4.1. Assessment of Remote Sensing Algorithms for the Retrieval of Inherent Optical Properties, and Chlorophyll a Concentrations in the Arctic Ocean
Ocean colour remote sensing has been used to study the spatial and temporal dynamics of phytoplankton abundance across a range of spatial and temporal scales in the global ocean, regional marine basins and fresh water bodies [
23]. However, a number of factors limit the capability of ocean colour remote sensing to provide daily synoptic maps of essential optical and biogeochemical variables at high latitudes. They include low Sun zenith angle, persistent cloudiness and fog, and ice cover that screens the under-ice algal dynamics, as well as further errors in remote sensing reflectance retrievals caused by the so-called ice adjacency effect, the occurrence of a deep chlorophyll maximum and, last but not least, the optical complexity of Arctic Ocean waters, especially over its shelves [
31,
74]. It has already been found that relationships between remote sensing reflectance band ratios and chlorophyll
a concentrations in polar regions differ markedly from those prevailing in lower latitude oceanic waters [
75,
76,
77]. The global remote sensing algorithms were adjusted for Arctic Ocean waters [
78] and the new OC4L algorithm, applied to Western Arctic Ocean shelf waters characterized by strong CDOM absorption, retrieved chlorophyll
a concentrations with the requisite accuracy for this parameter (RMSE < 35%) [
79]. Recently, new regional algorithms have been developed for the Chukchi Sea [
80] and Canadian Northeast Pacific and Northwest Atlantic waters for the use of data from the MODIS and VIIRS satellite sensors [
81]. Chlorophyll
a concentrations retrievals acquired using generic band ratio algorithms were encumbered with a lower uncertainty compared to the standard global ones.
To date there have been few studies for developing a regional algorithm for chlorophyll
a concentration retrievals in the polar North Atlantic and European Arctic, except the one by Stramska et al. (2003) [
32]. Those authors used a second-order polynomial band ratio algorithm based on the
/
ratio. Although we followed that approach in our study, we used a significantly larger in situ data set to develop our algorithms. The observed log-normal distribution and formulations of remote sensing algorithms for retrieving IOPs of optical water based on log-transformed data are consistent with previous research [
29]. High correlation coefficients with band ratios around 443–490 nm to 550–560 nm were expected. These wavelengths, centred near the chlorophyll
a absorption peak in the blue and the minimum in the green, are commonly used for developing remote sensing algorithms in open ocean water [
82]. It is interesting that estimations based on the blue channel around 490 nm seem to be more accurate and stable than those based on bands around 443 nm. Our observations are consistent with O’Reilly et al. (2000) [
29] who stated that the band ratio based on the bands close to
/
were the best overall index of chlorophyll
a concentration. This is particularly evident in the case of for the OLCI radiometer, which measures radiation in the narrower bands. Intensive absorption leads to a severe decrease in the light level at these wavelengths, which may result in higher signal fluctuations and bigger statistical errors owing to the limited sensitivity of measuring devices. Our samples showed that the a
peak was placed around 443 nm and was much higher than at 490 nm. The CDOM contribution to the total non-water absorption a
in the study area varied between 21% and 51% [
10] and decreased rapidly with increasing wavelength, and the
waveband was much less affected by CDOM absorption. For this reason, all the algorithms that included the 490 nm waveband performed better than those based on 443 nm. The total non-water absorption studied in this part of the Nordic Seas was dominated by both components of particulate absorption: absorption by phytoplankton pigments, a
and absorption by non-pigmented particles a
, especially at longer wavelengths. In the near infrared at 670 nm, the absorption budget is controlled almost entirely by phytoplankton pigment absorption alone [
10,
12]. Because absorption constituents strongly related to chlorophyll
a are dominant, we were able to establish reliable bio-optical relationships between the chlorophyll
a concentration and absorption by particulates and phytoplankton pigments at 443 nm and 670 nm [
10,
11,
12,
32] that can be used in semi-analytical remote sensing reflectance inversion models.
The chlorophyll
a concentration estimated from satellite-derived spectral reflectances had a bigger RMSE than the in-water based estimates. This suggests that the atmospheric correction still contributes significantly to the overall error (
Figure 8). The correction using SWIR bands for water masses of high turbidity [
62] simultaneously exhibits a smaller bias, but a higher statistical error (
Table 4), because MODIS images have a significant noise level in the SWIR part of the spectrum. Water carrying larger amounts of suspended matter scatters in a spectrally-uniform way and raises the radiation levels in the NIR bands, which are commonly used as a zero-level reference to exclude the influence of the atmosphere [
61]. In the marginal sea ice zone, the presence of melt water containing large amounts of suspended particles enhances scattering [
18,
68] and the assumption of negligible values water-leaving radiance in the NIR band may not be fulfilled locally.
The waters of the Nordic Seas are highly productive, especially in the West Spitsbergen Current (WSC), where phytoplankton growth is stimulated by ample supplies of nutrients and the less intensive light attenuation compared to the East Greenland Current EGC [
15]. Chlorophyll
a concentrations typically peak in the summer from May until July [
7,
9]. The majority of sampling in the WSC was done in June and July over three years, which could have been responsible for the greater sample homogeneity and the smaller number of extremes. The results obtained with the second-order polynomials were therefore almost as good as with the third-order ones, which may not be true for the whole year. The low level of uncertainty of our algorithms and their consistency in retrieving spectral values of particulate and total non-water absorption coefficients demonstrated that we were right to apply the waveband combination ratio using satellite bands at 490 nm in the empirical formulations. In the cases where there were rapid increases in CDOM absorption, the water-leaving radiance signal recorded by satellite wavebands at shorter wavelengths would be disturbed [
22]. New formulations of the empirical relationships defined for the MODIS and OLCI bands displayed certain similarities (
Table 2 and
Table 3). The band ratio indices were based on almost the same wavelengths. Second-order polynomials applied to log-transformed data turned out to be suitable in most cases for both sensors. Even though the satellite sensors have different bandwidths, statistical quality assessment revealed very close levels of algorithms performance. The algorithms for MODIS performed somewhat better, in particular for a
and a
. This is the result of the narrower bandwidths of the OLCI radiometer in relation to MODIS, which seem to be more sensitive to CDOM absorption. With both satellites, moreover, the accuracy of a
and a
retrieval was less than of a
and a
.
The semi-analytical algorithms for retrieving spectral values of IOPs like GSM [
27] or QAA [
71] were developed using a dataset in which the majority of data were gathered in temperate and subtropical waters [
31]. The accuracy of a semi-analytical algorithm depends closely on parametrizations between phytoplankton pigment absorption and chlorophyll
a concentration, and the backscattering spectrum and backscattering phase function. These parameters have changed considerably in the Nordic Seas and European Arctic Ocean. Stramska et al. (2003) [
32] reported that phytoplankton communities dominated by diatoms had a higher b
/b
ratio and that its spectrum was steeper with increasing chlorophyll
a concentration compared to phytoplankton communities dominated by dinoflagellates and coccolithophorids. Changes in b
/b
ratio impacted the blue–green reflectance ratio. Part of the Nordic Seas where phytoplankton communities were dominated by diatoms were characterized by relatively high blue-to-green reflectance ratio, compared to water where phytoplankton communities were dominated by diatoms were characterized by a relatively high blue-to-green reflectance ratio, compared to water in which phytoplankton communities were dominated by dinoflagellates, where the reflectance ratio was much lower. The last two decades have witnessed a significant northward expansion of temperate phytoplankton communities into Arctic Ocean waters [
15,
16] bringing about significant changes in their bio-optical properties [
10,
12,
83]. The poorer accuracy of semi-analytical algorithms in high latitude oceanic waters, documented by Clay et al. (2019) [
81], may be explained by changes in phytoplankton phenology and the associated bio-optical properties in these waters.
Figure 3 also clearly shows that reflectance spectra tend to congregate in two clusters with lower and higher values of R
in the 400–550 nm spectral range. The cluster of low R
values is clearly visible in the R
spectrum measured in summer 2014 and 2015. Sea surface temperatures in those years were significantly higher than in 2013 owing to the enhanced inflow of warm AW observed in Fram Strait at that time [
84]; the intensification of AW inflow contributes to the poleward expansion of temperate phytoplankton in the Arctic Ocean [
85].
Stramska et al. (2003) proposed a band ratio algorithm for the retrieval of a
and a
[
32]. The log-linear parametrization was used to estimate the values of both coefficients, utilizing the R
/R
and R
/R
band ratios. Our results demonstrated that this band ratio combination was also effective deriving total and particulate absorption at 443 and 670 nm for both the MODIS and OLCI satellite radiometers. The only difference was the second-order polynomial parametrization proposed in this paper for retrieving
Chla, a
and a
from MODIS and OLCI data. The exceptions were the formulas for retrieving a
and a
from OLCI. For these two cases we suggest linear parametrizations. All the absorption coefficients proposed in this paper are characterized by a low level of uncertainty, and the values of the respective coefficients are consistent with the data reported by Kowalczuk et al. 2017 and 2019 [
10,
12].
CDOM in AW dominating in the eastern part of Fram Strait has a relatively low level of absorption and is predominantly of marine origin [
9,
86]. The CDOM absorption coefficient at 350 nm a
is on average up to three times lower than of the water masses influenced by EGC [
7]. Statistically significant inter-annual variabilities in the level of CDOM absorption in Atlantic-dominated waters between 2009–2010 were reported by Pavlov et al. (2015) [
7] and between 2013–2015 by Makarewicz et al. (2018) [
9]. In the Nordic Seas, the IOPs of water on the opposite sides of the Polar front differ significantly as regards the proportions of absorbing constituents [
7,
13,
87]. The CDOM in the western part of Fram Strait and on the East Greenland shelf CDOM originates from the riverine input of Siberian Rivers to the Arctic Ocean [
7,
13,
14]. The CDOM absorption level in PW carried southwards with EGC, though much higher than in AW, is temporally more stable [
13,
88,
89] and could be effectively lowered by fresh, glacier melt waters from the Greenland ice sheet [
88,
89] or drift sea ice melt [
68]. In spite of the broad spatial and temporal variability in CDOM absorption in the Nordic Seas, the accuracy of
Chla,
,
,
and
retrievals in our algorithms was not sensitive to changes in this parameter.
4.2. The Role of Sea Ice Cover in the Spatial and Temporal Variability of Inherent Optical Properties in the Nordic Seas
Particulate absorption at 443 and 670 nm is a very reliable proxy of the chlorophyll
a concentration in the Nordic Seas [
10,
12,
32] and is a convenient tool for observing phytoplankton dynamics. The monthly averaged a
products from May to August in three consecutive years from 2013 to 2015 (
Figure 9) revealed a distinct temporal and spatial zonation of phytoplankton dynamics in the Nordic Seas. The bloom started in May along the sea ice marginal zone, along the Greenland shelf in the western part of Fram Strait, whereas in the eastern part of Fram Strait and along the West Spitsbergen shelf, phytoplankton abundance was relatively low. As summer progressed, the phytoplankton was displaced northwards along the sea ice marginal zone and eastwards towards the central part of Fram Strait, but the bloom intensity decreased (
Figure 9). In late summer (July and August), the phytoplankton concentration declined. It was distributed in the central and eastern parts of Fram Strait, and followed the northern part of the sea ice edge in the Nansen Basin and on the northern Spitsbergen Shelf. This zonation is consistent with the previous observations using the satellite imagery of chlorophyll
a concentrations [
90] and the modelling studies [
91]. Mayot et al. (2020) [
92] also found that the intensity of spring sea ice exported from the Arctic Ocean through Fram Strait had a significant influence on Greenland Sea phytoplankton abundance. He reported that in those years when sea ice export was high, stratification was stronger and phytoplankton blooms were more intense in the Greenland Sea. We, too, witnessed such an effect in the monthly averaged a
products when comparing July 2014 and 2015: in July 2014 the phytoplankton bloom in the central Fram Strait was more intense than in July 2015. This is wholly consistent with the in situ observations presented by Kowalczuk et al. (2019) [
10]. In 2014, the mostly southerly winds brought about advection of sea ice towards the central Fram Strait, intensifying stratification and producing a stronger phytoplankton bloom. In July 2015 the predominantly northerly winds caused sea ice to converge along the east coast of Greenland: the oceanic waters in the central and eastern parts of Fram Strait were less stratified and the phytoplankton bloom was weaker.
A narrow belt of less absorbing water is usually visible between the ice edge and the areas with elevated a
(
Figure 9). This is the effect of multiple processes due to the broad temporal averaging bracket (one month) used in mosaicking. Such a long time leads to strong fluctuations in the ever-moving ice edge position. Another process is the sea adjacency effect, which influences atmospheric correction performance. Last but not least is the vertical distribution of phytoplankton. The presence of melt water in the sea ice marginal zone generates pycnocline, which forces the phytoplankton to aggregate along the density gradient between the melt and oceanic waters, creating a deep chlorophyll
a maximum [
10,
93]: its presence means that part of the phytoplankton biomass is located at depths well below satellite radiometer detection limits [
90]. As summer progressed, the sea ice melted and vast areas of the Eastern Greenland shelf waters were exposed, but we did not record any elevated absorption. This must have been due to the optical properties of the melt water, which is low in phytoplankton pigments and containing diluted CDOM compared to the underlying CDOM-rich East Greenland Current waters [
13,
68]. The mineral particles that melt out from ice and are contained in the melt water made a greater contribution to scattering than to absorption [
18].