impact of surface conditions on thin sea ice concentration estimate from passive microwave...

15
Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations Mohammed Shokr a, , Lars Kaleschke b a Science and Technology Branch, Environment Canada, 4905 Dufferin St., Toronto, Ont. Canada, M3H 5T4 b Institute of Oceanography, University of Hamburg, Bundesstrasse 53, D-20146 Hamburg, Germany abstract article info Article history: Received 6 July 2011 Received in revised form 5 January 2012 Accepted 8 January 2012 Available online 17 February 2012 Keywords: Sea ice concentration Passive microwave Ice algorithms Thin ice Snow on sea ice Ice concentration retrieved from spaceborne passive microwave observations is a prime input to operational sea ice monitoring programs, numerical weather prediction and global climate models. However, it is usually underestimated by existing algorithms due to surface conditions, especially in case of young ice types. Eval- uation of those algorithms identies errors in concentration estimates but does not necessarily link them to the adverse surface conditions. The present study is an attempt to establish those links for young ice (b 25 cm) thick. It uses measurements of microwave emission from articially grown sea ice in an outdoor tank and calculates ice concentration using ve established algorithms: NT, Bootstrap (BSA), NT2, ASI and ECICE. Since the actual concentration is known (100%), then any deviation from this value is considered an error and can be linked to the observed surface conditions, which are usually caused by weather events. Those conditions were acquired on hourly or daily basis. Results identify key conditions that lead to under- estimation of ice concentration. They include surface refreezing, slush, snow settling following fresh snowfall, and falling precipitation in different forms. The study shows also that NT and NT2 are most affected by surface processes while BSA performs better. ASI is much less affected because it uses the high frequency channel (e.g. SSM/I 85 GHz), which is sensitive only to processes within the top snow layer. ECICE, with its probabi- listic and ensemble approach shows also good results under most surface conditions. Dry or wet snow does not lead to signicant difference in ice concentration estimate. The study also aims at validation of ECICE. © 2012 Elsevier Inc. All rights reserved. 1. Introduction One of the most important sea ice parameters from marine navi- gational, weather prediction and climatic viewpoints is ice concentra- tion. Passive microwave sensors are most suitable for retrieving this parameter because of their ability to penetrate both cloud cover and polar darkness. Commonly used algorithms for sea ice concentra- tion retrieval include NASA Team (NT) (Cavalieri et al., 1984), The Bootstrap algorithm (BSA) (Comiso & Sullivan, 1986), Enhanced NASA Team (NT2) (Markus & Cavalieri, 2000), and ARTIST Sea Ice (ASI) (Kaleschke et al., 2001). A more recent algorithm, called Environment Canada's Ice Concentration Extractor (ECICE), has been developed to determine total and ice type concentration (Shokr et al., 2008). All algorithms employ brightness temperature (Tb) for a given frequency f and polarization p (can be horizontal hor vertical v) and/or the derived parameters of polarization ratio (PR), and the gradient ratio (GR) which are dened as follows: PR f ¼ Tb fv Tb fh Tb fv þ Tb fh and GR f 1 pf 2 p ¼ Tb f 1 p Tb f 2 p Tb f 1 p þ Tb f 2 p Retrieval of concentration of thin ice, which is dened in this study as ice less than 25 cm thick, from microwave data is particularly difcult for two reasons. First, large uctuations of Tb and hence PR usually characterize the transition from open water to ice as well as the unsettled thin ice surface conditions especially during snowfall or freezing rain (Shokr et al., 2009). Since lower frequency channels (e.g. 19 GHz) are associated with larger penetration depth and hence larger uctuations, then their use (which is the case in most algorithms) will cause wrong estimate of ice concentration. Secondly, Microwave emission from snow-covered sea ice is mainly affected by the subsurface composition and properties rather than the bulk ice properties. Consequently, the retrieved ice concen- tration may not be correct if surface conditions produce uncharacter- istic values of Tb. This is more likely to happen in case of thin ice because it exhibits a wider range of surface properties and processes that varies signicantly at small spatial and temporal scales. The rea- son for this variation is twofold: (1) the steep temperature gradient Remote Sensing of Environment 121 (2012) 3650 Corresponding author. Tel.: + 1 416 739 4906; fax: + 1 416 739 4221. E-mail address: [email protected] (M. Shokr). 0034-4257/$ see front matter © 2012 Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2012.01.005 Contents lists available at SciVerse ScienceDirect Remote Sensing of Environment journal homepage: www.elsevier.com/locate/rse

Upload: mohammed-shokr

Post on 28-Oct-2016

214 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations

Remote Sensing of Environment 121 (2012) 36–50

Contents lists available at SciVerse ScienceDirect

Remote Sensing of Environment

j ourna l homepage: www.e lsev ie r .com/ locate / rse

Impact of surface conditions on thin sea ice concentration estimate from passivemicrowave observations

Mohammed Shokr a,⁎, Lars Kaleschke b

a Science and Technology Branch, Environment Canada, 4905 Dufferin St., Toronto, Ont. Canada, M3H 5T4b Institute of Oceanography, University of Hamburg, Bundesstrasse 53, D-20146 Hamburg, Germany

⁎ Corresponding author. Tel.: +1 416 739 4906; fax:E-mail address: [email protected] (M. Sho

0034-4257/$ – see front matter © 2012 Elsevier Inc. Alldoi:10.1016/j.rse.2012.01.005

a b s t r a c t

a r t i c l e i n f o

Article history:Received 6 July 2011Received in revised form 5 January 2012Accepted 8 January 2012Available online 17 February 2012

Keywords:Sea ice concentrationPassive microwaveIce algorithmsThin iceSnow on sea ice

Ice concentration retrieved from spaceborne passive microwave observations is a prime input to operationalsea ice monitoring programs, numerical weather prediction and global climate models. However, it is usuallyunderestimated by existing algorithms due to surface conditions, especially in case of young ice types. Eval-uation of those algorithms identifies errors in concentration estimates but does not necessarily link them tothe adverse surface conditions. The present study is an attempt to establish those links for young ice(b25 cm) thick. It uses measurements of microwave emission from artificially grown sea ice in an outdoortank and calculates ice concentration using five established algorithms: NT, Bootstrap (BSA), NT2, ASI andECICE. Since the actual concentration is known (100%), then any deviation from this value is considered anerror and can be linked to the observed surface conditions, which are usually caused by weather events.Those conditions were acquired on hourly or daily basis. Results identify key conditions that lead to under-estimation of ice concentration. They include surface refreezing, slush, snow settling following fresh snowfall,and falling precipitation in different forms. The study shows also that NT and NT2 are most affected by surfaceprocesses while BSA performs better. ASI is much less affected because it uses the high frequency channel(e.g. SSM/I 85 GHz), which is sensitive only to processes within the top snow layer. ECICE, with its probabi-listic and ensemble approach shows also good results under most surface conditions. Dry or wet snow doesnot lead to significant difference in ice concentration estimate. The study also aims at validation of ECICE.

© 2012 Elsevier Inc. All rights reserved.

1. Introduction

One of the most important sea ice parameters from marine navi-gational, weather prediction and climatic viewpoints is ice concentra-tion. Passive microwave sensors are most suitable for retrievingthis parameter because of their ability to penetrate both cloud coverand polar darkness. Commonly used algorithms for sea ice concentra-tion retrieval include NASA Team (NT) (Cavalieri et al., 1984), TheBootstrap algorithm (BSA) (Comiso & Sullivan, 1986), EnhancedNASA Team (NT2) (Markus & Cavalieri, 2000), and ARTIST Sea Ice(ASI) (Kaleschke et al., 2001). A more recent algorithm, calledEnvironment Canada's Ice Concentration Extractor (ECICE), has beendeveloped to determine total and ice type concentration (Shokret al., 2008). All algorithms employ brightness temperature (Tb)for a given frequency f and polarization p (can be horizontal “h” orvertical “v”) and/or the derived parameters of polarization ratio(PR), and the gradient ratio (GR) which are defined as follows:

PRf ¼Tbfv−TbfhTbfv þ Tbfh

+1 416 739 4221.kr).

rights reserved.

and

GRf1pf2p¼ Tbf1p−Tbf 2p

Tbf1p þ Tbf2p

Retrieval of concentration of thin ice, which is defined in thisstudy as ice less than 25 cm thick, frommicrowave data is particularlydifficult for two reasons. First, large fluctuations of Tb and hence PRusually characterize the transition from open water to ice as well asthe unsettled thin ice surface conditions especially during snowfallor freezing rain (Shokr et al., 2009). Since lower frequency channels(e.g. 19 GHz) are associated with larger penetration depth andhence larger fluctuations, then their use (which is the case in mostalgorithms) will cause wrong estimate of ice concentration.

Secondly, Microwave emission from snow-covered sea ice ismainly affected by the subsurface composition and properties ratherthan the bulk ice properties. Consequently, the retrieved ice concen-tration may not be correct if surface conditions produce uncharacter-istic values of Tb. This is more likely to happen in case of thin icebecause it exhibits a wider range of surface properties and processesthat varies significantly at small spatial and temporal scales. The rea-son for this variation is twofold: (1) the steep temperature gradient

Page 2: Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations

37M. Shokr, L. Kaleschke / Remote Sensing of Environment 121 (2012) 36–50

within the ice depth, and (2) the recurrent sharp variation of atmo-spheric temperature and precipitation that usually prevail duringthe early ice formation period. These factors are responsible for afew surface processes that affect microwave emission as describedin the following.

Thin sea ice surface may become wet or flooded if submergedunder heavy snow fall. This leads to underestimation of ice concentra-tion. Brine volume fraction may vary rapidly within the ice subsurfacelayer due to freeze up or melting of brine pockets in response to sharpvariation of atmospheric temperature (Weeks & Ackley, 1986). Sinceliquid brine is highly lossy medium, Tb decreases with increasingbrine volume. As the bulk ice temperature decreases, liquid brineinside brine pockets starts to freeze around the pocket's wall yielding10% greater volume. This increases the pressure inside the pocket andcauses brine expulsion downward to warmer temperature and up-ward to the surface if ice is very thin (Tucker et al., 1992). The latercreates a liquid brine film at the surface, which consequently causesa sharp drop in Tb (both horizontally and vertically polarized), withan increasing PR (Shokr et al., 2009; Wensnahan et al., 1993).

Frost flowers may also be formed at the thin ice surface as a resultof water vapor release from the ice subsurface under very cold atmo-spheric temperature (Martin et al., 1996). Once formed, they usuallywick up brine from the surface, hence causing Tb to decrease, but ren-der PR at the small value typical of those from plain ice surface(Grenfell & Perovich, 1994) (i.e. it does not depolarizes the signalany more than the depolarization caused by the plain ice surface).Therefore, this is not an adverse condition for the ice concentrationalgorithms that depend on PR.

Snow on sea ice also complicates the surface composition andhence the retrieval of ice concentration, especially in relativelywarm regions or on thin ice. However, it remains poorly studied, es-pecially from the microwave emission standpoint. An instructive re-view of geophysical, dielectrical and thermal properties of snow onsea ice is presented in Langlois and Barber (2007). More recent stud-ies on snow on first-year sea ice are reported in Langlois et al. (2008)and Yackel and Barber (2007). Dry snow is not a lossy medium so itallows more emission from the underlying ice to penetrate but italso causes more scattering. The net effect is that fresh snowfallcauses a decrease in Tb for both polarizations, especially for lower fre-quencies (Grenfell & Comiso, 1986; Perovich et al., 1998). However,PR remains generally small and independent of snow thickness.Hence the total concentration of ice under dry snow is usually esti-mated correctly by algorithms that employ PR of lower frequencies,especially when the ice is thicker than, say, 15 cm. The situation athigher frequencies (e.g. 85 GHz) is not as confirmed in the literature.While a drop in brightness temperature by 10 K (for both vertical andhorizontal polarization) is reported for snow depth>50 mm due tohigher volume scattering within the snow pack (Kern, 2001), moredrop (by about 5 K0 was observed for the horizontal polarization assoon as dry snow starts to fall on bare ice surface (Shokr et al.,2009). The latter means an increase in PR and therefore an underesti-mation of ice concentration using higher frequency channels.

A few geophysical processes within the snow also affect micro-wave emission. Densification of snowpack increases its permittivity,hence suppresses microwave emission from the underlying sea ice(Pullianen & Hallikainen, 2001). Cycles of melting and refreezingcause changes in snow grain size, shape, and liquid water content inthe snowpack. Microwave brightness temperature increases non-monotonically with snow grain size (Fuhrhop et al., 1998). Therefreezing of liquid water in the snow decreases attenuation and in-creases scattering of microwave emission, leading to an increase inTb (Tonboe et al., 2003). Refreezing snow also leads to the formationof ice layering (within the snowpack) and ice crust (near the surfaceof the snow base). This causes significant scattering and hence adecrease of Tb especially at lower frequencies (≤ 37 GHz) (Comisoet al., 1997). This effect is more pronounced for the horizontal

polarization at high incidence angles due to the lower penetrationdepth and stronger snow layering effects (Hallikainen, 1989, andLanglois & Barber, 2007). As a result, PR increases, which leads to un-derestimation of ice concentration from the algorithms that use lowersatellite microwave frequencies.

When the snow is exposed to a strong temperature gradient, acoarse-grained layer, called hoar layer, is formed at its base. Thegrain size distribution is usually highly mixed, with long dimensionsranging from 2 to 10 mm, and short dimensions ranging from 0.1 to0.3 mm (Sturm et al., 2006). The sensitivity of ice concentrationfrom NT algorithm (uses observations from the 19 GHz and 37 GHzchannels) to the correlation length of the hoar layer grains wasstudied by modeling microwave emission using different values ofcorrelation lengths (Tonboe et al., 2003). Results show almost no sen-sitivity to the grain size but higher sensitivity to the snow density ofupper layers. However, the presence of brine-rich layer at the snow/ice interface increases the dependence of the emitted radiation ongrain size. At higher frequency (90 GHz) the emissivity of the snow-covered sea ice was found to be sensitive to the hoar layer depth(Perovich et al., 1998). There is also evidence for strong decrease ofpolarization ratio if slush is formed within the snow pack (Garrity,1992; Kern, 2001; and Shokr et al., 2009). The effect of slush layerswithin the snow pack is most pronounced in the frequency rangefrom 5 to 35 GHz (Mätzler et al., 1984).

All of the above–mentioned processes produce radiometricallydifferent surfaces for the same thickness-based thin ice category.Their combined influence makes it difficult to retrieve ice concen-tration as well as other ice and snow parameters (for example, nomethod has been developed to estimate the snow–water equivalentover sea ice although it has been developed over land (Mätzler,2006)). A few existing methods of ice concentration attempted toaddress this surface complexity. For example, NT2 accounts for sur-face glaze and layering that are described in Comiso et al. (1997) byusing a threshold of a gradient ratio (Markus & Cavalieri, 2000).ECICE is generic enough so that it can identify any surface by provid-ing the probability distribution of the radiometric observations. Thepresent study is an attempt to identify conditions that permit or pre-clude the estimate of thin ice concentration using five different algo-rithms: NT, BSA, NT2, ASI and ECICE. The approach entails comparisonof the ice concentration results retrieved microwave radiation mea-surements from the surface of artificial sea ice grown in an outdoortank. The ice concentration is always 100%. This is the only valuethat is made in this comparison study. More material on validationof ECICE using natural sea ice in the Gulf of St. Lawrence is presentedin Shokr et al. (2008).

2. Experimental setup

The Ice Tank is a circular outdoor above-ground swimming pool of7.5 m diameter located at the National Research Council facility inOttawa, Canada. It was filled with 28 parts per thousand (ppt) of sa-line water to simulate sea water (Fig. 1). The experiment started onNovember 23, 2005 when freezing began and progressed gradually.It ended on January 10, 2006 when the ice thickness (including layersof snow and frozen slush) reached 24 cm. Hence, ice was consideredthin for the entire period but with a variety of surface conditionsthat included open water (OW), Slush (SL), bare or snow-coveredice with dry or wet snow (SN) and ice with metamorphosed snowthat includes ice0 layering, lenses or crust within or at the top ofthe snow (CR).

A Surface-Based Radiometer (SBR) system (Asmus & Grant, 1999)was used to record passive microwave emission from the ice surfaceevery 5 min. The system is a special purpose data acquisition platformdesigned solely for control and data acquisition using microwave ra-diometers. It allows for a real time graphical display of calibratedbrightness temperature data. The radiometers, manufactured by

Page 3: Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations

Fig. 1. The outdoor Ice Tank with the two radiometers shown in the background.

Table 1Sea ice types and surfaces used in the study.

Surface Acronym

Open water OWSlush SLBare or snow-covered ice (dry or wet) SNIce layering or crust within the snow CRIceb12 cm thickness THNIce>15 cm thickness THKMulti-year ice MYSame as CR C-type

38 M. Shokr, L. Kaleschke / Remote Sensing of Environment 121 (2012) 36–50

Radiometrics Ltd, (Boulder, CO), were installed in weatherproof tem-perature controlled enclosures allowing better instrument stability.They operated at 19 GHz, 37 GHz, and 85 GHz dual (horizontal andvertical) polarization with 6 degree beamwidth horn antennas col-lecting data at 50° incidence angle. The radiometers are self calibrat-ing with a “quick look” real time accuracy of approximately 1 K. Thefootprint on the ice surface was approximately 1 m2 from all chan-nels. Brightness temperature samples were obtained from the foot-print every 5 min.

Ice and snow temperature profiles were sampled every half hourat 1 cm to 4 cm intervals using 34 thermocouples made of chromealuminum, installed on a temperature dowel. Weather data were re-ceived from the Ottawa airport weather station on an hourly basis. Icethickness and snow depth were measured daily along with ice surfacesalinity and snow bulk salinity using an optical refractometer withaccuracy of 0.1 ppt. Visual observations of surface physical conditionswere also recorded once or twice every day. More details on measure-ments, data acquisition system and calibration of the radiometers areincluded in Shokr et al. (2009).

It should be noted, however, that snow grain size was not mea-sured. For many years, the passive microwave snow community hasstruggled with poor snow grain parameterization and its very impor-tant effect on Tb. Grain size is by far the most important parametercontrolling emission from dry snow. But as liquid water or brinestart to form within the snow volume they tend to dominate theemission. In this study, only qualitative description of snow grainsize was obtained. Snow grain size effect on emission from drysnow has been addressed in previous studies (Andreadis et al.,2008; Durand et al., 2009; Josberger et al., 1996; and Li et al., 2006).

Measurements by SBR constitute mainly surface emission withinsignificant contribution of the reflected downwelling atmosphericradiation. The downwelling radiation Tdown was measured every twoor three days from clear sky. Depending on the radiometer's frequen-cy, values ranged between 7 K and to 30 K. Those values are relativelynegligible compared to Tb from ice surface but not from open waterbecause the latter is radiometrically colder surface. The correctionaccounted for the reflected downwelling radiation (Tdown-ref)

Tdown−ref ¼ Tdown 1−εð Þeτ

Where ε is the emissivity of the ice surface (an average value of 0.9was selected) and τ is the atmospheric opacity (in this case it waszero since a surface-based radiometer was used). The correctionwas achieved by subtracting the results from the measured Tb. Theratio of this quantity with respect to Tb was typically 0.06 for OWand 0.01 for ice surface. Hence, only OW emission was corrected forthe reflected downwelling radiation.

Since there was only one thickness-based ice type during the ex-periment (i.e. thin ice), ECICE was used to identify concentrations ofsurface rather than ice types. Four surfaces are defined: open Water

(OW), Slush (SL), bare or snow-covered ice with dry or wet snow(SN), and ice with metamorphosed snow that includes ice layering,lenses or crust (CR). The latter is the same as the C-type identifiedin NT2. Table 1 includes acronyms of those ice and surface types.

3. Description of the algorithms

The five algorithms used in this study: NT, BSA, NT2, ASI and ECICEestimate total ice concentration, which is composed of the major icetypes: thin ice (THN), first-year ice (FY) and multi-year ice (MY). Inaddition, NT estimates concentration of MY ice separately from FYice and NT2 estimates concentration of either THN or C-type ice,which features ice layering or crust within the snow pack, separatelyfrom the summation of FY and MY ice. An underlying theme in all al-gorithms is that each observed radiometric parameter can be linearlydecomposed into components generated from each type within theheterogeneous cover within the footprint (e.g. OW, FY and MY ice).Each component is identified by a pre-determined typical value ofthe relevant radiometric observation (e.g. brightness temperaturefrom 19 GHz horizontal polarization) from each surface, multipliedby the unknown concentration of that surface. The typical radiomet-ric values are called tie points. The set of linear equations that repre-sent the observations is solved deterministically in NT, while a searchtechnique is used in NT2 to assign a concentration to the observedmeasurement such that it is closest to the concentration estimatedusing a radiative transfer model for many different concentrationsand cloud conditions.

Bootstrap algorithm uses two channels of microwave data andhas a flexibility to choose the best two (among the lower frequen-cies of 19 GHz and 37 GHz) for every situation of the footprint com-position. The 37 GHz channel has proven to produce better iceconcentrations. ASI is a combination of an algorithm developed bySvendsen et al. (1987) with a weather filter and validated duringthe Arctic Radiation and Turbulence Interaction Study (ARTIST).The method calculates the total sea ice concentration from polariza-tion difference (P=Tb85−Tb85h) near 90 GHz (or 85 GHz for SSM/I.This difference is small for most ice types and large for open water.Tie points P0 and P1 for open water and 100% ice cover, respectively,can be derived from reference measurements. The additional usageof lower frequency channels for weather filtering is a necessity inthe marginal ice zone. More details can be found in Kaleschke etal. (2001), Kern et al. (2003), Andersen et al. (2007), Spreen et al.(2008), and Maaß and Kaleschke (2010).

ECICE (Shokr et al., 2008) employs sets of probability distributionof the radiometric values for each parameter and each surface, ratherthan a single tie point. It also uses an optimization approach ratherthan a deterministic solution of a set of equations. It allows the useof any set of radiometric observations to estimate the concentrationof any given set of ice types or surfaces. The number of observationsmust be equal to or greater than the number of surfaces (excludingOW). In this study, ECICE was applied using three input parameters:Tbh, TbV and PR. The three parameters were used from each channelseparately: 19 GHz, 37 GHz, and 85 GHz. Hence, results are presented

Page 4: Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations

39M. Shokr, L. Kaleschke / Remote Sensing of Environment 121 (2012) 36–50

from using data from each channel. Instead of calculating partialconcentrations of ice types, concentration of each surface amongthe four surfaces that were observed during the experiment wasperformed: SL, SN, CR and OW. Therefore, it was possible, usingECICE, to assess the performance of each channel in identifying eachsurface and determining its concentration as well as the total iceconcentration.

All algorithms employ one form or another of “weather filter”.In this study, the filter was deactivated since the radiometer wasground-based.

4. Tie point estimation

This section introduces the tie points used in the four algorithms:NT2, BSA, NT and ASI, along with the distributions of radiometric pa-rameters used in ECICE, from which the ensemble of tie points is pro-duced. Data for all tie points were obtained or derived from theemitted radiation measured by the SBR from the simulated sea icein the Tank. The data are presented in Table 2 for the four surfaces:OW, THN ice, thick ice THK, and CR. The THN ice ties points were av-erages of samples obtained from ice thickness between 4 cm and10 cm, and the THK ice points were averages from samples ofthickness>15 cm).

NT was used with a set of tie points in the PR19–GR37V19V space.The algorithm requires tie points of three surfaces: OW, FY andmulti-year (MY) ice. For FY ice, tie points obtained for THK in thisstudy were used. For MY ice same tie points presented in Cavalieriet al. (1984) were used. To accommodate the ice type notations inNT, all ice types in the Tank were grouped under the category of FYice. The inclusion of MY ice in NT algorithm causes errors as will bediscussed in Section 5.4. An attempt was made to use tie points forslush instead of MY ice (hence allow NT algorithm to define slush sur-face instead of MY ice surface). This produced very large errors in iceconcentration estimates. The reason was that the microwave radio-metric values from slush are located almost on the line connectedthe values from OW and FY ice in the PR19–GR37V19V space. Becauseof this linear trend only the fractions OW and FY ice (but not slush)could be retrieved.

NT2 was applied using a set of tie points for OW, THK ice, THN ice,and CR surface. These correspond to tie points from OW, FY ice, THNice and C-type ice, respectively, in the original application of thealgorithm to the Arctic sea ice data (Markus & Cavalieri, 2000).Since clouds did not directly affect the Tb observations recorded bythe SBR system, the correction for the 85 GHz observations was notnecessary. This was done in the original algorithm using the atmo-spheric lookup table generated from a radiative transfer model fordifferent cloud conditions. The absence of atmospheric influencesmakes the Ice Tank data suitable to assess the utility of the 85 GHzdata in determining ice concentration under different ice surfaceconditions since complications due to atmospheric influences areeliminated.

ASI requires tie points of polarization difference in 85 GHz for OWand ice. The original values, denoted P0 and P1 respectively, are 47.0 Kand 7.5 K (Kaleschke et al. (2001). Ice Tank data generated P0=76.0 Kand P1=9.8 K and 24.2 K for ice less than 6 cm and greater than

Table 2Tie points used in the tested algorithms for the given surfaces that were encountered duringpoints are presented in brightness temperature Tb (K), and polarization ratio PR.

Type Source Tb19h Tb19v PR19 Tb3

OW Tank 100.7 174.3 0.268 121THN ice Tank 205.2 253.6 0.105 218THK ice Tank 222.6 257.5 0.073 230CR Tank 182.8 245.1 0.146 194

16 cm thickness, respectively. The latter value of P1 was selected be-cause it produced more accurate results although it underestimatedthe concentration when the ice was very thin (b6 cm thickness).

As mentioned, ECICE was applied to identify and determine theconcentration of four radiometrically distinctive surface types: OW,SL, SN and CR. These categories are different from the categoriesused in NT and NT, which are based on ice age. Samples from eachsurface type were selected to establish the distribution of Tb and po-larization ratios. The selection satisfied three criteria: (1) the sampleshould be obtained from a vertically homogeneous snow profile; i.e.it cannot be obtained from surface that features snow on top ofslush, (2) the ice thickness must be greater than 8 cm in order toguarantee that the low frequency channel is not affected by the un-derlying water, and (3) periods during or immediately after rain,snow or freezing rain fall should be avoided.

Fig. 2 shows the distributions of Tb (vertical and horizontal) aswell as polarization ratio from 19 GHz data for the four surfaces. Asexpected, OW is well separated from the three other surfaces. SLand SN surfaces are also separated from each other. However, CR sur-face overlaps with both of them and this causes error in identifyingthe CR surface as explained in Section 5.4. The corresponding distri-butions from the 37 GHz (not shown) are similar to those presentedin Fig. 2. But the distributions from the 85 GHz data show more fluc-tuations for the three ice surfaces. This indicates the sensitivity of thischannel to surface processes in response to variation of atmospherictemperature.

It should be recalled that the tie points for the C-ice type in NT2 isthe same as the mean value of the distribution of CR samples in Fig. 2.Similarly, the tie points for the FY ice in the four algorithms (otherthan EICE) are the same as the mean values of the SN distribution.

5. Results and discussions

5.1. Total ice concentration

Comparison between total ice concentration estimates from thefour algorithms is shown in Figs. 3–6. The horizontal axis showsthe day of the year, hour and minute in the format (ddd:hh.mm).The top graph includes the evolution of Tb from the six radiometricchannels, annotated with atmospheric temperature, surface salinityand composition as well as selected weather events. The middlepanel includes results from NT2, ASI, BSA, and NT using the Tank tiepoints (Section 4). The bottom panel includes results from ECICEusing the three sets of input data from 19 GHz, 37 GHz and 85 GHzseparately as indicated above. In order to separate the results fromthe different algorithms when they all output 100% ice concentration,the 100% value is displayed as 102%, 104% and 106% for results fromASI, BSA and NT, respectively in the middle panel; and 102% and104% for results from ECICE using 37 GHz and 85 GHz data, respec-tively in the bottom panel.

Fig. 3 includes results from the beginning of the freezing period.Ice thickness was between 2 and 4 cm with slushy surface most ofthe time. The sharp drop in Tb (e.g. observed at 329:07.01) followedby a gradual increase is a manifestation of cyclic pattern of surfacewetness followed by freezing. This cyclic pattern was observed for

the experiment. Data were obtained from the Ice Tank radiation measurements. The tie

7h Tb37v PR37 Tb85h Tb85v PR85

.0 201.8 0.250 160.7 236.7 0.191

.5 255.4 0.078 234.2 258.6 0.05

.6 260.5 0.061 225.5 256.9 0.065

.4 231.0 0.086 201.6 226.1 0.057

Page 5: Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations

Distribution Tb-19v

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

0.50

100 120 140 160 180 200 220 240 260 280 300

Brightness Temp. (K)

Pro

pb

abili

tyP

rop

bab

ility

Pro

pb

abili

ty

OW

SL

SN

CR

Distribution Tb-19h

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

0.50

80 100 120 140 160 180 200 220 240 260 280 300

Brightness Temp. (K)

OW

SL

SN

CR

distribution PR19

0

0.1

0.2

0.3

0.4

0.5

0.6

-0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

PolarizationRatio

OW

SL

SN

CR

Fig. 2. Probability distributions of the 19 GHz radiometric parameters used in ECICE from four surfaces: OW, SL, SN, and CR.

40 M. Shokr, L. Kaleschke / Remote Sensing of Environment 121 (2012) 36–50

iceb6 cm thickness (Shokr et al., 2009). Thin ice surface is usuallycovered with a thin layer of brine formed due to vertical brine expul-sion from the ice subsurface layer as the ice temperature drops

(Weeks & Ackley, 1986). That instigates the drop in Tb. When snowfalls on this surface it melts immediately, causing a significant de-crease in the salinity of the brine layer at the surface. The decrease

Page 6: Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations

Fig. 3. Evolution of Tb and calculated total ice concentration from the tested algorithms for the period November 24 to 29, 2005. Upward arrows indicate the time associated withthe information next to the arrow.

41M. Shokr, L. Kaleschke / Remote Sensing of Environment 121 (2012) 36–50

of salinity is associated with an increase of the freezing temperature,which speeds up ice formation and subsequently a gradual increase ofTb.

The figure shows that ice concentration estimates from all algo-rithms is affected by this cyclic pattern. The NT2, ASI, BSA and NT al-gorithms produce concentration values much less than 100%, withnoticeable improvement from AS and BSAI. ECICE, on the otherhand, produces the expected 100% concentration especially fromusing the 37 GHz or 85 GHz data. The only exceptions are the shortperiods following the sharp drop in Tb immediately after snow orfreezing rain fall. The recovery from lowest wrong value of ice con-centration that marks the onset of snowfall back to the correct 100%value is faster when using the 37 GHz or 85 GHz data. Data from the

19 GHz channel underestimate the ice concentration as it couldhave reached the underlying simulated sea water due to its largerpenetration depth (about 3 cm).

On November 29, 2005 (day 333) the atmospheric temperaturereached 8.6 °C, causing ice to disintegrate (candled ice). Rain alsostarted and caused sharp fluctuations in Tb.

All methods show ice concentration estimates less than 100% withheavy fluctuations that echo the variation in Tb. Both ASI and ECICEusing 85 GHz data produce more stable results of the correct 100%concentration with short periods of less that 100% concentration.This observation points to the advantage of using 85 GHz data fromspaceborne sensors, provided that a reliable scheme of atmosphericcorrection is developed.

Page 7: Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations

Fig. 4. Evolution of Tb and calculated total ice concentration from the tested algorithms for the period December 6 to 11, 2005. Upward arrows indicate the time associated with theinformation next to the arrow.

42 M. Shokr, L. Kaleschke / Remote Sensing of Environment 121 (2012) 36–50

Fig. 4 covers another period of thin ice (between 3.5 cm and 5 cm)with relatively stable surface conditions until the snowfall event(343.15:25) that dumped 6 cm snow. The temperature was relativelycold ((≅−6.4 °C). This was followed by slushy surface when the tem-perature rose to −2.3 °C on day 244. During the colder period withsnow-free surface NT outputs an average concentration of 79% in-stead of the correct 100%, while NT2 shows better output (averageof 89%). The underestimation of ice concentration from these two al-gorithms can be attributed to the use of the 19 GHz channel with itsrelatively large penetration depth. BSA, which uses 37 GHz data, pro-duced a better average concentration of 93%, while ASI, which usesdata from 89 GHz that has penetration depth ofb1 cm in YI, producedthe correct 100% data most of the time. ECICE output 100% concentra-tion using any channel except for two very short periods when theconcentration from using the 19 GHz data drops. The first period is

marked with a heavy snow squall event that lasted for 30 min(340.17:24) and that caused the concentration to drop to 72%. Inthe second period (341.10:48) the concentration dropped lightlybelow 90% but no observation could be used to explain this drop.

Immediately following the snow fall on day 343, a sharp drop in Tbis observed with a corresponding increase in PR (PR19 rose from 0.03to 0.21 and PR85 from 0.03 to 0.11). The temperature rose gradually to−2.3 °C, and the surface was covered by 5 cm snow that turned into1 cm slush. During this period, which featured another minor snow-fall event (345.04:47), all algorithms underestimated ice concentra-tion with heavy fluctuations. This was caused by the continuoussnow metamorphism as it was settling on the ice surface and thenturning to slush. NT and NT2 show the lowest concentration values(10%–40%) while concentrations from ASI and BSA fluctuate around60%. Results from ECICE are better (reach the correct 100%) but also

Page 8: Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations

Fig. 5. Evolution of Tb and calculated total ice concentration from the tested algorithms for the period December 13 to 18, 2005. Upward arrows indicate the time associated withthe information next to the arrow.

43M. Shokr, L. Kaleschke / Remote Sensing of Environment 121 (2012) 36–50

fluctuate after the minor snowfall indicated above. Results are thesame from using any channel.

Fig. 5 shows results from day 347 to day 352. This period followeda period of almost four days of sub-zero air and surface temperaturewith snow fall on day 343 (see Fig. 4). The data can be grouped intotwo periods. The first is from day 347 to the evening of day 349. It fea-tures snow-free ice surface with thin slushy layer that continued tofreeze as the atmospheric temperature decreased to −12 °C. The icethickness was growing steadily from 9 cm to 12 cm. These conditionsappeared to be ideal for accurate ice concentration estimate from anyalgorithm. However, NT, NT2 and ASI underestimate the ice concen-tration, while BSA produces the much better values, including theexpected 100% concentration. This can be attributed to the apparenthigh values of the polarization difference from all channels (topgraph in Fig. 5). The PR values are higher than their corresponding

tie points for ice (Table 2) and so are the polarization difference,which is used in the ASI algorithm (tie points are presented inSection 4). No visual observations of the surface provided any clueto explain such high values except for the presence of a very thinlayer (a fewmillimeters) of frozen saline-free slush on the ice surface.Surface salinity varied between 11 ppt and 17 ppt. The inaccurate iceconcentration estimates in this case point to the disadvantage ofusing a single tie point to characterize a surface. ECICE, with itsensemble of concentration estimate (2000 simulates) reproducedsuccessfully the 100% concentration.

The second period started on day 349 when a major winter stormbegan and prevailed until 350.15:25, dumping 25 cm of snowwith icepallets and freezing rain. Atmospheric temperature rose from−5.8 °Cto −1.9 °C. NT2, BSA and NT underestimate the concentration withheavy fluctuations as the snow continued to settle on the surface.

Page 9: Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations

Fig. 6. Evolution of Tb and calculated total ice concentration from the tested algorithms for the period December 21 to 26, 2005. Upward arrows indicate the time associated withthe information next to the arrow.

44 M. Shokr, L. Kaleschke / Remote Sensing of Environment 121 (2012) 36–50

The striking observation is that the polarization difference of the85 GHz channel dropped significantly to its typical tie point of ice(Table 2) as soon as the precipitation started to fall on the surface.Hence, the ice concentration estimates from ASI became correct.ECICE, using 85 GHz data, also produces the correct concentration,while it underestimates the concentration at some points, particularlyfrom using the 19 GHz data. That is probably due to the less penetra-tion depth of 85 GHz signal within the snow pack with respect to the19 GHz. (Hallikainen et al., 1986), which makes it less sensitive toprocess in the snowpack. While this demonstrates the advantage ofusing 85 GHz data over the other two channels, it should be notedthat the observations from the surface-based radiometer used inthis study were not affected by atmospheric influences (except forthe minimum influence of the reflected downwelling radiation asmentioned above). In airborne and satellite-borne observations,

correction for such influences on the observed 85 GHz emission isusually neither comprehensive not accurate (Kern, 2001). The use ofthe 85 GHz data from the SBR in this study allowed linking the obser-vations directly to the surface conditions.

Fig. 6 shows results from day 355 to day 360. The ice was 15 cmthick with complex surface that started with snow and frozen slush,followed by wet snow and slush under warm temperature (2.4 °C),and finally with snow that featured very large crystals and ice layer-ing under sub-zero temperatures. Variety of weather events were ob-served, ranging from snowfall, rain, ice palette and freezing drizzle.With these complex weather and surface compositions NT producesconcentrations that deviate significantly at times from the true 100%for short periods. Upon comparison with the top figure of Tb, it canbe seen that the wrong concentration output echoes almost exactlythe evolution of the Tb19h. NT2, when produces better results because

Page 10: Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations

45M. Shokr, L. Kaleschke / Remote Sensing of Environment 121 (2012) 36–50

Tb37h is less sensitive to the snow processes as seen, once again, in thetop figure. BSA produces even better results, while ASI produces theperfect 100% concentration since it uses the 85 GHz data, which areleast sensitive to the snow processes. For the entire period, ECICE pro-duces the expected 100% concentration regardless of the complexsurface characteristics and with using data from any channel. Theonly exception is the concentration drop to 82% at the end of day

Fig. 7. Evolution of Tb and calculated total ice concentration from the tested algorithmsDecember 1st (starting 335.08:46).

360, which resulted from using the 19 GHz data. During that timeTb19h dropped from 260 K to 180 K, with similar drop observedfrom all other channels. It should be noted that a drop in Tb19h andTb37h occurred on days 357 and 359 but did not affect the resultsfrom ECICE using either the 19 GHz or the 37 GHz channels. No corre-sponding drop is observed from the 85 GHz channels during thesetwo events, which means that they were triggered by a mechanism

during the complete ice melt on December 1 and 2, 2005 when rainfall occurred on

Page 11: Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations

Table 3Mean and standard deviation of total ice concentration estimates from the four testedalgorithms for two ice thickness ranges (regardless of the surface conditions).

Algorithm Iceb120 mm Ice>150 mm

Mean (%) stdev. (%) Mean (%) stdev. (%)

NT 41.6 34.0 88.2 19.5NT2 49.4 32.7 98.4 4.7BSA 57.2 28.3 98.8 1.9ASI 62.5 32.7 99.0 3.4ECICE-19 GHz 65.7 37.8 99.0 4.8ECICE-37 GHz 74.1 36.8 99.9 0.8ECICE-85 GHz 77.4 35.1 99.4 2.9

46 M. Shokr, L. Kaleschke / Remote Sensing of Environment 121 (2012) 36–50

different than that which caused the drop in Tb from the other twochannels at end of day 360.

A final note about the effect of rainfall over OW on the estimatedice concentration is worthwhile mentioning. Here the true ice con-centration is 0%. Rain over OW causes an increase of Tb particularlyfrom the horizontal polarization and mainly from the higher frequen-cy channels. This causes a corresponding increase in estimation of iceconcentration, which otherwise should be 0%. Fig. 7 shows Tb and thederived ice concentration using the four tested methods during arainfall event over OW on December 1st, 2005 (day 335). Note thesignificant increase in Tb from Tb37h and Tb85h. The maximum valuesof Tb19h, Tb37h and Tb85h during the peak of the rain are 122.0 K,161.1 K and 239.9 K; respectively. The corresponding PR values are0.190, 0.155, and 0.043. These values are significantly higher thanthe corresponding values from OW under no rain (Table 2); namely0.27, 0.25 and 0.19; respectively. It is obvious then that using the po-larization difference (or ratio), particularly from the 85 GHz channel,will cause significantly high ice concentration (false estimate) in caseof rainfall over OW. This is obvious upon comparing results fromECICE using data from each channel. For the other three algorithms,BSA demonstrates the best performance under OW with rainy condi-tion and ASI is the worst 9since it uses 85 GHz data). Note theGaussian-like distribution of the anomalous Tb as the rain intensifiesthen abates. For practical applications, 85 GHz data should be exclud-ed from ice concentration calculations if they are identified withrainfall.

Table 3 summarizes the accuracy of each algorithm in estimatingtotal ice concentration of thin ice (b120 mm thick) or thick ice(>150 mm thick), regardless of the surface conditions. It is obviousthat the problem of underestimating the 100% ice concentration inthis data set is proven to be in case of thin ice. In this case note thegradual improvement of the performance of the algorithms startingfrom NT and ending with ECICE using 85 GHz data. It should be em-phasized, once again, that the apparent advantage of using 85 GHzdata might be lost when observations are obtained from space sincethis channel is influenced by atmospheric effects, which are not

Table 4Total ice concentration (%) estimates from different algorithms under different surface con

Surface conditions Ice thick(mm)

ECICE

19 GHz

Very thin ice 5 65.6Bare dry thin ice 10–30 79.1Slush—no ice 15–20 74.6Slush (10 mm) 50 77.2Frozen slush 60 90.8Dry surface 40–60 100.0Dry surface: thin fresh-water ice layer at top 100–120 100.0Wet snow 150 100.0Wet snow srfc. flooding 150 89.3Rain on snow on THK ice 150 100.0Crust and layering in snow 150 94.4Candled ice 30 64.7

easy to account for. For thick ice, all algorithms, except NT performwell in general thought not necessarily under the detailed weatherand surface conditions as presented in the above paragraphs.

5.2. Effect of surface conditions on ice concentration estimates

Previous studies confirmed that Tb and PR are affected by snowconditions on sea ice especially during early spring (Grenfell et al.,1994; Langlois & Barber, 2007). Those anomalies impair ice concen-tration estimates although they represent useful information to iden-tify snow processes such as wetness, melt onset and transition frompendular to funicular regimes (Hwang et al., 2007). Table 4 summa-rizes total ice concentration estimates from different algorithmsunder different ice thickness ranges and surface conditions. Thistable forms the basis of the following discussions.

For very thin ice (≤ 5 mm) with bare surface, NT, NT2, BSA andASI return average concentrations 27.2%, 34.0%, 65.7% and 51.4%; re-spectively. ECICE returns higher concentrations, though it was rela-tively low (65.6%) from using the 19 GHz data. The 85 GHz channelwith its 83.4% concentration seems to be more appropriate for thinice because it does not “see” the underlying water (bearing in mindthat no atmospheric correction was needed for the present data).For thicker ice (10–30 mm thick) with bare dry surface, ECICE stilloutperforms the other three algorithms, which provide more-or-lesssame concentration estimate in this case (with better results fromBSA).

During the onset of freezing, 15–20 cm of slush was formed withno solid ice. ECICE produces total concentration between 74.6% and91.1% from using 19 GHz and 85 GHz data, respectively. These are sig-nificantly higher than the results from the other four algorithms(23.5%, 31.9%, 66.4% and 59.7% from NT, NT2, BSA and ASI; respective-ly). The output from BSA and ASI may be viewed as most realisticsince slush is a middle state between OW and solid ice. In the 91.1%total concentration output from ECICE using 85 GHz data, 69.6% is de-fined as slush and the rest is assigned as bare or snow-covered surface(NS) (not shown in Table 4). This is also a reasonable estimate. Datafrom slushy surface (10 mm slush on 50 mm solid ice) also resultsin significantly low ice concentration using NT and NT2 (19.5% and31.1%; respectively). BSA and ASI produce higher values of 61.3%and 48.8%, respectively. ECICE produces highest values (i.e. closestto the true 100% concentration) between 72.2% and 77.2%.

For frozen slush surface on top of 60 mm ice there is noticeable im-provement in the BSA and ASI output (61.2% and 49.9%, respectively)compared to NT and NT2 (22.4% and 33.9%, respectively). Recall thatthe refreezing of slush starts at the surface and propagates downward,so 85 GHz channel is the first to respond to this process.

Dry surface (bare snow or snow cover) is most favorable conditionfor ice concentration retrieval from any algorithm. That is because itdoes not cause change in PR though it causes Tb (in both horizontal

ditions.

NT NT2 BSA ASI

37 GHz 85 GHz

73.8 83.4 27.2 34.0 65.7 51.498.8 82.7 57.3 54.3 52.3 50.188.4 91.1 23.5 31.9 66.4 59.776.2 72.2 19.5 31.1 61.3 48.892.2 99.2 22.4 33.9 71.2 49.9

100.0 100.0 85.8 91.9 92.2 97.8100.0 99.4 66.4 69.8 68.8 70.8100.0 100.0 77.3 94.9 89.3 100.097.6 100.0 87.5 90.1 89.9 99.4

100.0 100.0 94.3 100.0 100.0 100.093.8 98.5 41.8 90.0 92.3 93.474.6 98.3 34.8 55.3 66.6 90.4

Page 12: Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations

Table 5Ratio (in %) of THN or THK ice concentration to total ice concentration from NT2 for dif-ferent ice thickness ranges.

Thick. range(mm)

THN/total THK/total Stdev

b40 4.3 95.7 13.150–60 4.7 95.3 20.090–120 23.8 76.2 39.6150–180 27.9 72.1 33.6>200 10.6 89.4 22.5

47M. Shokr, L. Kaleschke / Remote Sensing of Environment 121 (2012) 36–50

and vertical polarization) to drop due to scattering (Grenfell &Comiso, 1986; Perovich et al., 1998). This is demonstrated in thedata from 40 to 60 mm ice thickness (Table 4). The lowest concentra-tion produced in this case is 85.8% from NT. This observation also ap-plies to thicker ice (>120 mm thick) (not shown in the table). Whenthicker ice (100–120 mm thick) is covered with a layer of refrozenfresh water, ECICE reproduces the expected 100% concentrationwhile NT2, BSA and ASI produce nearly 70%, while NT produces 66.4%.

Snow wetness was identified through visual observations duringthe experiment. For the case of wet snow (150 mm accumulation)on top of 150 mm thick ice, both ASI and ECICE (using any channel)produces the expected 100% concentration, while NT, NT2 and BSAproduce 77.3%, 94.9% and 89.3%; respectively. The use of the 85 GHzdata continues to be advantageous in the case of surface flooding(observed when atmospheric temperature was between 1 °C and3 °C for more than 24 h).

Light rain fell on snow-covered ice (150 mm snow on top of150 mm thick ice) for 5 h on December 25, 2005 while atmospherictemperature was around 0 °C. It caused gradual drop in Tb19h (10 K)and Tb37h (5 K) with no change in Tb from the vertical polarization.This light rainfall did not affect the performance of any algorithm exceptNT. Ice crust and surface layering within the snow pack impair ice con-centration calculations from NT (produces 41.8% concentration). How-ever, when any other algorithm was used it produced much bettervalues (above 90%). Recall that NT2 incorporates a provision to mini-mize the effect surface layering effects.

Finally, candled ice of 30 mm thickness is correctly identifiedusing the 85 GHz data as seen from the results of ECICE and ASI. NT,NT2 and BSA produce concentration values 34.8%, 55.3%, and 66.6%respectively. ECICE results from using 19 GHz and 37 GHz data arestill inaccurate (64.7% and 74.6%, respectively).

A few weather events also affect the estimation of ice concentra-tion. As mentioned in Section 5.1, (Fig. 3), the beginning of the snow-fall on thin ice (b6 cm thick) causes significant drop in concentrationestimates from all algorithms (b20%). ECICE, using 37 GHz or 85 GHzdata, shows full recovery to 100% concentration within 3–10 h afterthe onset of snowfall. BSA also shows full recovery though after a lon-ger period. As mentioned before, rain on OW surface makes all algo-rithms, especially those which use 85 GHz data, misidentify OW asice. Rain over candled ice represents a most difficult situation for iceconcentration retrieval as mentioned in Section 5.1. Snow, ice pellets,and freezing rain on top of the ice surface causes a drop in Tb with asignificant increase in PR from the 19 GHz channel and minimum in-crease from the 85 GHz channel. The two algorithms that produce thecorrect ice concentration in this case are ASI and ECICE using 85 GHzdata. NT2 and ECICE using 19 GHz data underestimate the concentra-tion by 40%–60%, while NT and BSA underestimate it by 60%–80%.Surface flood causes a sudden and very sharp drop in ice concentra-tion using NT or NT2 algorithms. The above forms of precipitationsare observed to have less negative effect on the ice concentrationestimate if the ice is thicker than 15 cm. NT is the most affectedalgorithm under these conditions. Periods of near-zero atmospherictemperature followed by colder temperature cause metamorphismof the snow. NT and NT2 are not able to recover the true 100% ice con-centration in this situation.

5.3. Ice type concentration

As mentioned in Section 4, NT produces the concentration of FYand MY ice. The original tie point of MY (obtained from Arctic data)is used in this study. Since MY ice did not exist in the Tank, the resultscan be used to show the statistics when MY ice was incorrectly iden-tified. In 62% of the samples, NT produces non-zero value of MY ice.The average value of those erroneous estimates is 25.1% with stan-dard deviation of 22.8%. Those values emerged from all surface condi-tions, i.e. no particular condition contribute to the error.

The concentration of MY ice is also derived from the total ice con-centration output from ASI. ASI does not identify MY ice on its own.So, its concentration was determined based on a simple equation pre-sented in Lomax et al. (1995):

Cmy ¼ic 1:613TB85v−0:613Tb85h−194:215ð Þ

248:6−194:215ð Þ

Where ic is total ice concentration from ASI. The results areclipped between 0 and 1.

Application of this equation shows that 21% of the samples containMY ice (i.e. misidentification of that ice type) with an average concen-tration of 40.1% and standard deviation 29.1%.

As mentioned in Section 3, NT2 was applied to estimate partialconcentration of two ice types: THK ice (>150 mm thickness) and ei-ther THN ice (b120 mm thickness) or C-type ice. The decision toswitch between THN and C-type is based on a threshold on thegradient ratio: GR37V19Vb−0.01. Visual observations of the C-typecondition, along with daily measurements of ice thickness, are usedto verify the results. NT2 identified correctly 82.45% of the observedC-type samples. On the other hand, it assigned this type incorrectlyto 8.77% of the observed samples from snow-covered ice only; i.e.no mistaken identification of C-type as slush, bare ice or OW surfaces.This result confirms the usefulness of the gradient parameter to iden-tify the C-type ice.

The ability of NT2 to discriminate between THN and THK types(recall that the tie points for both types were obtained from theTank data) is demonstrated in the results in Table 5. The tableshows the two ratios of thin ice to total ice and thick ice to total ice,for different ice thickness ranges. The standard deviation of thisratio is also shown. Ideally, the THN/total ratio should be 1 for anythicknessb120 mm, and the THK/total ratio should be 100% for anythickness>150 mm. However, the table shows very small THN/totalratio in case of thin ice. This is most likely because the tie point forthe thin ice was obtained from dry ice surface of thickness between40 mm and 100 mm. Hence the poor results in the table confirm thedifficulty of identifying thin ice using NT2 approach because this icetype cannot be represented by a single tie point. It also points to thesensitivity of the microwave radiometric properties of thin ice toweather and surface conditions. The table shows also that identifica-tion of thick ice, on the other hand, is feasible.

5.4. Retrieval of surface conditions using ECICE

In this section, ECICE was used to identify surface conditions. Thequestion of which microwave frequency is most suitable to identifythe four tested surfaces (OW, SL, SN and CR) is addressed. ECICEwas applied to 579, 415, 1960, and 610 samples obtained fromthose surfaces; respectively. None of the samples was used to estab-lish the distributions of the radiometric parameters from each surface(see Section 4). All samples were obtained from ice thicker than 6 cmin order to ensure that Tb was not affected by the underlying water.For each sample the algorithm outputs the concentration of each sur-face type, using one set of observations: Tbh, Tbv and PR from a given

Page 13: Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations

48 M. Shokr, L. Kaleschke / Remote Sensing of Environment 121 (2012) 36–50

frequency. Hence, for a given surface, the results show the best fre-quency to identify the surface.

Results are presented in Fig. 8. Each graph represents results fromapplying ECICE to data from a certain surface using observations froma certain frequency channel (data from the OW samples are notshown). Discrete distributions of the output concentration of eachsurface (OW, SL, SN, and CR) are shown in four concentration ranges:0%–25%, 26%–50%, 51%–75%, and 76%–100%. The vertical axis in eachgraph represents the number of occurrences of the concentrationvalues that falls into the relevant range, divided by the total numberof samples from the relevant surface (in percentage). For example,when the 415 samples from the slush surface were used against the19 GHz observations (top left graph), the percentage of occurrencesof OW, SL, SN and CR concentrations within the concentration range0%–25% becomes 76.8%, 5.5%, 99.5%, and 91.3%; respectively. Thesame percentages for the 76%–100% concentration range are 5%,64.4%, 0% and 0%. Ideally, since each group of sample data wasobtained from a homogeneous surface, ECICE should produce 100%concentration of that surface and 0% of the other surfaces. However,the figure shows some cases closer and others father from this ideal-ized expectation.

Starting with the slush surface (left graphs); the figure shows thatslush surface can best be identified using the 19 GHZ or 37 GHZ ob-servations. Almost 65% of the slush samples are correctly identified(i.e. located within the 76%–100% concentration range) when anyone of these channels are used (top two figures at the left). Slush isslightly confused with OW and CR surfaces in the 19 GHz data. Onthe other hand, 85 GHz data identified only 52% of the slush samplescorrectly (again, judging from the result in the 76%–100% bin). Thischannel confuses SL with CR (nearly 28% of the SL samples produceCR concentration in the range 25%–50%). The figure also shows thatneither 37 GHz nor 85 GHz channels confuse SL with OW, though19 GHz channel slightly does. This conclusion is rather favorable

0102030405060708090

100 SN surface u

OWSLSNCR

0102030405060708090

100 SN surfac

OWSLSNCR

0102030405060708090

100 SL surface using 19 GHz data

OWSLSNCR

0102030405060708090

100 SL surface using 37 GHz data

OWSLSNCR

51% - 75%26% - 50% 76% - 100%0% - 25% 26% - 50%0% - 25%

51% - 75%26% - 50% 76% - 100%0% - 25% 526% - 50%0% - 25%

51% - 75%26% - 50% 76% - 100%0% - 25% 526% - 50%0% - 25%0

102030405060708090

100 SL surface using 85 GHz data

OWSLSNCR

0102030405060708090

100 SN surfac

OWSLSNCR

Fig. 8. ECICE results using input data from three types of homogeneous surface: slush (SL),from each microwave frequencies: 19 GHz, 37 GHz, and 85 GHz are used separately. For eachage number of occurrences of the concentration values with respect to the total number of

because it allows the use of ECICE to retrieve total ice concentrationin presence of slushy surface (but perhaps not in case of floodedsurface).

Moving to the bare dry ice or snow-covered ice surface (SN)(middle graphs); it can also be seen that ECICE can identify thissurface better with 19 GHz or 37 GHz data. The percentage of SN sam-ples that have the concentration of SN surface falling in the range76%–100% is 72%, 73% from these two frequencies; respectively. Thepercentage decreases to 61% when 85 GHz data are used. The outputconcentrations of the other two ice surfaces (SL and CR) are small(falling mostly in the range 0%–25%). While SN surface is not con-fused with OW using any frequency channel, the 85 GHz channel con-fuses it slightly with the slush surface. It is also interesting to notethat SN surface is not confused with the CR surface (which is alsosnow cover but with surface crust or ice layering).

The CR surface is, in fact, the most difficult to identify (rightgraphs). ECICE misidentifies this surface completely from using the19 GHz data (the percentage of occurrences of CR from the CR sam-ples is virtually 0% in the 75%–100% range, where ideally it shouldbe 100%). The algorithm confuses this surface with both SL and SNsurfaces. The situation is better for the 37 GHz and 85 GHz data, asthe percentage of occurrence of CR concentration falling in therange 76%–100% (when CR samples were used) is 58.6% and 53.1%;respectively. It should also be mentioned that while all samplesfrom CR surface produce total ice concentration of nearly 100%when 37 GHz or 85 GHz data are used, only 87% of those samples pro-duce total ice concentration in the range 76%–100% when 19 GHzdata are used (this can be inferred from the 87% OW occurrence inthe 0%–25% range in the top right figure). This confirms the underes-timation of total ice concentration by the 19 GHz channel in presenceof surface crust as reported in (Comiso et al., 1997). Ice layering insnow also causes significant decrease in Tb19h and hence an increasein polarization ratio (Hallikainen, 1989). This observation has been

sing 19 GHz data

e using 85 GHz data

0102030405060708090 CR surface using 19 GHz data OW

SLSNCR

0102030405060708090

100 CR surface using 85 GHz data

OWSLSNCR

51% - 75% 76% - 100% 51% - 75%26% - 50% 76% - 100%0% - 25%

1% - 75% 76% - 100% 51% - 75%26% - 50% 76% - 100%0% - 25%

51% - 75%26% - 50% 76% - 100%0% - 25%1% - 75% 76% - 100%

e using 37 GHz data

0102030405060708090

100 CR surface using 37 GHz data

OWSLSNCR

snow-covered (SN) and crust (CR), with observations of Tbh, Tbv and PR. Observationsconcentration range shown on the horizontal axis, the vertical axis shows the percent-samples from the relevant surface.

Page 14: Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations

49M. Shokr, L. Kaleschke / Remote Sensing of Environment 121 (2012) 36–50

confirmed in the present study. It should be noted also that none offrequencies used confuse CR surface with OW surface.

6. Conclusions

Evaluation of five ice concentration retrieval algorithms: NT, BSA,NT2, ASI and ECICE have been conducted to determine their abilityin identifying concentration of thin ice with different surface condi-tions. The ice was artificially grown in an outdoor experiment, calledthe Ice Tank, to a thickness of 24 cm, subjected to different weatherconditions. The experiment period was from November 23, 2005 toJanuary 10, 2006. The ice concentration in the Tank was 100% thoughthe surface exhibited a variety of conditions: snow-free, snow-covered, slush andmetamorphosed snowwith crust layers and coarsegrains.

Any deviation of the algorithm's estimate from the 100% concen-tration is considered an error (underestimation). This is the premisebehind using the microwave radiation from artificial ice of knownconcentration to evaluate the results from the algorithms. The studyaims also at linking errors in concentration estimates from eachalgorithm to weather and consequently surface conditions. Thoseconditions and compositions were recorded on hourly or daily basis.Microwave brightness temperature was sampled every 5 min usingsix channels of a ground-based radiometer operated at frequencies19 GHz, 37 GHz and 85 GHz.

In order to make the comparison between algorithm results valid,the tie points needed for NT, NT2, BSA and ASI as well as the distribu-tion of radiometric measurements needed for ECICE were all obtainedfrom the same data of the Ice Tank experiment. These data representthe ice age as required by NT, ASI and BAS as well as ice surface con-ditions as needed by NT2 and ECICE.

The study confirmed that the challenge in determining ice concen-tration is associated with thin ice (b12 cm thickness). It also showsthat advantage of using ECICE, ASI and to some extent BSA over NT2and NT. ECICE was applied with using three separate sets of micro-wave data obtained from the three frequency channels of 19 GHz,37 GHz and 85 GHz.

For very thin ice (b6 cm), NT, NT2 and ASI underestimate the con-centration. BSA and ECICE (from using any microwave frequencydata) return better results in this case. When snow falls on thin iceit changes the surface immediately and triggers a cycle of sharpdrop of Tb followed by gradual increase to the typical ice values.ECICE (using 85 GHz and to some extent 37 GHz) maintains the cor-rect 100% concentration as the surface returns to a stable conditionafter the snow settles. The second best results are obtained fromBSA. When rain falls on thin ice and turned the surface into candledice, ASI and ECICE when used with 85 GHz data produce the near100% ice concentration.

In general the total concentration output from NT echoes almostexactly the evolution of Tb19h. This channel is most sensitive tosnow cover and its metamorphism. NT2 and BSA, with their incorpo-ration of 37 GHz data, are less sensitive to the surface conditions. ASI,with its incorporation of 85 GHz observation, is least sensitive. Theseconclusions are based on the groundmeasurements from the SBR sys-tem but it should be noted that 85 GHz observations from space mustbe corrected to account for atmospheric influences.

Accuracy of each algorithm in estimating total concentration ofthin ice (b6 cm) and thicker ice (>12 cm) was established, includingall surface conditions. For thinner ice, gradual improvement of perfor-mance of algorithm is noticed, with NT as the least accurate followedby NT2, BSA, ASI and ECICE. For thicker ice, all algorithms performequally well when the ice surface is bare or snow covered (withoutsnow metamorphosis). However, NT and NT2 underestimate ice con-centration significantly in presence of melted or slushy surface andunder any form of precipitation (snowfall, ice pallets, or freezingrain). BSA demonstrates better performance. The best performance

is shown by ASI and ECICE using 85 GHz data. If data from thischannel is appropriately corrected for appropriately for atmosphericinfluences, the channels should be superior compared to the lowerchannels.

ECICE was used to identify surface conditions (instead of tradi-tional ice types) and assess the difficulty at which each surface canbe identified. Ice layering within the snow pack impedes retrieval ofice concentration from the 19 GHz data, as confirmed in previousstudies. Slushy surface can best be identified using the 19 GHz or37 GHz observations. Results identify also key conditions that leadto underestimation of ice concentration. They include surface meltand refreezing, slush, ice layering within the snow pack, snow settlingfollowing fresh snowfall, and falling precipitation in different forms.

The study shows also that NT and NT2 are most affected by surfaceprocesses while ASI and BSA are less affected. This is probably be-cause ASI uses the high frequency channel (e.g. SSM/I 85 GHz),which is sensitive only to processes within the top snow layer. More-over, BSA includes filters that account for surface conditions. ECICEproduces better results under most surface conditions. This is perhapsdue to its use of the possible range of radiometric observations foreach surface, as reflected in their probability distributions, insteadof using a single tie point. Additionally, it could be due to the imple-mentation of an ensemble and optimization approaches instead ofsolving a set of algebraic equations.

Acknowledgments

The authors would like to acknowledge the financial support ofthe Canadian Ice Service (CIS) towards conducting the Ice Tank ex-periment. Mr. Ken Asmus of CIS arranged for the experimentalsetup, provided logistic support and carried out data acquisition.The authors would also like to thank the three anonymous reviewerswhose comments have certainly contributed to the improving themanuscript.

References

Andersen, S., Tonboe, R., Kaleschke, L., Heygster, G., & Pedersen, L. T. (2007). Inter-comparison of passive microwave sea ice concentration retrievals over thehigh-concentration Arctic sea ice. Journal of Geophysical Research, 112, C08004.doi:10.1029/2006JC003543 (J. Geophys. Res., 112).

Andreadis, K. M., Liang, D., Tsang, L., Lettenmaier, D. P., & Josberger, E. G. (2008). Char-acterization of errors in a coupled snow hydrology-microwave emission model.Journal of Hydrometeorology, 9, 149–164.

Asmus, A. W., & Grant, C. (1999). Surface Based Radiometer (SBR) data acquisitionsystem. International Journal of Remote Sensing, 20(15/16), 3125–3129.

Cavalieri, D. J., Gloersen, P., & Campbell, W. J. (1984). Determination of sea ice param-eters with the Nimbus-7 scanning multichannel microwave radiometer. Journal ofGeophysical Research, 89(D4), 5355–5369.

Comiso, J. C., & Sullivan, C. W. (1986). Satellite microwave and in situ observations oftheWeddell Sea ice cover and its marginal ice zone. Journal of Geophysical Research,91, 9663–9681.

Comiso, J. C., Cavalieri, D. J., Parkinson, C. L., & Gloersen, P. (1997). Passive microwavealgorithms for sea ice concentration- A comparison of two techniques. RemoteSensing of Environment, 60, 357–384.

Durand, M., Kim, E. J., & Margulis, S. A. (2009). Radiance assimilation shows promisefor snowpack characterization. Geophysical Research Letters, 36, L02503. doi:10.1029/2008GL035214.

Fuhrhop, R., Grenfell, T. C., Heygster, G., Johnsen, A., Schlussel, P., Schrader, M., et al.(1998). A combined radiative transfer model for sea ice, open ocean and atmo-sphere. Radio Science, 33(2), 303–316.

Garrity, C. (1992). Characterization of snow on floating ice and case studies of bright-ness temperature change during onset of melt. In F. D. Carsey (Ed.), Microwaveremote sensing of sea ice. AGU Monograph 68. (pp. 313–328) Washington, DC:American Geophysical Union.

Grenfell, T. C., & Comiso, J. C. (1986). Multifrequency passive microwave observationsof sea ice grown in a tank. IEEE Transactions on Geoscience and Remote Sensing,GE-, 24(6), 826–830.

Grenfell, T. C., Comiso, J. C., Lange, M. A., Eicken, H., & Wensnahan, M. R. (1994). Passivemicrowave observations of the Weddell Sea during austral winter and early spring.Journal of Geophysical Research, 99(C5), 9995–10, 010.

Grenfell, T. C., & Perovich, D. K. (1994). Analysis of surface based passive microwaveobservations during LEADEX 1992. Proceedings of Geoscience and Remote SensingSymposium, 2, 0005–1007.

Page 15: Impact of surface conditions on thin sea ice concentration estimate from passive microwave observations

50 M. Shokr, L. Kaleschke / Remote Sensing of Environment 121 (2012) 36–50

Hallikainen, M. T. (1989). Microwave radiometry of snow. Advances in Space Research,9, 267–275.

Hallikainen, M. T., Ulaby, F. T., & Abdelrazik, M. (1986). Dielectric properties of snow inthe 3 to 37 GHz range. IEEE Transactions on Antennas and Propagation, 34(11),1329–1340.

Hwang, B. J., Ehn, J. K., Barber, D. G., Galley, R., & Grenfell, T. C. (2007). Investigations ofnewly formed sea ice in the Cape Bathurst polynya: 2. Microwave emission. Journalof Geophysical Research, 112. doi:10.1029/2006JC003703.

Josberger, E. G., Gloersen, P., Chang, A., T.C., & Rango, A. (1996). The effects of snowpackgrain size on satellite passive microwave observations from the Upper ColoradoRiver Basin. Journal of Geophysical Research, 101(C3), 6679–6688.

Kaleschke, L., Lupkes, C., Vihma, T., Haarpainter, J., Bochert, A., Hartmann, J., et al.(2001). SSM/I sea ice remote sensing for mesoscale ocean–atmosphere interactionanalysis. Canadian Journal of Remote Sensing, 27(5), 526–537.

Kern, S. (2001), A new algorithm to retrieve the sea ice concentration using weather-corrected 85 GHZ SSM/I measurements, Ph.D Dissertation, University of Bremen,Germany.

Kern, S., Kaleschke, L., & Clausi, D. A. (2003). A comparison of two 85-GHz SSM/I iceconcentration algorithms with AVHRR and ERS-2 SAR imagery. IEEE Transactionson Geoscience and Remote Sensing, 41(10), 2294–2306.

Langlois, A., & Barber, D. G. (2007). Passive microwave remote sensing of seasonalsnow-covered sea ice. Progress in Physical Geography, 31(6), 539–573.

Langlois, A., Scharien, R., Geldsetzer, T., Iacozza, J., Barber, D. G., & Yackel, J. (2008).Estimation of snow water equivalent over first-year sea ice using AMSR-E and sur-face observations. Remote Sensing of Environment, 112, 3656–3667.

Li, Z., Tan, Y., & Tsang, L. (2006). Modelling the passive microwave remote sensing ofsnow using dense media radiative transfer theory with NMM3D rough surfaceboundary conditions. Microwave and Optical Technology Letters, 48(3), 557–562.

Lomax, A. S., Lubin, D., & Whritner, R. H. (1995). The potential for interpreting totaland multiyear ice concentrations in SSM/I 85,5 GHz imagery. Remote Sensing ofEnvironment, 54(1), 13–26.

Maaß, N., & Kaleschke, L. (2010). Improving passive microwave sea ice concentrationalgorithms for coastal areas—Applications to the Baltic Sea. Tellus A, 62(4),393–410.

Markus, T., & Cavalieri, D. J. (2000). An enhancement of NASA Team sea ice algorithm.IEEE Transactions on Geoscience and Remote Sensing, 38(3), 387–1396.

Martin, S., Yu, Y., & Drucker, R. (1996). The temperature dependence of frost flowergrowth on laboratory sea ice and the effect of the flowers on infrared observationsof the surface. Journal of Geophysical Research, 101(C5), 12111–12126.

Mätzler, C. (2006). Thermal microwave radiation—Applications for remote sensing. IETElectromagnetic Waves Series, London, UK (pp. 310–330).

Mätzler, C., Ramseier, R. O., & Svendsen, E. (1984). Polarization effects in sea–ice signa-tures. IEEE Journal of Oceanic Engineering, 9(5), 333–338.

Perovich, D. K., Longarce, J., Barber, D. G., Maffione, R. A., Cota, G. F., Mobley, C. D., et al.(1998). Field observations of the electromagnetic properties of first-year sea ice.IEEE Transactions on Geoscience and Remote Sensing, 36(5), 705–1715 (Part 2).

Pullianen, J., & Hallikainen, M. (2001). Retrieval of regional snow water equivalentfrom space-borne passive microwave observations. Remote Sensing of Environment,75, 76–85.

Shokr, M. E., Asmus, K., & Agnew, T. A. (2009). Microwave emission observations fromartificial thin sea ice: The Ice Tank experiment, IEEE. Transactions on Geoscience andRemote Sensing, 47(1), 325–338.

Shokr, M. E., Lambe, A., & Agnew, T. A. (2008). A new algorithm (ECICE) to estimate iceconcentration from remote sensing observations: An application to 85 GHz passivemicrowave data. IEEE Transactions on Geoscience and Remote Sensing, 46(12),4104–4121.

Spreen, G., Kaleschke, L., & Heygster, G. (2008). Sea ice remote sensing using AMSR-E 89-GHz channels. Journal of Geophysical Research, 113, C02S03. doi:10.1029/2005JC003384.

Sturm, M., Maslanik, J. A., Perovich, D. K., Stroeve, J. C., Richter-Menge, J., Markus, T.,et al. (2006). Snow depth and ice thickness measurements from the Beaufort andChukchi Seas collected during the AMSR-Ice03 campaign. IEEE Transactions on Geo-science and Remote Sensing, 44(11), 3009–3020.

Svendsen, E., Mätzler, C., & Grenfell, T. C. (1987). A model for retrieving total sea iceconcentration from a spaceborne dual-polarized passive microwave instrumentoperating near 90 GHz. International Journal of Remote Sensing, 8(10), 1479–1487.

Tonboe, R., Andersen, A., and Toudal, L. (2003), Anomalous winter sea ice backscatterand brightness temperature, Danish Meteorological Institute, Scientific Report03–13, Copenhagen, ISBN 87-7478-489-7.

Tucker, W. B., III, Perovich, D. K., Gow, Weeks, W. F., & Drinkwater, M. R. (1992). Phys-ical properties of sea ice relevant to remote sensing. In F. Carsey (Ed.), MicrowaveRemote Sensing of Sea Ice. Geophysical Monograph 68. (pp. 9–28) WashingtonD.C.: American Geophysical Union.

Weeks, W. F., & Ackley, S. F. (1986). The growth, structure and properties of sea ice. InN. Untersteiner (Ed.), The geophysics of sea ice. NATO ASI Series B: Physics, vol. 146,New York: Plenum Press.

Wensnahan, M., Grenfell, T. C., Winebrenner, D. P., & Maykut, G. A. (1993). Observa-tions and theoretical studies of microwave emission from thin saline ice. Journalof Geophysical Research, 98(C5), 8531–8545.

Yackel, J., & Barber, D. (2007). Observations of snow water equivalent change onlandfast first-year sea ice in winter using synthetic aperture radar data. IEEETransactions on Geoscience and Remote Sensing, 45(4), 1005–1015.