2.1.

Study Site

The study site was located in the northern coast of Estonia, the Gulf of Finland, the Baltic Sea (Fig. 1). The study area encompassed $\sim 200{\mathrm{km}}^2$ covering the coast of the mainland and two Pakri islands—Suur Pakri and Väike Pakri. The islands are separated from the mainland to the south by the shallow Kurkse Strait. The Pakri islands and surrounding marine area belong to the Natura 2000 network, which is the European network of protected areas.

Fig. 1

Location of the (a) study area in the northern coast of Estonia, the Gulf of Finland, the Baltic Sea and (b) study area imaged by the CASI hyperspectral sensor on May 31, 2016.

The water depth in the study area remains generally below 5 m, but water depth may reach near 100 m in some parts of the open sea. Higher plants, such as Potamogeton pectinatus, Zannichellia palustris, and Zostera marina, as well as charophytes dominate the soft sandy substrate found in the sheltered areas between islands and between the mainland and islands. Hard substrates (rock, boulders, stones, etc.) are found in more exposed areas dominated by either brown algae Fucus vesiculosus or filamentous green algae Cladophora glomerata communities.

2.2.

In Situ Field Sampling

Ground reference data were collected from May 30 to June 3, 2016, at the same time as the remote sensing data collection. Underwater video records for characterizing benthic habitats were collected from almost 200 field stations. The locations of the field stations were determined previously using aerial photographs from the public web map application of the Estonian Land Board. The video analysis method described in Ref. 48 was used to estimate the substrate type, the total coverage of benthic vegetation, and the coverage of visually distinguishable vegetation species. Coordinates of each field station were recorded by the global positioning system. Bathymetry was measured with the boat echo sounder.

Above water reflectance measurements were acquired using two Ramses (TriOS) hyperspectral sensors. Measurements were taken in 10 field stations above homogeneous substrate or vegetation patches estimated visually from the boat. Multiple reflectance measurements (5 to 10) were performed in each field station, from which the average spectrum was calculated. Above water reflectance measurements provided ground reference for the assessment of image preprocessing steps. The Ramses two-sensor system included irradiance and radiance sensors. The irradiance sensor was used to measure downwelling spectral irradiance $E_d$ ($\mathrm{W}{\mathrm{m}}^{-2}{\mathrm{nm}}^{-1}$), and the radiance sensor was used to measure upwelling spectral radiance $L_u$ ($\mathrm{W}{\mathrm{m}}^{-2}{\mathrm{nm}}^{-1}{\mathrm{sr}}^{-1}$). Above water remote sensing reflectance was calculated as the ratio of $L_u/E_d$. Ramses remote sensing reflectance was multiplied by the $Q$-factor, which converted it to the irradiance reflectance, making it thereafter comparable to the outcome of the atmospheric correction. The $Q$-value may range from 3 to 6 steradians,⁴⁹ but as defining its value was outside of the scope of the current study, the $Q$-factor was considered equal to $\pi$.

Depth profiles for downwelling irradiance ($E_d$) were acquired with Ramses radiometers in four field stations around Pakri islands on the same day as the image collection. Ramses was lowered through the water column and measurements taken every 0.5 m from the surface to the depth of 2.5 m. The underwater $E_d$ depth profiles were used to calculate the diffuse downwelling attenuation coefficient ($K_d$), which was required to perform the water column correction (WCC). Water samples were collected concurrently with depth profiles, which were later analyzed for the concentrations of $a_{\mathrm{CDOM}}$ (${\mathrm{m}}^{-1}$), total suspended matter (TSM, $\mathrm{g}{\mathrm{m}}^{-3}$), and chlorophyll a (Chl-a, $\mathrm{mg}{\mathrm{m}}^{-3}$). Laboratory water sample analyses are described in Ref. 50.

2.3.

Airborne Remote Sensing

Airborne hyperspectral imagery was acquired on May 31, 2016, over the Pakri marine area with a CASI-1500 sensor produced by Itres, Canada (Fig. 1). The data collection was conducted by the Latvian Institute for Environmental Solutions (IES). CASI sensor covers a spectral range between 367 and 1045 nm. Bandwidths of the CASI-1500 sensor are programable, and different bandwidths were selected to optimize the signal-to-noise ratio (SNR) across the entire spectrum. The sensor has lower sensitivity in the blue and near-infrared (NIR) regions of the spectrum, and broader ($\sim 10\mathrm{nm}$) bandwidths were selected for that region. Narrower bandwidths ($\sim 5\mathrm{nm}$) were selected for the 550- to 750-nm spectral range. A total of 93 spectral bands were used for the data collection.

Altogether 12 flight lines were collected from the Pakri study area. The data were collected at the altitude of 2100 m, resulting in 1.0 m spatial resolution. Unfortunately, image quality suffered from brightness unevenness, resulting in the visible contrast in the central region of each flight line.

2.4.

Image Processing

A critical step in remote sensing data analysis is the preprocessing of images. Radiometric correction and geo-correction of the CASI data were conducted by the IES. After that, ENVI^® software (Research Systems Inc.) was used to apply subsequent preprocessing procedures: (1) atmospheric correction, (2) minimum noise fraction (MNF) transform, (3) glint correction, and (4) WCC.

2.5.

Benthic Habitat Classification

Collected underwater video records were analyzed for sediment properties, as well as for benthic vegetation species and coverage. As a result, a benthic habitat classification scheme in which prevailing habitat types were grouped into classes representing the pattern of the ecosystem was developed. Seven benthic classes were defined and labeled as (a) bare substrate, (b) filamentous green algae, (c) sparse higher order vegetation, (d) dense higher order vegetation, (e) charophytes, (f) brown macroalgae, and (g) deep water (Table 1 and Fig. 2). Red macroalgae were identified in some field stations, but as they were found in low quantities they were excluded from the classification scheme. The classification scheme for the current study was defined based on the dominant class. If, for example, the coverage of brown macroalgae in a particular sampling point was estimated to be 60% and the coverage of red macroalgae 20%, it was considered a point belonging to the brown macroalgae class. We cannot say that red macroalgae were completely absent in the Pakri study area, but they were not dominant in our sampling sites.

Table 1

The species and sediment composition contained within each of the seven benthic habitat classes.

Fig. 2

Classes defined for benthic habitat classification: (a) bare substrate, (b) filamentous green algae, (c) sparse higher order vegetation (vegetation $<30\%$), (d) dense higher order vegetation (vegetation $>30\%$), (e) charophytes, and (f) brown macroalgae.

A supervised classification procedure, which automatically divides all image pixels into previously defined habitat classes, was applied to the imagery. Information acquired by underwater video records was used for selecting training regions of interest (ROIs) for supervised classification. The quality of the classification results was assessed by the quantitative process of accuracy assessment utilizing ground reference data from video records. The ground reference data used for the validation of classified maps were nonoverlapping from the data used for training the classification algorithms.

2.6.

Bathymetry

In situ water depth measurements were used for the site-specific bathymetry algorithm development proposed by Stumpf et al.³³ The basic premise of the model is that, because water absorptivity is wavelength dependent, the upwelling radiance measured in a spectral band experiencing greater absorption will be less than that measured in a band with weaker absorption.⁵⁹ Thus with increasing water depth, reflectance values decrease proportionally faster in a band with greater absorption and slower in a band with weaker absorption. As depth increases, the reflectance ratio of low absorption band to high absorption band should show an increase and need to be scaled to the actual depth according to Ref. 33:

In our previous study,³⁷ different band combinations of Hyspex hyperspectral sensors were tested to specify spectral regions most suitable for the bathymetry mapping in the Baltic Sea conditions. The study showed that the ratio of green band to yellow band achieved the best correlations with measured water depths. Therefore, in this study, we concentrated on the green/yellow spectral regions and defined specific CASI bands, which allowed for the best correlation with the water depth. The best correlation was achieved using a wavelength ratio of $520/608\mathrm{nm}$. In addition, Vahtmäe and Kutser³⁷ showed that the best processing level for bathymetry retrieval was imagery with atmospheric correction. For that reason, the bathymetry model was applied to the atmospherically and MNF corrected CASI image.

The retrieved bathymetry was smoothed with a $3\times 3$ low-pass filter to reduce noise, which would otherwise potentially reduce the accuracy of WCC.

3.1.

Sea Surface Reflectance

Sea surface reflectance was retrieved from the CASI hyperspectral sensor after completing three image correction steps: atmospheric correction, MNF transform, and glint correction. The performance of those preprocessing steps was assessed by comparing the spectrum of the corrected image pixel with field spectral measurements of the corresponding pixel taken using the Ramses spectrometer. Some selected spectra of the CASI preprocessed spectra and Ramses in situ measurements are shown in Fig. 3.

Fig. 3

Comparisons of Ramses ground-truth spectra with corresponding different level corrected CASI spectra: (casi_1) atmospherically corrected spectra; (casi_2) atmospherically and MNF corrected spectra; and (casi_3) atmospherically, MNF and glint corrected spectra. (a) Sand 0.4 m, (b) brown macroalgae 1.0 m, (c) green macroalgae 0.5 m, and (d) optically deep water.

In general, reflectance spectra derived from the CASI sensor captured the overall spectral shapes of spectra measured with the Ramses spectrometer. The surface reflectance spectra retrieved after applying atmospheric correction (Fig. 3, casi_1) contained significant levels of noise. Brighter spectra (e.g., sand) contained lesser amounts of noise compared with the darker spectra (e.g., optically deep water) as brighter objects exhibit higher signal levels. The MNF transformation produced much smoother spectra (Fig. 3, casi_2). This shows that MNF transform may help to remove the residual noise from hyperspectral data.

Glint correction reduced reflectance for all visible bands (Fig. 3, casi_3). In general, glint corrected spectra became more similar to in situ measured Ramses spectra in absolute values than atmospherically corrected spectra. Nevertheless, for example, in situ measured green macroalga spectra [Fig. 3(c)] showed higher reflectance values than glint corrected spectra. The difference may be conditioned from the high spatial heterogeneity characteristic of Estonian coastal waters, namely, as the CASI pixel area is $1{\mathrm{m}}^2$, different vegetation species as well as substrate may contribute to the total signal in heterogeneous areas. The Ramses measurement area is only a few ${\mathrm{cm}}^2$ (depending on the measurement distance). Due to the different measurement areas, reflectances captured by different sensors do not always match, even though measured in the same location.

Some negative values below 720 to 750 nm were retrieved after the glint correction procedure. The reason for this is that the method used for the sun-glint removal assumes a negligible water-leaving signal in the NIR spectral region. In practice, however, there is always some “residual” radiance in the NIR part of the spectrum and the NIR signal can be significant if the water is optically shallow.^7,41,60 Therefore, in the current case, it can be assumed that some benthic NIR-reflectance contributed to the water-leaving NIR signal and caused overcorrection of reflectance values.

In addition, neither atmospheric nor glint correction managed to correct blue wavelengths of the CASI optically deep water spectra. Shorter wavelengths (below 520 nm) of optically deep water remained under-corrected as shown in Fig. 3(d).

3.2.

Diffuse Attenuation Coefficient

$K_d$ was calculated from in situ Ramses vertical profile measurements [Eq. (6)]. Vertical profiles were measured at four field stations, and for each station $K_d$ values along the water column were averaged. Finally, an average $K_d$ value between all profiles was obtained (Fig. 4). $K_d$ values increased exponentially beyond 700 nm due to the increasing water absorption. CDOM absorbs strongly in the blue spectral region and Chl-a in the blue and red spectral regions, which resulted in increased $K_d$ values in those spectral ranges.

Fig. 4

Downwelling diffuse attenuation coefficients, $K_d$ (${\mathrm{m}}^{-1}$), derived from $E_d$ in-water profiles at four field stations around the Pakri marine area.

It is seen that $K_d$ spectra had low variance throughout all sampled sites, suggesting that the water optical properties were relatively constant over the study site (Table 2). The concentrations of water column constituents clearly characterized a case 2 water type. CDOM concentrations were similar to the open parts of the Baltic Sea. TSM concentrations were rather high, probably due to resuspension of sediments by the wind. Chl-a concentrations were also typical of the Baltic Sea waters, remaining below the bloom limit ($5\mathrm{mg}/{\mathrm{m}}^3$).

Table 2

Water column optical constituents for Pakri marine area field sites.

3.3.

Benthic Substrate Detectability Limits

Detectability limits [Eq. (7)] were calculated for five main benthic substrate classes present in the Baltic Sea: red macroalgae, brown macroalgae, green macroalgae, higher order vegetation, and bare substrate. Each class was represented by species widely spread over the Baltic Sea and/or considered a key species in the Baltic Sea: Furcellaria lumbricalis (red macroalgae), Fucus vesiculosus (brown macroalgae), Cladophora glomerata (green macroalgae), and Potamogeton pectinatus (higher order vegetation). Bare substrate was represented by a sand spectrum. End-member spectra of the substrates were taken from our spectral library. If the end-member library contained several measurements per species, then the mean spectrum was calculated.

Figure 5 shows detectability limits for selected benthic species at given water attenuation properties (expressed as $K_d$, Fig. 4). It is seen that invalid values were retrieved for almost all substrates in cases of blue wavelengths (below 520 nm) while using Eq. (7) for detectability limits calculations. This was caused by the fact that the optically deep water spectrum, used in the calculation, was extracted from the CASI sea surface reflectance image. However, blue wavelengths of this extracted spectrum [Fig. 3(d)] exhibited higher spectral values than end-member spectra as they were under-corrected during atmospheric correction.

Fig. 5

(a) End-member spectra of selected substrates used in the detectability calculations and (b) the depth at which the CASI sensor can no longer differentiate a substrate from optically deep water. Refer to Fig. 4 for an average $K_d$.

The depth limits for benthic substrate detection varied depending on the magnitude of substrate reflectance. The substrate with the highest reflectance (e.g., sand) was discernible from deep water at the furthest depths. At the wavelength of deepest light penetration (near 570 nm), the depth restriction for CASI benthic substrate detection extended to 7.6 m for sand, 5.0 m for green macroalgae Cladophora, 3.0 m for higher order vegetation Potamogeton, and 3.1 m for brown macroalgae Fucus. In the red spectral region, the maximum depth of detection for most substrates remained around 2 to 3 m and in the NIR spectral region around 1 m.

Reflectance values of red macroalga Furcellaria remained lower than an average optically deep water spectrum almost throughout the entire spectrum, providing invalid results in the case of using Eq. (7).

3.4.

Bathymetry Mapping

Altogether, 175 in situ measured depth points from a depth range of 0 to 4 m were collected from the study area. Half of the measured depth points were used for the calibration and half for the validation of the bathymetry model. Figure 6 displays a scatterplot comparison of modeled water depths retrieved by the bathymetry model versus in situ measured depths. The validation of the derived bathymetry map revealed a correlation coefficient ($R^2$) value of 0.88 with a root-mean-square error (RMSE) of 0.32 m.

Fig. 6

Comparison of modeled water depths obtained by Stumpf et al. bathymetry model and in situ measured depths.

The water depth map obtained by the bathymetry model across the study site is shown in Fig. 7. Shallower areas (depth under 3 m) were found mostly in between two Pakri islands and south of islands. Some interesting features included, for example, sand dunes fringing the coast of the mainland and sand waves between Western Island and the mainland in 0.5 to 2 m of water. Water depth increased fast in the seaside coasts of the Pakri islands.

Fig. 7

Water depth map of the Pakri marine area obtained by Stumpf et al.: (a) bathymetry model from the atmospherically and MNF corrected CASI image and (b) CASI true color image from the Pakri study area.

Sudden depth variations visible in the bathymetry map (Fig. 7) were caused by across-track brightness gradients between multiple CASI flight lines. However, these artifacts were not important from the benthic habitat mapping point of view as they occurred in waters too deep for habitat mapping. Such brightness artifacts can be compensated for or at least reduced by applying brightness gradient correction methods such as the cross-track correction procedure in ENVI. The procedure should be applied to the individual radiometrically corrected flight lines prior to geocorrection. In the current case, CASI data were received from the IES that were already geometrically corrected. Therefore, the cross-track correction procedure was not applied to the data. Some negative depth values [marked with white in Fig. 7(a)] were also retrieved by the bathymetry model. These areas were mostly located in extremely shallow ($<0.1\mathrm{m}$), near-coastal water.

3.5.

Retrieved Bottom Reflectance

The performance of the WCC was evaluated by comparing the corrected spectra with measured substrates spectra from our spectral library (Fig. 8). The lack of substrate reflectances measured in situ in the study area impeded a quantitative estimation of WCC accuracy. Therefore, we are not able to say whether the overall magnitude of retrieved reflectance values was in agreement with the actual substrate reflectance spectra. Generally, the application of WCC increased reflectance of substrate spectra compared with the surface spectra. Our results showed that there was a closer agreement between end-member and WCC spectra in the case of shallower depths. Magnitudes of WCC reflectance spectra started to increase with increasing depth [Fig. 8(a)]. In addition, negative values in blue and red wavelengths were retrieved in the case of benthic vegetation at depths over 2 m.

Fig. 8

Reflectance spectra of different substrates at varying water depths extracted from the water column corrected CASI image and corresponding endmember spectra from our spectral library. Absolute reflectance values are displayed in (a) and standardized reflectance values in (b).

To enable better the comparison of spectral shapes, reflectance spectra were standardized based on the mean and standard deviation of the spectra [Fig. 8(b)]. Spectral shapes of bottom spectra retrieved by removing water column effects over green macroalgae (Cladophora sp.) and charophytes (Chara sp.) were consistent with end-member spectra—Chara displaying flatter green peaks compared with Cladophora. Sand end-member spectra increased in reflectance from 450 to 700 nm without showing any remarkable absorption features, while retrieved CASI sand spectra exhibited a chlorophyll absorption feature near 675 nm. Spectral shapes of retrieved higher order vegetation and corresponding end-member Potamogeton differed from each other to some extent. Reflectance maxima of in situ measured Potamogeton was shifted toward longer wavelengths (630 nm) compared with WCC spectra, which showed reflectance maxima near 570 nm. Shapes of retrieved brown algae spectra did not correspond to end-member Fucus spectra. Fucus exhibits peaks near 600 and 650 nm and a shoulder near 580 nm, whereas retrieved brown algae spectra, similar to green algae, displayed reflectance maxima near 570 nm and a distinctive drop in reflectance from 570 to 680 nm.

Our results showed that the Maritorena WCC method was unsuccessful for sites deeper than 2 m. In addition, some vegetation spectra (e.g., brown algae, higher order vegetation) were not retrieved correctly with the current WCC method even in very shallow water ($<1\mathrm{m}$). This suggested that the WCC method was not very useful in optically complex waters, and it was better to use above water reflectance imagery for benthic habitat mapping.

3.6.

Benthic Habitat Mapping

Previous studies^16,61 have shown that the spectral signal of the same bottom type varies depending on the water depth. Therefore, training ROIs for each benthic habitat class were selected from different water depths.

We used a two-level classification approach. In the first classification step, bare substrate (sand, gravel, limestone, etc.) was differentiated from benthic vegetation and deep water using a minimum distance (MD) classification algorithm. This was performed by generating 13 initial supervised classes (different bottom types at different depths), which were later merged into the two habitat classes of interest (bare substrate and vegetated bottoms). It was relatively straightforward to delineate bare substrate as sand exhibited comparatively brighter spectra than vegetation (Fig. 5). The resulting bare substrate class was masked out from the CASI image.

As a second step, a new classification procedure was applied to divide vegetated habitats into more detailed classes. Once again supervised classification algorithms, such as MD, maximum likelihood, spectral angle mapper (SAM), etc., were tested. The SAM classification algorithm provided the best result in vegetation mapping as indicated by the initial accuracy assessment. Therefore, it was used in the final classification.

ROIs for each vegetation class were selected from multiple depths considering the depth distribution limits of each substrate class. For example, during the field studies, charophytes were found only in shallower depths and ROIs for the “charophytes” class were selected from depths of 0.2, 0.5, and 1.0 m. At the same time, higher order vegetation was found from broader depth limits, and ROIs for this class were selected from depths of 0.5, 1.0, 1.5, and 2.5 m. The amount of ROIs was gradually increased for those classes, which showed significant confusion with other classes after initial classification. Altogether, 32 training ROIs were used (each ROI corresponds to one pixel, $1{\mathrm{m}}^2$), which generated 32 habitat classes. Finally, each habitat class at various water depths was merged into the single habitat class, resulting in the water depth-independent benthic habitat map. The benthic habitat map obtained after the two-level classification is presented in Fig. 9.

Fig. 9

Benthic habitat map of the Pakri marine area obtained from the CASI airborne imagery using supervised classification and seven benthic classes described in Table 1 and Fig. 2.

The classification results showed that the bare substrate was widespread in the area between the Pakri islands and mainland, alternating with dense and sparse higher plants communities. Dense higher vegetation was also mapped between two Pakri islands, mostly in the depths over 1.5 m. Charophytes were found in shallower waters between islands and on soft sandy shores near the coast. Field studies indicated that Charophytes were often covered with rather dense filamentous green algae, particularly on the coast of the mainland. Therefore, it is not surprising that these areas were classified as filamentous green algae by the classification algorithm. Brown algae were estimated to occur mainly on the outer sides of the Pakri islands.

The performance of the classification was assessed using the confusion matrix method, where classified benthic habitats were compared with in situ observed benthic habitat classes. 166 field control points ($<3\mathrm{m}$ water depth) were used for the accuracy assessment, resulting in the overall accuracy of 80% (Table 3).

Table 3

Classification accuracy of the hyperspectral CASI image.

The confusion matrix method produces also producer and user accuracies, which report individual class accuracies (Table 3). The lowest user accuracy was encountered for the class “sparse higher vegetation” (50%), meaning that only half of image pixels classified as “sparse higher vegetation” were actually this habitat class and another half belong to some other habitat class. It also means that this class was the most overestimated in the Pakri study area. The lowest producer accuracy was identified for the class “brown macroalgae” (60%), which means that 60% of pixels identified as “brown macroalga” during the field work were classified to the correct habitat class. The lowest producer accuracy indicates that the class “brown macroalgae” was the most underestimated in our study area.