultures of the two groups were misclassified, due to the known similar pigment composition and absorption spectra of these two groups. Most chlorophyte species showed identical spectral features within the group and, thus, were well-clustered together. The five cryptophyte cultures showed distinct spectral features compared to other groups. The six cyanobacteria species (11 cultures) were spectrally rather different from each other, and included green, blue-green, and red colored cultures. For instance, some cultures showed distinct spectral differences in their absorption, especially at 500–600 nm, induced by several phycobilin pigments (Figure 1). Nevertheless, the current study only focused on the differentiation of phytoplankton taxonomic groups but not yet intended to investigate details at the species level. With significant optical differences to the other groups, cyanobacteria are probably easily identifiable using hyperspectral data ([7], and references therein).

For the perspective of applications using data from hyperspectral sensors, HydroLight-simulated R_rs(λ) for different Chl concentrations were tested to differentiate phytoplankton taxonomic groups with this approach. Results revealed that the SI for the fourth derivative of simulated R_rs(λ) spectra varies largely, as it is highly influenced by the Chl concentrations (Figure 5) used in the simulations, leading to an increase of uncertainty in detecting the different taxonomic groups. From the SI variation in Figure 5, a threshold of Chl = 1 mg∙m⁻³ was primarily determined to more efficiently differentiate the groups when based on R_rs(λ). The main reason for this less efficient differentiation at low Chl concentrations is the dominating influence of spectral features of pure water absorption. Further verification of this threshold was done by using simulated R_rs(λ) for water types with discrete Chl concentrations. Results indicated that four main groups can be differentiated only when Chl ≥1 mg∙m⁻³ (Table 1). This is also consistent with the SI variation shown in Figure 5, meaning that the derivative analysis and clustering approach using R_rs(λ) is promising in waters with Chl ≥ 1 mg∙m⁻³ for differentiating chlorophyta, cryptophyta, cyanobacteria, and heterokontophyta/dinophyta/haptophyta when they are dominating. So far this finding has been noted for the first time by the present study for differentiating multiple phytoplankton taxonomic groups simultaneously; no similar results have been reported in the literature and, thus, cannot be compared. It still remains difficult to distinguish phytoplankton groups with similar optical signatures (e.g., heterokontophyta and dinophyta) purely depending on reflectance spectra. Approaches to discriminate these two groups during bloom events have been developed using combined data sets of R_rs(λ), chlorophyll anomaly, absorption, and backscattering spectra from both in situ measurements and space, e.g., [42,70,71,72]. It remains unclear, however, if these techniques are effective in waters beyond their study regions or in non-bloom waters. As an important water components in natural waters, CDOM and TSM were also considered in our simulated R_rs(λ). A test of R_rs(λ) for water type with Chl = 1 mg∙m⁻³, CDOM = 0.1 m⁻¹, and TSM = 1 g∙m⁻³ showed similar differentiation results with that for water types without CDOM and TSM (Figure 6e), indicating that as expected CDOM and TSM have an insignificant influence on the differentiation performance also when using simulated R_rs(λ). The exponential spectral shapes of CDOM and non-algal particles absorption do not influence the derivative analysis.

It is noteworthy that the difficulty in differentiating phytoplankton groups using R_rs(λ) also has to do with the fundamental difference between the inherent optical properties (IOPs) and apparent optical properties (AOPs), where the latter depend on the ambient light field. According to the discussions above, differentiation of phytoplankton taxonomic groups based on simulated R_rs(λ) spectra is less effective as that based on a_ph(λ) spectra. As far as hyperspectral sensors are concerned, one can obtain absorption spectra from R_rs(λ) by using bio-optical inversion models. However, the differentiation performance of using inverted absorption spectra is much less precise due to uncertainties and errors introduced by the inversion models. The effects seemed to be larger than those induced by the pure water absorption spectral features in R_rs(λ). Though the QAA-inverted a_pg(λ) were often in good agreement with the measured ones (Table 2), the QAA showed limitations resulting in the lacking absorption spectra at longer wavelengths for waters with low Chl concentrations due to the influence of the absorption by water itself [4,55]. Compared to simulated R_rs(λ) for water types I and II, the same approach applied to the QAA-inverted a_pg(λ) showed slightly poorer differentiation performance, as the QAA does not allow to use longer wavelengths, thus only giving satisfactory inversions of absorption in the range of 400–560 nm when Chl is low. This has caused the loss of important pigment information for certain taxonomic groups. For instance, from the measured $a_{ph}^n\left(\lambda \right)$ and the corresponding fourth-derivative spectra (Figure 1), it was clearly seen that the main spectral difference between cryptophyta and other groups was found at 560–600 nm. This deteriorated information at longer wavelengths by QAA resulted in a weak ability to distinguish cryptophyta from the three mixed groups (heterokontophyta, dinophyta, and haptophyta) (Table 1 and Figure 8a) and, thus, reduced the differentiation accuracy. The test of an extreme water type III (Chl = 50 mg∙m⁻³) suggested that the use of QAA-inverted a_pg(λ) partly restored the differentiation accuracy when Chl is high enough. However, the overall performance of using the inverted absorption spectra was restricted due to the limitations of the inversion model. A recent study in estimating the dominance of diatoms by Isada et al. [27] using the derivative spectroscopy approach also suggested that the QAA-inverted absorption spectra is less precise than in situ measured absorption. The performance of directly using in situ R_rs(λ) spectra was, however, not assessed in their study.

5. Conclusions and Outlook

In this study we tested the differentiation of phytoplankton taxonomic groups from hyperspectral data by using remote sensing reflectance R_rs(λ) directly versus using absorption spectra derived from R_rs(λ) by inversion algorithms. This was done to help future implementations of applications for hyperspectral satellite sensors like EnMAP.

When looking at direct differentiation capabilities, the fourth-derivative spectra of measured phytoplankton absorption performed more effectively than that of simulated R_rs(λ) for six major phytoplankton groups, but the QAA-inverted absorption spectra were less precise than R_rs(λ). The discrimination of phytoplankton taxonomic groups using R_rs(λ) data for a spectral analysis is limited by the strong influence of spectral features from pure water absorption. The inversion of R_rs(λ) to receive pigment absorption and remove water absorption influence is not giving better results due to errors induced by the inversion algorithm. This might change in the future with improved inversion algorithms. Therefore, the use of current hyperspectral remote sensing reflectance directly from hyperspectral sensors for phytoplankton group differentiation is suggested. Additionally, there are difficulties in very low algal concentrations and in discriminating heterokontophyta, dinophyta, and haptophyta due to their similar pigment composition.

The present study is restricted by the use of HydroLight simulations under ideal circumstances with stable CDOM and TSM concentrations. Furthermore, it is noteworthy that reflectance spectra of natural waters are far more complex than theoretical simulations, and the quality of the measured spectra are also a matter of the sensor’s spectral resolution, the radiometric calibration, and the atmospheric correction, to mention only the most prominent impacts. All of the above issues induce uncertainties and difficulties in identifying phytoplankton groups for natural waters. Future work will include the differentiation capability assessment on phytoplankton group mixtures, the utilization of R_rs(λ) both from natural waters and simulated EnMAP hyperspectral images, as well as a more elaborate investigation on the impact of varying water constituent concentrations on R_rs(λ) spectral features.