Sea-ice concentrations derived from satellite microwave brightness
temperatures are less accurate during summer. In the Arctic Ocean the lack of
accuracy is primarily caused by melt ponds, but also by changes in the
properties of snow and the sea-ice surface itself. We investigate the
sensitivity of eight sea-ice concentration retrieval algorithms to melt ponds
by comparing sea-ice concentration with the melt-pond fraction. We derive
gridded daily sea-ice concentrations from microwave brightness temperatures
of summer 2009. We derive the daily fraction of melt ponds, open water
between ice floes, and the ice-surface fraction from contemporary Moderate
Resolution Spectroradiometer (MODIS) reflectance data. We only use grid cells
where the MODIS sea-ice concentration, which is the melt-pond fraction plus
the ice-surface fraction, exceeds 90 %. For one group of algorithms,
e.g., Bristol and Comiso bootstrap frequency mode (Bootstrap_f), sea-ice
concentrations are linearly related to the MODIS melt-pond fraction quite
clearly after June. For other algorithms, e.g., Near90GHz and Comiso
bootstrap polarization mode (Bootstrap_p), this relationship is weaker and
develops later in summer. We attribute the variation of the sensitivity to
the melt-pond fraction across the algorithms to a different sensitivity of
the brightness temperatures to snow-property variations. We find an
Sea-ice area and extent are derived from the sea-ice
concentration, i.e., the fraction of a given area of the ocean covered with
sea ice. Observations of the brightness temperature by satellite passive
microwave sensors have been the backbone of sea-ice concentration retrieval
for more than 35 years because these are independent of daylight and are
quite insensitive to the cloud cover. These satellite sensors measure the
brightness temperature at window frequencies between 6 and
One potential reason for the reduced accuracy is the change in microphysical properties inside the sea ice, for instance, the desalination of the sea ice during the melt process or the flushing of air voids in multiyear ice with meltwater and other melt processes, as for example described in Scharien et al. (2010). Another potential reason is the change in surface properties of the sea ice. The three key surface features of summer melt on Arctic sea ice are a metamorphous, wet snow cover, a porous, wet sea-ice surface, and melt ponds. During summer, the snow cover on sea ice is usually wet or even saturated with meltwater (Garrity, 1992). Its density is usually considerably larger during summer than during winter (Warren et al., 1999; Maykut and Untersteiner, 1971). Diurnal melt–refreeze cycles, i.e., episodes of intermittent melting and refreezing of the snow, which is a common phenomenon during late spring, result in an increase in the snow grain size.
Wet snow is an efficient absorber of microwave radiation and has a microwave emissivity close to 1. It can effectively block microwave emission from underneath, and thereby masks differences in volume scattering between first-year and multiyear ice. Therefore microwave brightness temperatures of sea ice covered with wet snow usually are close to its physical temperature during melt (e.g., Stiles and Ulaby, 1980; Eppler et al., 1992; Hallikainen and Winebrenner, 1992; Garrity, 1992).
During the melt phase of melt–refreeze cycles, coarse-grained snow can be
regarded to behave similarly to wet snow due to its wetness. During the
refreeze phase, however, when it is dry, it absorbs less microwave radiation
than wet snow, and there is more scattering from within the snow. Therefore,
dry coarse-grained snow does not block or mask microwave emission and volume
scattering differences of the sea ice and/or snow underneath as efficiently
as wet snow does. The amount of volume scattering depends on microwave
frequency and polarization, and on the vertical location of the
coarse-grained snow layers relative to the snow surface. Because the
electromagnetic wavelengths are closer to the snow grain size at higher
frequencies, i.e., at 37–90
Melt ponds are puddles of meltwater on top of sea ice. They form during
summer from melting snow and sea ice. Their areal fraction, size and depth is
determined by the onset, length and severity of the melting season, the
sea-ice type and topography, and the snow-depth distribution at the beginning
of melt (Landy et al., 2014; Polashenski et al., 2012; Petrich et al., 2012;
Eicken et al., 2004; Perovich et al., 2002). The melt-pond water salinity is
close to 0
Typical values of changes in brightness temperature due to changes in snow wetness. Abbreviations TB, PR, GR, V, and H denote brightness temperature, normalized brightness-temperature polarization difference (“polarization ratio”), normalized brightness-temperature frequency difference (“gradient ratio”), vertical, and horizontal (polarization), respectively. Abbreviations E92, W14, and G92 refer to Eppler et al. (1992), Willmes et al. (2014), and Garrity (1992), respectively.
Typical values of changes in brightness temperature due to changes in snow density. For abbreviations TB, PR, GR, V, and H see Table 1. Abbreviations F98 and B15 refer to Fuhrhop et al. (1998) and Beitsch (2014), respectively.
The penetration depth into liquid water of microwave radiation at the
frequencies used here, i.e., between 6 and 89
This has consequences for climate research. For example, the sea-ice area, which is defined as the sum of the area of all sea-ice covered grid cells weighted by the sea-ice concentration, will be underestimated from case A but will be overestimated from case B. The ambiguity in the actual surface properties related to the sea-ice concentration value of 60 % in the example above is also challenging for the initialization and evaluation of numerical models, and the assimilation of sea-ice concentration data into such models. An unambiguous sea-ice concentration is required for, e.g., the correct computation of the sea-ice volume. In the terminology of the more advanced thermo-dynamic and dynamic sea-ice models or model components which treat leads and melt ponds separately (e.g., Holland et al., 2012; Flocco et al., 2010), the fraction of sea ice covering the open ocean is called sea-ice concentration and includes melt ponds. The fraction of the latter is given separately as the area of the sea-ice surface covered by melt ponds and is called melt-pond fraction. It is obvious that such models would have difficulties using a sea-ice concentration product which is biased like described above for cases A and B. Even numerical models, which are not as advanced and which do not treat melt ponds separately, would have difficulties using such a product.
Approaches have been developed, which permit the melt-pond fraction on sea ice to be derived from satellite observations in the visible/near-infrared frequency range (Istomina et al., 2015a, b; Zege et al., 2015; Rösel et al., 2012a). Their results could be used to correct the above-mentioned ambiguity by quantifying how much of the open water seen (30 % in the example above) is actually caused by melt ponds. However, the time series of melt-pond fraction data computed so far (2002–2009 and 2002–2011) are too short to apply such a correction for the entire sea-ice concentration data set, over 35 years long, from satellite microwave radiometry. In addition, such data may have limitations due to cloud cover and the viewing geometry at high latitudes (see Sect. 2.1).
Typical values of changes in brightness temperature due to changes in snow grain size. For abbreviations TB, PR, GR, V, and H see Table 1. Abbreviations F98, W14, and G92 refer to Fuhrhop et al. (1998), Willmes et al. (2014), and Garrity (1992), respectively.
Spatial distribution of the MODIS sea-ice parameter data set
superposed with the fraction of first-year ice
This calls for a better quantification of the uncertainty and/or of potential biases in the sea-ice concentration. How sensitive are present-day sea-ice concentrations algorithms to the melt-pond fraction? How do these algorithms differ with respect to the expected bias due to melt ponds, and how can we explain these differences? We hypothesize that microwave brightness temperatures and sea-ice concentrations derived from them change linearly with the increase in surface-water fraction or the decrease in net sea-ice surface fraction due to melt ponds. To the authors' best knowledge, an intercomparison of different algorithms which incorporates contemporary information of the melt-pond fraction and an independent sea-ice concentration estimate, as is the aim of this study, has not previously been carried out.
In the present paper we illustrate how satellite microwave brightness-temperature measurements vary with the net sea-ice surface fraction derived
from satellite visible/near-infrared (VIS/NIR) imagery. We compare the
sea-ice concentration obtained with different sea-ice concentration retrieval
algorithms from these brightness temperatures with the sea-ice concentration
and with the net sea-ice surface fraction from VIS/NIR imagery. We isolate
the influence of melt-pond fractions on the net sea-ice surface fraction by
limiting our analysis to VIS/NIR imagery sea-ice concentrations
The paper is organized as follows. Section 2 describes the data sets and methods used for the intercomparison of brightness temperatures and sea-ice concentrations derived with several algorithms and the melt-pond fraction. In Sect. 3 we are going to present the results of this intercomparison, which we discuss in Sect. 4. Section 5 concludes our findings.
The paper focuses on the melt season, i.e., the months of June, July, and August, of the year 2009. The spatial domain of our investigations is a region of the Arctic Ocean (see Fig. 1). This region is determined by the area and data which we chose to compute the sea-ice cover parameters from satellite VIS/NIR imagery, described in Sect. 2.1.
We derive the open-water fraction, melt-pond fraction, and net sea-ice
surface fraction from reflectance measurements of the Moderate Resolution
Imaging Spectroradiometer (MODIS) aboard the Earth Observing System (EOS)
satellite TERRA. We use the MODIS Surface Reflectance daily L2G Global 500 m
and 1 km product (MOD09GA,
The quality of MODIS reflectance measurements carried out at high latitudes
may be degraded from high sun zenith angles, long pathways through the
atmosphere, cloud shadows, and, in addition, shadows caused by ridges in the
sea-ice cover. We use only reflectance values with the highest quality. This
ensures that cloudy pixels and pixels with cloud shadows, pixels with sun
zenith angles
Histograms of MODIS melt-pond fraction
Mäkynen et al. (2014) hypothesized that our daily MODIS melt-pond
fractions are positively biased by about 5–10 % during early melt. In
situ observations carried out north of Greenland revealed a melt-pond
fraction of 0 % and a sea-ice concentration of 100 % during the first
2 weeks of June 2009 (Mäkynen et al., 2014). Melt onset dates given in
Perovich et al. (2014) support this observation. In order to confirm this
notion, we derived histograms of the MODIS melt-pond fraction and the MODIS
sea-ice concentration using the data with 12.5 km grid resolution for
latitudes north of 83
Even though a state-of-the-art cloud masking scheme has been applied to the
MODIS reflectance data before the MODIS sea-ice parameter retrieval
(Rösel et al., 2012a), there is still a substantial number of
misclassified grid cells. It has been demonstrated that even with a
multi-channel instrument such as MODIS, cloud classification is a challenge
over bright surfaces such as sea ice or snow (Chan and Comiso, 2013; Karlsson
and Dybbroe, 2010). In order to mitigate the influence from
misclassifications due to residual clouds, we only use 100 km grid cells
with a cloud cover
The potential misclassification of one of the surface types is more important. The reflectance values used are fixed for the entire summer season and the entire Arctic domain. Therefore the MODIS sea-ice parameter retrieval does not account for the spatiotemporal variability in the spectral properties of the melt ponds or the non-ponded sea ice. These spectral properties change as a function of ice type and melt-season duration. The spectral properties of melt ponds on first-year ice are likely to approach those of leads and openings as the melt season progresses, while for melt ponds on multiyear ice these change less due its larger thickness and different internal structure. This could result in an overestimation of the melt-pond fraction relative to the open-water fraction for first-year ice or vice versa because the spectral space between sea ice and water is larger than between melt ponds and open water (leads). Such a misclassification would have, however, no implications for the net sea-ice surface fraction. It would affect only the melt pond or the open-water fraction. Therefore such a misclassification does not likely influence the main results of the present paper, but should be kept in mind when interpreting MODIS sea-ice concentrations (see Sect. 3.2).
Rösel et al. (2012a) report root mean square difference (RMSD) values
between MODIS melt-pond fraction and independent melt-pond fraction
observations of 4–11 %. We compare the MODIS sea-ice concentration with
visual ship-based sea-ice concentration observations from seven ship
expeditions into the Arctic Ocean, and obtain an average RMSD of (
At the time of our analysis and writing, this MODIS product was the best we could have, despite the above-mentioned limitations due to cloud cover and spatiotemporal variation of the ice-type-dependent spectral properties of the summer sea-ice cover. The results of our quality analysis and the results of Marks (2015) confirm that we can take the MODIS sea-ice parameters as kind of the ground truth against which we compare brightness temperatures and sea-ice concentrations in Sects. 3 and 4.
We use brightness temperatures measured by the Advanced Microwave Scanning
Radiometer aboard the EOS-TERRA satellite: AMSR-E. The AMSR-E data used
are from the 6.9, 10.7, 18.7, 36.5, and 89.0
We compute sea-ice concentrations from this set of co-located AMSR-E brightness temperatures (Sect. 2.2) using eight selected sea-ice concentration algorithms investigated in the European Space Agency Climate Change Initiative – Sea Ice (SICCI) project. The full suite of sea-ice concentration algorithms used in the SICCI project is documented in the SICCI project reports: PVASR (Ivanova et al., 2013) and ATBD (Ivanova et al., 2014), together with the tie points for open water and sea ice. The tie points represent winter conditions. The motivation for this is twofold. One is our wish to intercompare the eight algorithms independently of individual tie points being specifically selected in the original algorithms. We want to use one universal set of tie points (see also Ivanova et al., 2015). This implies the second reason as to why we use winter tie points in the present study. For the derivation of the sea-ice tie points, Ivanova et al. (2015) used high ice concentration areas of convergent ice motion during winter. This ensures that (i) the areas from which tie points are retrieved are large enough and (ii) the areas have indeed 100 % sea-ice concentration. Such an approach does not work under summer conditions because openings/leads in the ice cover do not freeze over. In the present study we focus on a selected number of different (representative) types of algorithms (Ivanova et al., 2015), and do not include algorithms where a methodology is duplicated. The selected algorithms are summarized in Table 4.
We categorize the algorithms into four types based on the way brightness temperatures are used: (1) algorithms based on one polarization and one frequency (e.g., One_channel 6H); (2) algorithms based on different frequencies but with the same polarization, such as the frequency mode of the Comiso bootstrap algorithm (Bootstrap_f); (3) algorithms based on different polarizations but with the same frequency, such as the polarization mode of the Comiso bootstrap algorithm (Bootstrap_p); (4) algorithms based on at least two frequencies and/or polarizations, like the NASA Team algorithm (NASA_Team).
The sea-ice concentration algorithms.
A fifth type of algorithms is given by the so-called hybrid algorithms. These
combine two or more of the above-mentioned types of algorithms, such as the
EUMETSAT OSI-SAF algorithm (Eastwood et al., 2012) or SICCI (Ivanova et
al., 2015), which combine Bristol and Bootstrap_f or CalVal, which is
identical to Bootstrap_f, and the Arctic version of the Comiso bootstrap
algorithm (Comiso et al., 1997; Comiso, 2009), which combines Bootstrap_f
and Bootstrap_p. For the high sea-ice concentrations we focus on in this
paper, these two hybrid algorithms are almost identical to the algorithm
that is employed at high sea-ice concentrations; this is Bristol in the case of
the OSI-SAF (zero weight at
Time series of open-water and sea-ice fractions for all MODIS grid cells used in the present study for 1 June 2009 to 31 August 2009.
Brightness-temperature changes over Arctic sea ice are different for
first-year ice (FYI) and multiyear ice (MYI) (Eppler et al., 1992). In order
to separate these two sea-ice types, we use the Arctic sea-ice age data set
(Tschudi et al., 2016). This data set is available with weekly temporal
resolution, has a grid resolution of
Statistical parameters of the comparison AMSR-E sea-ice
concentration vs. MODIS sea-ice concentration (see Fig. 4) for June 2009.
Each column gives the value for all grid cells with MODIS sea-ice
concentration
We show the temporal development of the daily sea-ice parameters obtained
with MODIS (Sect. 2.1) for June to August 2009 in Fig. 3. These include MODIS
sea-ice concentration, the net sea-ice surface fraction, the net
surface-water fraction, which is the open-water fraction plus the melt-pond
fraction, and the melt-pond fraction for each day and each co-located grid
cell. No further averaging is applied, and we show all grid cells regardless
of ice type. Gaps in the time series and the varying number of data points
are caused by daily variations in cloud cover and the decrease in sea-ice
cover from June to August. Only grid cells with MODIS sea-ice concentration
During the first 2–3 weeks, the MODIS melt-pond fraction in our data set remains near zero. Then the melt-pond fraction starts to increase, first slowly: days 20–30 (fifth and sixth 5-day period or pentad of June), then rapidly: days 30–45 (first to third pentad of July). After a short plateau, where the melt-pond fraction remains near 35 %, it first declines rapidly to about 20 % at days 55–60 (last pentad of July) and then more slowly to about 15 % until the end of our study period (31 August). Throughout June, MODIS sea-ice concentrations are close to 100 % until day 30, and then there is more variability around 90–95 % after day 55. Net total water fraction and net sea-ice surface fraction are linked to the previous two parameters and add up to 100 %.
AMSR-E sea-ice concentration computed with six of the eight
algorithms listed in Table 2 vs. MODIS sea-ice concentration for all grid
cells with MODIS sea-ice concentration
Statistical parameters of the comparison AMSR-E sea-ice
concentration vs. MODIS sea-ice concentration (see Fig. 4) for July 2009.
Each column gives the value for all grid cells with MODIS sea-ice
concentration
Statistical parameters of the comparison AMSR-E sea-ice
concentration vs. MODIS sea-ice concentration (see Fig. 4) for August 2009.
Each column gives the value for all grid cells with MODIS sea-ice
concentration
We first compare sea-ice concentrations derived with the algorithms listed in
Table 4 from AMSR-E brightness temperatures (Sects. 2.2 and 2.3) with MODIS
sea-ice concentrations (Sect. 2.1), with the aim of illustrating how summertime
AMSR-E sea-ice concentrations compare to an independent sea-ice concentration
estimate. We include all data with MODIS sea-ice concentrations
We take the slope (the closer to 1 the better), the correlation (the higher the better), and the RMSD (the lower the better) as a quality measure and find the NASA Team algorithm to outperform all other algorithms listed in Table 4 for June (Table 5) – no matter whether we use all grid cells or only FYI or MYI grid cells (see Sect. 2.4). For July (Table 6), the NASA Team algorithm is as good as the Near90_lin algorithm. For August (Table 7), best slopes are obtained for the Bootstrap_p algorithm, while lowest RMSD values are obtained for the NASA Team algorithm. Note that the number of FYI grid cells is extremely low for August and that the numbers given in Table 7 for FYI should not be over-interpreted.
The average correlation, computed from six algorithms, decreases from June:
Statistical parameters of the comparison AMSR-E sea-ice
concentration vs. MODIS ice-surface fraction for MODIS sea-ice concentration
Statistical parameters of the comparison AMSR-E sea-ice
concentration vs. MODIS ice-surface fraction for MODIS sea-ice concentration
Statistical parameters of the comparison AMSR-E sea-ice
concentration vs. MODIS ice-surface fraction for MODIS sea-ice concentration
AMSR-E sea-ice concentration computed with six of the eight
algorithms listed in Table 2 vs. MODIS ice-surface fraction for all grid
cells with MODIS sea-ice concentration
We compare AMSR-E sea-ice concentration (Sect. 3.2) with the MODIS
ice-surface fraction (Sect. 2.1) for grid cells with MODIS sea-ice
concentration
Parameter spaces for NASA Team
We obtain slope, correlation coefficient, and RMSD values of all eight
algorithms (see Table 4) separately for (i) all grid cells, (ii) only the FYI
grid cells, and (iii) only the MYI grid cells (see Sect. 2.4), and summarize
these in Tables 8–10 for June, July, and August. For August we exclude all
values obtained for FYI grid cells because of their low count of 44
(Table 10). We find an increase in the slopes from June to July for all
algorithms, which is followed by a decrease for 6H, Bootstrap_f and
Bristol algorithms but a further increase for Bootstrap_p and
Near90_lin algorithms from July to August. Correlations between AMSR-E
sea-ice concentrations and MODIS ice-surface fractions are below 0.4 in June
(Table 8). In contrast, for July (Table 9) we obtain correlations
We carried out the same intercomparison using a MODIS sea-ice concentration threshold of 98 % (not shown) instead of 90 %. By using 98 %, no results can be obtained for August because of too few valid data. For June and July, slopes remain similar to those in Tables 8 and 9. For June, correlations are considerably smaller compared to using 90 %. Correlations are a bit higher for July. Despite this better correlation in July, the peak melting period (see Fig. 3), we decided to keep the 90 % threshold to ensure a large enough number of data points. The results of the previous paragraph remain the same for 90 and 98 % MODIS sea-ice concentration threshold.
We conclude the following: for one type of algorithm, AMSR-E sea-ice concentration is linearly related to the MODIS ice-surface fraction, as we hypothesized in the introduction; i.e., AMSR-E sea-ice concentrations are sensitive to the melt-pond fraction. These are the 6H, Bootstrap_f, and Bristol algorithms. For the other algorithms investigated, such a linear relationship is increasingly less pronounced in the following descending order: NASA Team, Near90_lin, Bootstrap_p.
To explain the different sensitivities to the melt pond fraction (Sect. 3.2),
we start with an illustration of the distribution of AMSR-E brightness
temperatures and contemporary MODIS ice-surface fractions of July 2009 in the
parameter spaces of four of the algorithms (Fig. 6). These algorithms are
NASA Team, ASI, or Near90_lin, as both rely on brightness temperatures near
90 A wintertime AMSR-E brightness-temperature distribution is shown for open
water (black dots) and AMSR-E NT2 sea-ice concentration Wintertime open-water (white cross) and sea-ice (black crosses)
tie points obtained from Ivanova et al. (2015) and used to compute the AMSR-E
sea-ice concentration (see Sect. 2.3 for an explanation of why we use winter
tie points) are shown. Red arrows denote the direction of increasing sea-ice concentration. AMSR-E brightness temperatures of our data set, i.e., only for MODIS
sea-ice concentration A red line connecting FYI and MYI tie points denotes the ice line.
For the NASA Team algorithm (Fig. 6a), summer data points from July 2009 are
located well within the cloud of winter data points (see I). The NASA Team
tie-point triangle (Cavalieri et al., 1990) is approximated by the dashed
white lines and the red (ice) line (see V). Many summer data points are
located to the left of the ice line. For these data points, NASA Team sea-ice
concentrations exceed 100 %, and MODIS ice-surface fractions are between
80 and 100 % (see the color scale). To the right of the ice line, summer
data points coincide with MODIS ice-surface fractions of
For the ASI or Near90_lin algorithm (Fig. 6b) summer data points from July
2009 are also located well within the cloud of winter data points. A
considerable number of the summer data points are located above the ice line.
For these data points, ASI or Near90_lin sea-ice concentrations exceed
100 %. Most of the summer data points located below the ice line
correspond to ASI or Near90_lin sea-ice concentrations between 80 and
100 %. The associated MODIS ice-surface fractions decrease from
For the Bootstrap_f algorithm (Fig. 6c), a substantial number of summer data points from July 2009 fall outside the winter data-point cloud. The majority of the summer data points are located above the winter ice line (red: our tie points, black: Comiso et al., 1997). The locations of these data points relative to the open-water tie point, the winter ice lines, and the tie points for MYI and FYI suggests that Bootstrap_f sea-ice concentrations exceed 100 % by up to a few tens of percent (compare Fig. 5b). The distance between the open-water tie point and the winter ice lines increases from left (MYI tie point) to right (FYI tie point). Similar MODIS ice-surface fractions tend to intersect the winter ice lines. Therefore, the overestimation of Bootstrap_f sea-ice concentration decreases with decreasing MODIS ice-surface fraction (see also Fig. 5b).
For the Bootstrap_p algorithm (Fig. 6d), only few summer data points from
July 2009 are located closely above the winter ice lines (see also
Sect. 4.1.3). Consequently, Bootstrap_p sea-ice concentrations do not
exceed 110 % (compare Fig. 5e). Similar to the Bootstrap_f algorithm
(Fig. 6c) only very few summer data points are located close to the MYI tie
point. The majority of those data points which are associated with MODIS
ice-surface fractions
Top row: winter tie points for first-year ice (FYI) and multiyear ice (MYI) expressed as normalized brightness-temperature polarization difference (PR); other rows: summer tie points derived as outlined in the text expressed as PR and brightness temperature (TB) at vertical (TBV) and horizontal (TBH) polarization. Brightness temperatures are given together with 1 standard deviation.
We used open-water and sea-ice tie points representative of winter conditions (Sect. 2.3). We are not aware of summer sea-ice tie points for the ASI or Near90_lin and the NASA Team algorithms, but they do exist for the bootstrap algorithm. The solid cyan line in Fig. 6c denotes the summer sea-ice tie point for the Bootstrap_f algorithm taken from Comiso et al. (1997). For the Bootstrap_p algorithm (Fig. 6d), the solid and dashed cyan lines denote the summer sea-ice tie points for the periods 1–18 July and 19 July–4 August, respectively. For the period after 4 August, the summer ice line coincides with the winter ice line (black line in Fig. 6d).
We use MODIS ice-surface fractions of the period 20 June to 5 July to derive
summer tie points from our summer brightness temperatures. We only select
data of MODIS ice-surface fractions
The potential impact of using summer instead of winter sea-ice tie points will be discussed in the following subsection.
Background: AMSR-E brightness-temperature frequency difference
(gradient ratio, GR) at 37 and 19
During the melting season, changes in the snow and sea-ice microphysical properties, the associated variations in AMSR-E brightness temperatures, and the retrieved AMSR-E sea-ice concentrations can occur within a few days. It is likely that Figs. 4–6 do blur such temporal variations, which, we think, need to be discussed to understand the observed differences in the sensitivity of the AMSR-E sea-ice concentration algorithms to the melt-pond fraction. Therefore we subdivide the MODIS and AMSR-E data sets used into pentads and discuss the temporal evolution for the four algorithms shown in Fig. 6.
For the NASA Team algorithm, in the first pentad (1–5 June, Fig. 7a), most
summer data points are located at PR19
NASA Team sea-ice concentrations exceed 100 % on 16–20 June and
especially 1–5 July (Fig. 7b, c) with values up to 120 %. We find only
few values
For the Bootstrap_f algorithm, in the first pentad (1–5 June, Fig. 8a), most
summer data points are associated with MODIS ice-surface fractions
The good agreement between Bootstrap_f sea-ice concentration and MODIS
ice-surface fraction in the first pentad of June breaks down during June and
re-emerges during July. Between the third pentad of July and the second
pentad of August, average correlations are
Background: vertically polarized AMSR-E brightness temperature at
19
It is difficult to quantify how this result would change by using summer
sea-ice tie points, which we did not use to compute AMSR-E sea-ice
concentrations with the two bootstrap algorithms for the reasons given in
Sect. 2.3, but did include in Fig. 6c, d as cyan lines. The distance between
the cyan line and the winter ice lines in proximity to the FYI tie point,
measured along the dashed white line (Fig. 6c), suggests that we would reduce
Bootstrap_f sea-ice concentrations by 10–15 %. Therefore, on the FYI
side of the parameter space, Bootstrap_f sea-ice concentrations would be
The temporal evolution of Bootstrap_p sea-ice concentrations in relation
to the MODIS ice-surface fraction during June is similar to the
Bootstrap_f algorithm (Fig. S1a, b in the Supplement). One principal
difference is the smaller slope we obtain with Bootstrap_p sea-ice
concentrations compared to the Bootstrap_f algorithm:
How does this result change if we use summer sea-ice tie points (Fig. 6d,
cyan lines, and Sect. 4.3.2)? The early summer ice line (Fig. 6d, solid cyan
line) is steeper than the winter ice lines, and intersects them close to the
FYI tie point. Therefore, close to the FYI tie point and to the right, all
summer data points are below the summer ice line, causing Bootstrap_p
sea-ice concentrations
The temporal development of brightness temperatures, sea-ice concentrations,
and ice-surface fractions obtained with the Near90_lin algorithm
(Figs. S2, 6b) is comparable to that obtained with the Bootstrap_p
algorithm (Fig. S1). The scatter in summer data points and the scatter
between Near90_lin sea-ice concentrations and MODIS ice-surface fractions
is a little less pronounced and peaks earlier. We attribute the scatter again
to snow-property variations (Tables 1, 2, 3). For the smaller electromagnetic
wavelength at 89
With respect to the 6H algorithm and the Bristol algorithm, we state that both algorithms reveal a temporal development of slopes and correlations between AMSR-E sea-ice concentrations and MODIS ice-surface fractions (Figs. S3, S4, Tables 8, 9, 10), which are similar to the Bootstrap_f algorithm. Both algorithms, 6H more than Bristol, are sensitive to melt ponds.
A MODIS ice-surface fraction value of 60 % can, in reality, be anything
between case A, 100 % sea ice with 40 % melt-pond fraction, and
case B, 60 % sea ice with 0 % melt-pond fraction, as laid out in the
Introduction. Slopes between the AMSR-E sea-ice concentration and the MODIS
ice-surface fraction obtained, for example, for the NASA Team algorithm, of
1.31 (Fig. 7f) would convert 60 % MODIS ice-surface fraction into
78 % NASA Team sea-ice concentration. In case B this would be an
Slope of the linear relationship and correlation between AMSR-E
sea-ice concentrations and MODIS ice-surface fractions for the six
algorithms, which do not cut off sea-ice concentrations, averaged over the
six pentads from 11–15 July to 6–10 August. For each algorithm, the average
value
We compute the average slope and correlation values of all algorithms, except ASI and NT2, for the six pentads 11–15 July to 6–10 August together with a resulting over- or underestimation of the actual sea-ice concentrations of cases A and B for which we chose ice-surface fractions of 60 and 80 %. The Bootstrap_f algorithm is most sensitive to melt ponds (highest slope), followed by the Bristol and 6H algorithms (Table 12). The Bootstrap_p algorithm is least sensitive to melt ponds (lowest slope), followed by the NASA Team algorithm. This sensitivity is most pronounced for the Bristol algorithm (largest correlation), followed by the Bootstrap_f algorithm. The sensitivity is least pronounced for the Near90_lin algorithm (smallest correlation), followed by the Bootstrap_p algorithm. Most pronounced means that snow and sea-ice property variations as well as the weather influence have a comparably small influence. These variations have a larger influence on AMSR-E sea-ice concentrations retrieved with an algorithm with a less pronounced sensitivity to melt ponds. The algorithms with the largest sensitivity to melt ponds interestingly provide the smallest underestimation of the concentration of melt-pond-covered sea ice and the largest overestimation of the concentration of non-ponded sea ice (e.g., the Bootstrap_f and Bristol algorithms, Table 12). The algorithms with the smallest sensitivity to melt ponds provide the largest underestimation of the concentration of melt-pond-covered sea ice and the smallest overestimation of the concentration of non-ponded sea ice (e.g., Bootstrap_p, Table 12).
Using summer sea-ice tie points for the Bootstrap_f algorithm would presumably reduce the mean slope as discussed in Sect. 4.3.2, leading to a smaller under- and overestimation of the sea-ice concentrations of cases A and B, compared to Table 12. Using the midsummer tie point for the Bootstrap_p algorithm would, in contrast, presumably increase the mean slope as discussed in Sect. 4.3.3, leading to a larger under- and overestimation of the sea-ice concentrations of cases A and B, respectively, compared to Table 12.
We investigate the sensitivity to melt ponds of eight sea-ice concentration retrieval algorithms based on satellite microwave brightness temperatures, by comparing contemporary daily estimates of sea-ice concentration and melt-pond fraction. We derive gridded daily sea-ice concentrations from the Advanced Microwave Scanning Radiometer aboard the Earth Observation Satellite (AMSR-E) brightness temperatures of June–August 2009. We use a consistent set of tie points to aid intercomparison of the algorithms. We derive the gridded daily fraction of melt ponds, open water between ice floes, and ice-surface fraction from contemporary Moderate Resolution Spectroradiometer (MODIS) reflectance measurements with a neural network based classification approach. We discuss potential uncertainty sources of these data, and conclude that MODIS ice-surface fractions are as accurate as 5–10 %. We carry out the comparison of AMSR-E and MODIS data sets at 100 km grid resolution.
AMSR-E sea-ice concentrations agree fairly well with MODIS sea-ice concentrations, the sum of the ice-surface fraction, and the melt-pond fraction, with slopes of a linear regression between 0.90 and 1.16. However, for some algorithms, AMSR-E sea-ice concentrations scatter widely for MODIS sea-ice concentrations larger than 80 %. We note that the eventual overestimation of the concentration of the sea ice in between the melt ponds, to produce seemingly “correct” sea-ice concentrations that include the melt ponds, will result in incorrectly overestimating the concentration of sea ice in areas with real open water.
We isolate the influence of melt ponds by only comparing AMSR-E sea-ice concentrations with MODIS ice-surface fractions for grid cells with MODIS sea-ice concentrations above 90 %. By doing so, we can use the ice-surface fraction instead of the melt-pond fraction as a measure of the impact of melt ponds and can keep the effect of potential misclassification between the two spectrally close surface types, open water and melt ponds, as small as possible. For most of June, we find a nonlinear relation between both data sets. We attribute this to the influence of snow-property variations impacting the microwave brightness temperatures and a melt-pond fraction that is still small. After June, for one group of algorithms, e.g., the Bristol and Comiso Bootstrap frequency mode (Bootstrap_f) algorithms, sea-ice concentrations are linearly related to MODIS ice-surface fractions. For other algorithms, e.g., Near90GHz and Comiso Bootstrap polarization mode (Bootstrap_p), the linear relationship is weaker and develops later in summer.
We take the degree of correlation between AMSR-E sea-ice concentration and
MODIS ice-surface fraction as a measure of an algorithm's sensitivity to the
melt ponds, and use the obtained linear regression slope to estimate
differences between actual and retrieved sea-ice concentration. All
algorithms underestimate the sea-ice concentration of 100 % sea ice with
an open-water fraction of 40 % due to melt ponds (case A) by between
14 % (Bootstrap_f) and 26 % (Bootstrap_p). The underestimation
reduces to 0 % for a melt-pond fraction of
One next step would be to extend the analysis to more years to confirm the results of our case study with a larger number of data. Currently, at pentad scale, the number of data is too small to use a higher MODIS sea-ice concentration threshold of, e.g., 98 % to better isolate the influence of melt ponds. Based on a substantially smaller number of data, using a threshold of 98 % at monthly instead of a pentad scale, we find that for the month of July, the correlation between AMSR-E sea-ice concentrations and MODIS ice-surface fraction increases from 0.86 to 0.92 (Bootstrap_f), from 0.85 to 0.91 (Bristol), and from 0.67 to 0.76 (NASA Team), while the slopes of the linear regression remain similar.
For reasons outlined in the description of the algorithms, we use a consistent set of sea-ice tie points derived for winter conditions. By applying published summer sea-ice tie points for the bootstrap algorithms, we find that the slopes of the linear regression would be reduced for Bootstrap_f but not for Bootstrap_p. As a result, Bootstrap_f would underestimate sea-ice concentrations for case A less, but overestimate sea-ice concentrations for case B more.
We suggest that algorithms that are more sensitive to melt ponds could be easily optimized further because the influence of snow and sea-ice surface property variations, of which distribution is unknown, seems to be less pronounced, while methods to derive melt-pond fraction, which would be needed for the optimization, have been developed. According to our results, this applies to the Bootstrap_f, Bristol, and Near90_lin algorithms, and the CalVal algorithm, which is similar to the Bootstrap_f mode and is used in the SICCI algorithm. The Bootstrap_p and NASA Team algorithms seem to be less suitable for further optimization. While these seem to have the lowest sensitivity to melt ponds, and therefore lowest underestimation for case A, they seem to overestimate the sea-ice concentration for case B most among the algorithms investigated.
The following data are used in the present paper:
AMSR-E/Aqua L2A global swath spatially resampled brightness temperatures
data set, version 3:
MODIS surface reflectance daily L2G global
500 m and 1 km product (MOD09GA: NT2 sea ice concentration
from the AMSR-E/AQUA daily L3 12.5 km brightness temperature, sea ice
concentration and snow depth polar grids product, version 3
( EASE Grid sea ice age, NSIDC, DAAC,
The daily MODIS melt pond fraction data set co-located with AMSR-E brightness temperatures and sea ice concentrations is owned by the ESA SICCI project and is available upon request by sending an email to stefan.kern@uni-hamburg.de or icdc.cen@lists.uni-hamburg.de.
The work leading to this paper was funded by ESA/ESRIN via the Climate Change Initiative under the SICCI project for all authors but Stefan Kern. Stefan Kern acknowledges support from the Center of Excellence for Climate System Analysis and Prediction (CliSAP), University of Hamburg, Germany. We thank the named and the anonymous reviewers for very constructive and helpful comments which helped to substantially improve the manuscript. Edited by: E. Hanna Reviewed by: G. Flato, G. Heygster, and one anonymous referee