Sea Ice Extent and Classification Mapping with the Moderate Resolution Imaging Spectroradiometer Airborne Simulator
George A. Riggs1, Dorothy K. Hall2, and Steven A. Ackerman3
1 Research and Data Systems Corp. 7833 Walker Drive, Suite 550, Greenbelt, MD, 20770.
2 Code 974, NASA Goddard Space Flight Center, Greenbelt, MD, 20771.
3 Cooperative Institute for Meteorological Satellite Studies, University of Wisconsin-Madison,
Space Science and Engineering Center, 1225 West Dayton Street, Madison, WI 53706.
ABSTRACT
An algorithm for mapping sea ice extent and generalized
classification of sea ice by reflective and temperature characteristics with Moderate
Resolution Imaging Spectroradiometer (MODIS) Airborne Simulator (MAS) data is presented.
The algorithm was tested using a MAS scene over the Bering Sea near St. Lawrence Island,
Alaska, USA, acquired 8 April 1995. Clouds were masked with the University of Wisconsin
cloud masking algorithm. Ice surface temperature was estimated with a split-window
technique. Sea ice extent and generalized type of sea ice were identified based on
reflective characteristics and estimated ice surface temperature using a grouped criteria
technique. Resulting maps were consistent with visual interpretation and with sea ice
extent and type information reported in prior studies of the region. Published by Elsevier
Science Inc.
Introduction
At its seasonal maximum, sea ice covers roughly 10% of the
Earth's oceans (Carsey, et al., 1992). It varies greatly in both extent and thickness due
to the combined effect of dynamic and thermodynamic processes. In the Arctic it varies
from a March maximum of approximately 15 x 106 km2, to a September
minimum of about 8 x 106 km2 (Parkinson and Gloersen, 1993). In the
Antarctic, a maximum extent of approximately 20 x 106 km2 occurs in
September while a minimum of 4 x 106 km2 occurs in February (Parkinson
and Gloersen, 1993). The presence of sea ice influences atmospheric and oceanic temperature
and circulation patterns, and reduces the amount of solar radiation absorbed in the ocean.
Sea ice and its associated snow cover serves as a strong insulator, restricting exchanges of
heat, mass, momentum and chemical constituents between the atmosphere and the ocean. With an
ice cover, only 30-50% of the incident solar radiation is absorbed. Without sea ice, 85-95%
of the incident solar radiation is absorbed by the oceans (Washington and Parkinson, 1986).
Satellite remote sensing is an important tool for global mapping of sea ice extent, concentration,
type and temperature. Two classes of sensors, microwave and multi-spectral radiometers, are
typically used. The ability of microwave instruments, e.g. the special sensor microwave imager
(SSM/I) and synthetic aperture radars (SARs), to collect data through cloud cover and polar darkness
makes them well suited for monitoring of sea ice extent and concentration. Passive-microwave sensors
acquire data at spatial resolutions of 25-50 km, while SARs may acquire data at 25 m resolution.
However, microwave instruments cannot collect data on spectral reflectance or surface temperature which
is needed for analysis of energy exchange between the ocean and atmosphere. Thus, multispectral
radiometers also play an important role in sea ice research.
Multi-spectral radiometers, e.g., the Advanced Very High Resolution Radiometer (AVHRR), the Moderate
Resolution Imaging Spectroradiometer (MODIS) and the MODIS Airborne Simulator (MAS) can be used for
determination of sea ice extent, temperature, albedo, movement, type, and concentration at resolutions
of 250 m - 1 km. After its launch in 1999 on the Earth Observing System (EOS) AM-1 platform, MODIS
will provide, data from 36 spectral bands from visible to thermal infrared (0.45 to 14.4 µ m)
at a spatial resolution of 250 - 1000 m. However, cloud cover and polar darkness preclude continuous
spatial and temporal coverage with those sensors.
The prototype of the MODIS sea ice algorithm is described in this paper, which used MAS to develop
the sea ice mapping methodology. The algorithm will generate maps of sea ice extent as determined
by reflectance characteristics and estimated ice surface temperature (IST) and will include the
estimated IST as part of the data product. The product will be produced globally at 1 km spatial
resolution. In the EOS era, MODIS will be used to generate a standard data product in the Earth
Observation System Data Information System (EOSDIS).
Data Description
MODIS Airborne Simulator
The MAS is a multi-spectral imaging spectrometer that measures reflected solar and emitted thermal
radiation between 0.55 and 14.2 µ m in 50 narrow-band channels, that is flown aboard a NASA ER-2
aircraft. Nineteen of the 50 channels of the MAS have comparable channels on the MODIS
(Figures 1a and 1b). MODIS channels 8-16, designed for ocean color
applications, are shown in Figures 1a and 1b for completeness but may
be of limited use over sea ice due to potential sensor saturation. At a nominal aircraft altitude of
20 km, MAS provides imagery with a swath width of 37 km and a pixel spatial resolution of 50 m at nadir.
The MAS was built to support development of MODIS remote-sensing algorithms. A detailed description of
the MAS, some MAS applications and field campaigns can be found in King et al. (1996) or at the MAS World
Wide Web site (URLs are listed under the Relevant World Wide Web sites heading). A description of the
MODIS instrument can be found in Barnes et al. (1998) or at the MODIS World Wide Web site.
Test Scene
A MAS scene acquired on 8 April 1995 during the Alaska-April95 aircraft campaign was selected for its
good combination of sea ice types, open water and clouds. The scene is located in the Bering Sea with
a segment of the southwest coast of St. Lawrence Island included (Figure 2)
and has a center location of approximately 63.42° north latitude, 172.10° west longitude. The
flight direction is oriented northeast to southwest. A three channel (MAS channels 1,7,10) color
composite of the scene is shown in Figure 3 and is used as a reference
image on which algorithm results are overlaid.
Sea Ice Algorithm
The primary purpose of the MODIS sea ice algorithm is to map sea ice extent. A potential secondary
purpose is the creation of thematic maps showing generalized sea ice classes. The algorithm identifies
sea ice by reflective characteristics and ice surface temperature, generating a four part data product.
A grouped criteria methodology is employed to identify sea ice on a per pixel basis. The rationale for
using the grouped criteria methodology is that the criteria are robust, can be easily modified or changed
and can be dynamically linked with geographic or seasonal parameters affecting identification of sea ice,
and is amenable to enhancements derived from future studies and independent validation. A flow diagram
of the algorithm is depicted in Figure 4. The University of Wisconsin
cloud mask (UWCM) algorithm was chosen to mask clouds because it will produce the standard MODIS cloud
mask data product. The UWCM algorithm uses select channels in the visible, infrared and thermal channels
in various cloud detection tests (single band thresholds, ratios, channel differences and others) with
threshold rules for detection of cloud. Experience gained with the UWCM using the MAS is directly
applicable to cloud masking with MODIS. A brief background discussion of sea ice reflectance characteristics
and ice surface temperature calculation is presented, followed by descriptions of groups of criteria applied
in the algorithm.
Sea Ice Reflective Characteristics
Observations of sea ice reflectance reported in the literature commonly cover the 0.4 - 1.0 µ
m range (e.g. Perovich, 1990; De Abreu et al., 1995) and sometimes extend out to approximately 2.4
µ m (e.g. Grenfell and Perovich 1984; Perovich et al., 1986). Snow-covered, opaque, white sea ice,
thick first-year ice and multiyear ice typically exhibit maximum reflectance between 0.4 - 0.8 µ m,
decreasing into the near-infrared, reaching a reflectance minimum at approximately 1.6 µ m, then rising
again to a peak at approximately 1.9 µ m. Field observations show that sea ice is predominately
snow-covered (Schlosser, 1988) and that this snow is usually present within a few days of formation
(Perovich, 1990). Young sea ice (e.g., grease ice, nilas, young gray) has lower spectral albedos (in the
range of 0.10 - 0.40) than older sea ice over the 0.4 - 0.8 µ m region. At longer wavelengths
(› 1.0 µ m) the spectral albedo of young sea ice decreases to a value similar to older sea
ice (Schlosser, 1988; Perovich, 1990; Allison et al., 1993). Sea ice undergoing ablation and containing
melt ponds has decreasing reflectance from 0.6 to 0.8 µ m, followed by steady uniform low reflectance
to approximately 1.6 µ m (Grenfell and Perovich, 1984; De Abreu et al., 1995). Reflectance curves
(Figure 5) constructed from the mean reflectance of each MAS band from
two spatially separated 5 x 5 pixel samples band of ice, cloud and water in the scene show reflectance
features consistent with those reported in the sea ice observation studies discussed above. Using these
sample curves from the scene and above studies as background, the sea ice reflectance criteria were
developed to identify snow-covered sea ice and new sea ice.
Ice Surface Temperature
Split-window techniques are employed to determine sea-surface temperature (e.g. McMillin and Crosby,
1984; Minnett, 1990) or ice-surface temperature (e.g. Key and Haefliger 1992; Massom and Comiso, 1994;
Yu et al., 1995; Key et al., 1997). Split-window techniques allow for correction of atmospheric effects,
primarily water vapor, and yield reasonably accurate sea surface temperature (SST) or ice surface
temperature (IST) estimates.
Instrument channels used for estimation of surface temperature with the split-window technique are centered
at approximately 11.0 and 12.0 µ m. Several investigators (e.g. Key and Haefliger, 1992; Lindsay and
Rothrock, 1993; Massom and Comiso, 1994; Yu et al., 1995; Key et al., 1997) have employed variations of the
split-window technique for estimation of IST in polar regions using the AVHRR and other instruments. In most
studies, accuracy of the split-window technique has been enhanced by adding predictor variables to account for
effects such as sensor scan angle. Accuracy of the estimate of IST may be increased by regression of the
estimated IST with temperatures modeled from radiative transfer models or with observed surface temperatures
(e.g. Minnett, 1990; Key and Haefliger, 1992; Lindsay and Rothrock, 1993; Key et al., 1997). IST accuracy
of 0.5 - 1.5 K relative to measured or modeled surface temperatures has been reported by investigators using
the split-window technique.
The technique of Key et al. (1997) was adapted for estimating IST using MAS data. MAS channels 45 and 46,
centered approximately at 11.0 and 12.0 µ m, respectively (Figure 1b),
are used to determine IST. IST for MAS data is estimated using the equation
IST = a + bT45 + c(T45 - T46) + d[(T45
- T46) * sec(theta - 1)], (1)
where
IST -- ice surface temperature (K)
a,b,c,d -- are coefficients determined from multilinear regression of brightness temperatures to estimated surface temperatures (4)
T45 -- brightness temperature of MAS channel 45 (11.0 m
m)
T46 -- brightness temperature of MAS channel 46 (12.0 m
m)
theta
-- sensor scan angle (0 - 43°
for MAS)
When MODIS becomes operational, MODIS channels 31 and 32, centered at approximately 11.0 and 12.0 µ m,
respectively (Figure 1b), are specified to be used in estimating IST with (1).
Data Processing
Radiance and conversion data required for calculation of reflectance
in visible channels and for calculation of brightness temperature in thermal channels were extracted from
the MAS data file. MAS channel radiance data stored as scaled-radiance data, were descaled then converted
to reflectance or brightness temperature within the algorithm (Figure 4).
Equations used to descale and convert the MAS channel data are described in Gumley et al. (1994).
Reflectance was calculated as
Rlambda
= Ilambda
/ Io lambda
* cos(theta)(2)
where
Rlambda
= reflectance at lambda µ m)
Ilambda
= radiance measured in the channel (W m-2 sr-1 µ m-1)
Io lambda
= solar spectral irradiance (W m-2 µ m-1)
theta
= solar zenith angle.
MAS channels 1, 7, 10, 12 and 20, centered at 0.547 µ m, 0.869 µ m, 1.609 µ m, 1.723
µ m, and 2.1 µ m, respectively, were used for detection of sea ice reflectance characteristics.
Thermal channel data are converted from radiance to brightness temperature by the inversion of Planck's
equation (Gumley et al., 1994): [Eq. 3]
T= C2 / lambda ln [(C1 / lambda5R 106))+1.0] (3)
where
T = brightness temperature in K
C1 = 2hc2 = 1.1910439 × 10-16 W m-2
C2 = (hc)/k = 1.4387686 &3215 10-2 m K
lambda
= wavelength in m
R = Planck radiance (W m-2 sr-1 m-1)
Masking
Land: Land was masked using the 1 km resolution United States
Geological Survey (USGS) global land/water mask (USGS, 1997), as used in the UWCM algorithm. This
mask will be incorporated into the MODIS geolocated sea-ice data product.
Cloud: The UWCM algorithm was applied to the MAS scene with the resulting cloud mask used as
input to the sea ice algorithm (Figure 4). The UWCM initializes all
cloud-test flags to cloudy, creating the initial condition of complete cloud cover. Subsequently, the
algorithm clears the flags if no clouds are detected by the cloud tests. The single unobstructed
field-of-view (FOV) flag, essentially a clear or cloudy flag, was initially used as the cloud mask
for the scene. If the unobstructed FOV flag probability was >= 66% the pixel was interpreted
as clear. Only the flagged condition of "certain cloudy" was interpreted as a cloud.
Criteria Groups
Sea ice by reflectance: The function of this criteria group (Figure 4)
is to identify snow-covered sea ice. Beginning with the assumption that sea ice is snow covered and that
snow dominates the reflectance characteristics, the MODIS snow-mapping techniques for land (Hall et al., 1995;
Riggs et al., 1996) are applied to detect sea ice. The main criterion is the normalized difference snow index
(NDSI). The NDSI is used to detect the high reflectance of snow at visible wavelengths, and the low reflectance
at approximately 1.6 µ m (Figure 5).
NDSI = (visible reflectance - near-infrared reflectance) / (visible reflectance - near-infrared reflectance).
NDSI is calculated with MAS bands 1 and 10; NDSI = (1- 10)/(1+10). NDSI will be calculated with MODIS bands
4 and 6, respectively.
The second criterion in this group is high reflectance in a visible wavelength where snow has a high and
water has a very low reflectance. The purpose of this criterion is to separate water, which has high NDSI
values similar to snow (Hall et al., 1995; Riggs et al., 1996), from snow-covered sea ice. If only the NDSI
criterion was used, water would be confused with snow-covered sea ice. Sea ice is identified where NDSI
>= 0.4 and visible reflectance > 0.17 in MAS channel 1 (0.55 µ m).
Sea ice by IST:The function of this criteria group (Figure 4) is
to identify sea ice by the estimated IST. A vicarious approach was taken to estimate surface temperature
as a surrogate for observed temperature in determination of coefficients for IST (1). Assuming that the
polar atmosphere in the cold season under clear sky conditions is nearly transparent (Yu et al., 1995),
the surface temperature may be estimated directly from the observed brightness at 11.0 µ m (Key et
al., 1997).
At 11.0 µ m, atmospheric effects are minimal, thus a greater signal
from the surface is measured, relative to a channel at 12.0 µ m (Massom and Comiso, 1994).
Surface temperature was estimated directly by the general relationship
Ts = Tb/epsilon¼ (4)
where
Ts = surface temperature
Tb = brightness temperature
epsilon
= surface emissivity
The channel chosen for this calculation was MAS channel 45, centered at
11.0 µ m. If the difference of surface temperature at 11.0 and 12.0 µ m
(Ts45 - Ts46 ) is small, then the assumption of a clear atmosphere is reasonable.
In this scene, the mean temperature difference (Ts45 - Ts46 ) for the sea water
samples is 0.39 K. Use of the direct surface temperature estimate (4) gives an estimated surface
temperature adjusted to the assumed emissivity of the surface. Sea water emissivity is assumed to be
0.985 (Comiso, 1994). The mean Ts45 for sea water was estimated to be 271.6 K. Sensor
zenith angle is neglected because the equation is applied to essentially a point sampling of data.
Four five-by-five pixel samples of open water off the coast of St. Lawrence Island were extracted for
estimation of surface temperature and multi-linear regression using (1).
A multi-linear regression using the directly estimated surface temperature Ts45 [Eq. (4)]
and sensor brightness temperatures for sea water was run to determine the coefficients used in the split
window technique (1). Values for the coefficients are;
a = -0.0024
b = 1.0038
c = -1.27 × 10-6
d = 1.87 × 10-5
IST was then determined for the scene using these coefficients in (1).
Arctic sea water can have freezing points over the range 271.36 - 271.45 K
(Yu and Rothrock, 1996). For this scene a freezing point of 271.4 K was assumed based on direct estimates
of open water and ice in the scene. A pixel with a temperature less than 271.4 K was identified as sea ice;
any pixel with a higher temperatures was identified as open water.
Combined: The results of the preceding two criteria groups are combined to show where they agree
and disagree in identifying sea ice. Combination maps where sea ice is identified by both criteria groups,
one criteria group but not the other, and by neither criteria group are created using Boolean and relational
expressions.
Discussion and Results
The sea ice extent and ice type maps are in general agreement with
visual interpretation of sea ice in the scene, and previous studies of sea ice types in the same
region (e.g. Massom and Comiso, 1994). An ocean field campaign was not associated with the acquisition
of the MAS data during the April 1995 campaign, thus analysis is confined to descriptive and comparative
analysis with previous studies. Results of cloud masking are discussed followed by sea ice mapping by
reflectance characteristics and estimated IST.
Discrimination of clouds is particularly difficult in the polar regions because of the lack of contrast
between temperatures and reflectance characteristics of certain cloud types and sea ice. The UWCM product
contains information on the probability of cloud, individual cloud test results and the decision path taken
within the algorithm, on a per pixel basis. The UWCM algorithm has eight conceptual domains that apply
different cloud screening tests for several specified conditions related to geographical location, surface
conditions, and viewing conditions (Ackerman et al., 1996). This approach facilitates adapting the algorithm
for global application while seeking to maximize accuracy of cloud detection. Though four domains and their
cloud screening tests were possible depending on the surface in this MAS scene, we are concerned only with
the daytime ocean domain. Table 1 presents the spectral reflectance (Rlambda)
and brightness temperature (BTlambda) cloud screening tests used in domains appropriate for this
scene.
Using only the unobstructed FOV flag, clouds in the scene were masked. In this scene some areas of ocean
were masked as clouds (Figure 6). Clouds were identified over areas of new
ice or open water e.g., in the lower right quadrant, near the St. Lawrence Island coast, and in the left
third of the scene (Figure 6). Cloud misidentification appeared to occur
in the domain of daytime ocean (column one in Table 1), where the ice/snow background
flag was set to off. Investigation of the individual cloud tests done in that processing path revealed that
the most accurate mapping of clouds was obtained with a single cloud test, the 11 - 4 µ m brightness
temperature difference test. When only this test was used as the cloud mask, clouds were mapped with high
accuracy (Figure 7) as determined from visual analysis of the image.
The source of cloud misidentification is the visible reflectance cloud screening test. This is a single
channel reflectance test for discriminating bright clouds over dark surfaces. The channel and reflectance
threshold used in this test depend on the surface being viewed. In the daytime ocean domain
(Table 1) this visible reflectance test used MAS channel 6 (0.66 µ m) with a
reflectance threshold of 7.0%. Ocean pixels flagged as having no ice/snow background were misidentified as
cloud by this test. No other cloud tests applied in this processing domain were found to misidentify cloud.
The misidentification was caused by greater-than-expected reflectances from the ocean in MAS channel 6. The
mean reflectance of the ocean under the misidentified clouds is 8.1%. Since this exceeds the reflectance
threshold, these pixels were identified as cloud. Two possible reasons for the high ocean reflectances are:
1) the ocean surface was not totally ice-free, or 2) the sea surface was significantly wind-roughened.
Plots of reflectance for features in the scene (Figure 5) indicate that
some areas of the ocean had relatively high reflectance, assuming negligible atmospheric contribution to
the signal. Also confounding the test was the fact that some of the new ice types had reflectances near
those for cloud at visible wavelengths, but not in the infrared. The misidentification of cloud in this
situation does not invalidate the use of the UWCM. Rather, it highlights the need for the user to be aware
of and to make use of the individual cloud tests as necessary.
The ice by reflectance criteria group accurately (by visual interpretation) mapped the extent of sea ice,
snow-covered and new sea ice in the scene (Figure 7). These selected
reflectance criteria were based on sea ice reflectances reported in the literature and observed in the MAS
data. Only sea ice extent was determined by this criteria group. With the many channels on the MAS and the
MODIS sensors, it may also be possible to identify broad sea ice classes. Sea ice reflectance curves
(Figure 5) suggest that two generalized classes, namely new sea ice and
snow-covered sea ice, may be differentiated on the basis of reflectances. These classes are interpreted
to be related to the World Meteorological Organization (WMO) stages of development, with new ice corresponding
to stages that are < 10 cm thick and snow-covered ice corresponding to stages > 10 cm thick. New
sea ice types exhibited reflectances intermediate between snow covered sea ice and water over the 0.5 - 0.9
µ m region (Figure 5). New sea ice reflectance characteristics could
cause confusion with clouds in the scene (Figure 5), however the cloud mask
removed this potential confusion. Another difference between snow-covered sea ice and new ice was noted in
the 1.6 - 1.9 µ m region. In this spectral region, snow-covered sea ice has a local peak in reflectance
(Figure 5) while new sea ice has a relatively uniform low reflectance
(Figure 5).
These observed reflectance differences were used to create thematic maps of new, and snow-covered sea ice.
Two methods of classifying ice were implemented. The intermediate reflectance of new ice over the 0.6 -
0.8 µ m region was approximately 60% lower than snow-covered ice and approximately 30% lower than young
ice types (Figure 5). This intermediate reflectance (25% to 45%) was added
as a criterion to the reflectance criteria group to identify new ice. A pixel meeting both the original
criteria and this new ice criterion is identified as new sea ice. If it meets only the original criteria
it is identified as snow-covered sea ice (Figure 8a).
Also, snow-covered and new sea ice can be differentiated using the near-infrared ratio of MAS channel 12
(1.7 µ m) to channel 10 (1.6 µ m). Snow-covered sea ice has ratio values >= 1.1 and
new ice has ratio values of < 1.1. This criterion test (snow-covered sea ice if ratio >=
1.1 and new ice if < 1.1), along with a threshold test in MAS channel 1 (0.55 µ m) can be used
separate water from new ice and was implemented as a separate criterion group. This group resulted in a
reasonably accurate map of snow-covered sea ice (Figure 8b).
Both maps of sea ice classes (Figure 8a and Figure 8b)
agree with visual interpretation of ice type, and are in general agreement with general patterns of sea ice
extent and type for the region as reported in previous studies (e.g. Massom and Comiso, 1994). Absence of
in situ observations limit these results to qualitative discussion. However, they do show that
additional information about sea ice beyond sea ice extent is possible with the MAS and MODIS sensors. For
the MODIS sea ice algorithm only the criteria groups and tests given in Figure 4
will be used in the first version after launch.
Sea ice extent (Figure 9) from the IST criteria group was in agreement with
interpreted sea ice extent in Figure 3. The sea ice extent map created by
flagging any pixel with an IST less than 271.4 K fails to make full use of the IST information. Ice
classification is possible using IST because of the marked thermal contrast between open water, new ice and
first-year and older sea ice types (Massom and Comiso, 1994). Also, the surface temperature of ice is
strongly related, in principle, to ice thickness, allowing inference of ice thickness (Yu and Rothrock,
1996). Based on this information, a hypothetical three-classification sea ice map was created using the
temperature ranges of the classes shown in Table 2. The ice-type map for this
classification scheme is shown in Figure 9. New ice and young ice types
are the dominant ice types identified. Some first-year ice was identified along the coast of St. Lawrence
Island which is a mix of both fast ice and snow-covered land which was not masked by the land mask. Without
surface observations this classification cannot be verified, but it is a reasonable classification for the
region based on descriptions in Massom and Comiso (1994) for the area around St. Lawrence Island for the
same time of year during 1988. The observed consistency of sea ice type and distribution as determined by
the estimated IST method between the studies is encouraging.
Independent validation activities in support of the MODIS sea ice algorithm will be performed to
determine the coefficients, modify the IST equation, and possibly to include dynamically-adaptive
coefficients that can be adjusted for specific geographical regions, seasons, surface temperatures, or
a combination of parameters in order to increase accuracy of the IST.
The motivation for combining the results of the two criteria groups is to enhance analysis and
identification of sea ice, primarily in areas of new-ice formation when the reflectance characteristics of
new sea ice make it difficult to detect and where the surface temperature is near the freezing point of sea
water. Mapped sea ice extent was different in the reflectance and IST criteria groups. There are areas in
the scene, (e.g. along the left third and right middle third of the scene), that appear to be open water in
the color-composite image (Figure 3) that are not identified as ice by the
reflectance criteria group for ice (Figure 7) but are identified as new ice
by the IST (Figure 9) classification. Differences in sea ice extent between
the two techniques are areas where there may be a mix of sea ice and open water or where other factors were
affecting the surface characteristics. Horizontal striping in Figure 9 and
Figure 10 suggest sensor performance effects in the thermal bands. The most
certainty in the sea ice extent occurs where the two identification techniques (reflectance and IST
(Figure 4)) agree (Figure 10).
Summary
Using reflectance and thermal data from MODIS or MAS it is possible to map
sea ice extent and possibly classify sea ice type. The UWCM provided a reasonably accurate masking of clouds
over sea ice when only specific cloud tests were considered in this analysis. Thus use of the cloud mask in
polar regions may require an in-depth analysis of the cloud mask data content to maximize its effectiveness
and accuracy. Albeit limited in scope, this work has been used to develop the MODIS sea ice algorithm.
Algorithm development and analysis continues with MAS data, including comparison of results with SSM/I derived
sea ice extent. It is anticipated that a similar capability for mapping sea ice will be obtained with the
MODIS sea ice algorithm.
The MODIS sea ice algorithm is developed with the same criteria described here for mapping sea ice with
MAS data (Figure 4) and is fully automated. Classification of sea ice
type is not yet included in the MODIS algorithm. The algorithm will be bootstrapped with actual MODIS
data using the coefficients reported here or with better determined coefficients from other studies with
MAS. Initial estimated IST accuracies with MODIS data are anticipated to be poor and the coefficients
adjusted frequently during the first year to improve accuracy. Validation studies are expected to be
used to identify adjustments to the coefficients and to define sets of coefficients for use in specific
regions (e.g. regions within the Arctic, or Antarctica) and possibly seasonal subsets for regions. Sea
ice will be mapped both day and night with MODIS, though at night only the IST will be used. The MODIS
cloud mask will be used for both day and night products. The UWCM does night time detection of clouds
using thermal tests, however, performance of the cloud mask at night over sea ice is still being investigated
and is expected to be lower than its performance during daytime.
The MODIS sea ice data product will be archived in Hierarchical Data Format-EOS (HDF-EOS) composed of
scientific data sets and metadata. Sea ice derived from reflectance characteristics, estimated IST, sea
ice extent by IST and combined sea ice extent (Figure 4) are archived as
scientific data sets in the data product. Night time products will have only IST arrays. Extensive
metadata (information about the data product) supporting EOSDIS Core System (ECS) services and containing
quality assurance information and product summary information is also included with each product. Sea ice
products are to be archived at the National Snow and Ice Data Center (NSIDC) in Boulder, Colorado (Scharfen
et al., 1997). Sea ice products at 1 km spatial resolution are planned for swaths ("scenes") day and night
of MODIS data. Gridded daily products at 1 km and 0.25° spatial resolution in Lambert azimuthal
equal-area polar projection, and composite products of eight day and monthly at 0.25° spatial
resolution are planned.
Acknowledgments
Thanks to N. Larsen (U. Alaska, Fairbanks) for thought-provoking
discussions throughout the course of this project, and to A. Klein (Texas A &M University) for
assistance with image production and review of the manuscript. Thanks to K. Strabala, R. Frey and
L. Gumley for their important roles in coding the MAS cloud algorithm. Thanks also to anonymous
reviewers for useful comments and suggestions that much improved the manuscript.
Relevant World Wide Web sites:
References
Ackerman, S., Strabala, K., Menzel, P., Frey, R., Moeller, C., Gumley, L., Baum, B., Schaaf, C., Riggs, G., and Welch, R. (1996), Discriminating clear-sky from cloud with MODIS, Algorithm Theoretical Basis Document (MOD35), Version 3.0, 109 pp.
Allison, I., Brandt, R. E., and Warren, S. G. (1993), East Antarctica sea ice: albedo, thickness distribution, and snow cover. J. Geophys. Res. 98(C7):12,417-12,429.
Barnes, W. L., Pagano, T. S., and Salomonson, V. V. (1998), Prelaunch characteristics of the Moderate Resolution Imaging Spectroradiometer (MODIS) on EOS-AM 1. IEEE Trans. Geosci. Remote Sens. 36(4);1088-1100.
Carsey, F. D., Barry, R. G. and Weeks, W. F. (1992), Chapter 1. Introduction, In: Microwave Remote Sensing of Sea Ice (F. D. Carsey, Ed.), Geophysical Monograph 68, AGU, pp. 1-7.
Comiso, J. C. (1994), Surface temperatures in the polar regions from Nimbus 7 temperature humidity infrared radiometer. J. Geophys. Res. 99(C3):5,181-5,200.
De Abreu, R. A., Barber, D. G., Misurak, K., and LeDrew, E. F. (1995), Spectral albedo of snow-covered first-year and multiyear sea ice during spring melt. Ann. Glaciol. 21:337-342.
Grenfell, T. C., and Perovich, D. K. (1984), Spectral albedos of sea ice and incident solar irradiance in the Southern Beaufort Sea. J. Geophys. Res. 89(C3): 3,573-3,580.
Gumley, L. E., Hubanks, P. A., and Masuoka, E. J. (1994), MODIS Technical Report Series, Volume 3, MODIS Airborne Simulator Level 1B Data User's Guide, NASA Tech. Memo. 104594, Vol.3, 37 pp.
Hall, D. K., Riggs, G. A., and Salomonson, V. V. (1995), Development of methods for mapping global snow cover using Moderate Resolution Imaging Spectroradiometer data. Rem. Sens. Environ. 54:127-140.
Key, J. R., and Haefliger, M. (1992), Arctic ice surface temperature retrieval from AVHRR thermal channels. J. Geophys. Res. 97(D5): 5,885-5,893.
Key, J. R., Collins, J. B., Fowler, C., and Stone, R. S. (1997), High-latitude surface temperature estimates from thermal satellite data. Rem. Sens. Environ. 61:302-309.
King, M. D., Menzel, W. P., Grant, P. S., Myers, J. S., Arnold, G. T., Platnick, S. E., Gumley, L. E., Tsay, S., Moeller, C. C., Fitzgerald, M., Brown K. S., and Osterwisch, F. G. (1996), Airborne scanning spectrometer for remote sensing of cloud, aerosol, water vapor, and surface properties. J. Atmos. and Ocean. Technol. 13:777-794.
Lindsay, R., and Rothrock, D. (1993), The calculation of surface temperature and albedo of Arctic sea ice from AVHRR. Ann. Glaciol. 17:174-183.
Massom, R., and Comiso, J. C. (1994), The classification of Arctic Sea ice types and the determination of surface temperature using advanced very high resolution radiometer data. J. Geophys. Res. 99(C3):5201-5218.
McMillin, L. M., and Crosby, D. S. (1984), Theory and validation of the multiple window sea surface temperature technique. J. Geophys. Res. 89:3655-3661.
Minnett, P. J. (1990), The regional optimization of infrared measurements of sea surface temperature from space. J. Geophys. Res. 95:13497-13510
Parkinson, C. L., and Gloersen, P. (1993), Global sea ice coverage, In: Atlas of Satellite Observations Related to Global Change. (R. J. Gurney, J. L. Foster, and C. L. Parkinson, eds.) Cambridge University Press, pp. 371-384.
Perovich, D. K. (1990), The evolution of sea ice optical properties during fall freeze-up. SPIE Vol. 1302, Ocean Optics X. 520-531.
Perovich, D. K., Maykut, G. A., and Grenfell, T. C. (1986), Optical properties of ice and snow in the polar oceans. I: observations. SPIE Proceedings, Vol. 637, Ocean Optics VIII. 232-241.
Riggs, G. A., Hall, D. K., and Salomonson, V. V. (1996), Recent progress in development of the Moderate Resolution Imaging Spectroradiometer snow cover algorithm and product. In 1996 International GeoScience and Remote Sensing Symposium, Remote Sensing for a Sustainable Future, 27-31 May, 1996, Lincoln, NE. Vol 1:139-141.
Scharfen, G. R., Hall, D. K., and Riggs, G. A. (1997), MODIS snow and ice products from the NSIDC DAAC. SPIE Proceedings, Vol. 3117, In Earth Observing Systems II, (W.L. Barnes, ed.) 143-147.
Schlosser, E. (1988), Optical studies of Antarctic sea ice. Cold Regions Sci. Tech. 15:289-293.
United States Geological Survey. (1997), EROS Data Center Distributed Active Archive Center, http://edcwww.cr.usgs.gov/landdaac/1KM/land_sea_mask.html.
Washington, W. M., and Parkinson, C. L. (1986), An Introduction to Three-Dimensional Climate Modeling. University Science Books, Mill Valley, CA, 422 pp.
Yu, Y., and Rothrock, D. A. (1996), Thin ice thickness from satellite thermal imagery. J. Geophys. Res. 101:25,753-25,766.
Yu, Y., Rothrock, D. A., and Lindsay, R. W. (1995), Accuracy of sea ice temperature derived from the advanced very high resolution radiometer. J. Geophys. Res. 100(C3):4,525-4,532.
List of Tables
Table 1. UWCM domains and cloud screening tests used for the MAS scene.
The spectral reflectance (R) and brightness temperature (BT) are listed. An × mark indicates
that a test was done for that domain.
Domains
Cloud Screening Tests
Daytime
Ocean
Daytime
Land
Snow
Day
Coastline
Day
BT11
×
BT13.9
×
×
×
×
R1.38
×
×
×
×
BT8 - BT11 and BT11 - BT12
×
×
×
BT3.7 - BT11
×
×
×
×
R0.66
×
×
×
R0.87/R0.66
×
×
×
Spatial Variability
×
Table 2. Sea ice classes by temperature range.
Class
Temperature Range open water
IST >=
271.4 K
new ice
271.4 K < IST <=
268.0 K
young ice
268.0 < IST < 262.0 K
first-year ice
IST <=
262.0 K
List of Figures
Figure 1
Figure 2
Figure 3
Figure 4
Figure 5
Figure 6
Figure 7
Figure 8a
Figure 8b
Figure 9
Figure 10
ABSTRACT
An algorithm for mapping sea ice extent and generalized classification of sea ice by reflective and temperature characteristics with Moderate Resolution Imaging Spectroradiometer (MODIS) Airborne Simulator (MAS) data is presented. The algorithm was tested using a MAS scene over the Bering Sea near St. Lawrence Island, Alaska, USA, acquired 8 April 1995. Clouds were masked with the University of Wisconsin cloud masking algorithm. Ice surface temperature was estimated with a split-window technique. Sea ice extent and generalized type of sea ice were identified based on reflective characteristics and estimated ice surface temperature using a grouped criteria technique. Resulting maps were consistent with visual interpretation and with sea ice extent and type information reported in prior studies of the region. Published by Elsevier Science Inc.
Introduction
At its seasonal maximum, sea ice covers roughly 10% of the
Earth's oceans (Carsey, et al., 1992). It varies greatly in both extent and thickness due
to the combined effect of dynamic and thermodynamic processes. In the Arctic it varies
from a March maximum of approximately 15 x 106 km2, to a September
minimum of about 8 x 106 km2 (Parkinson and Gloersen, 1993). In the
Antarctic, a maximum extent of approximately 20 x 106 km2 occurs in
September while a minimum of 4 x 106 km2 occurs in February (Parkinson
and Gloersen, 1993). The presence of sea ice influences atmospheric and oceanic temperature
and circulation patterns, and reduces the amount of solar radiation absorbed in the ocean.
Sea ice and its associated snow cover serves as a strong insulator, restricting exchanges of
heat, mass, momentum and chemical constituents between the atmosphere and the ocean. With an
ice cover, only 30-50% of the incident solar radiation is absorbed. Without sea ice, 85-95%
of the incident solar radiation is absorbed by the oceans (Washington and Parkinson, 1986).
Satellite remote sensing is an important tool for global mapping of sea ice extent, concentration,
type and temperature. Two classes of sensors, microwave and multi-spectral radiometers, are
typically used. The ability of microwave instruments, e.g. the special sensor microwave imager
(SSM/I) and synthetic aperture radars (SARs), to collect data through cloud cover and polar darkness
makes them well suited for monitoring of sea ice extent and concentration. Passive-microwave sensors
acquire data at spatial resolutions of 25-50 km, while SARs may acquire data at 25 m resolution.
However, microwave instruments cannot collect data on spectral reflectance or surface temperature which
is needed for analysis of energy exchange between the ocean and atmosphere. Thus, multispectral
radiometers also play an important role in sea ice research.
Multi-spectral radiometers, e.g., the Advanced Very High Resolution Radiometer (AVHRR), the Moderate
Resolution Imaging Spectroradiometer (MODIS) and the MODIS Airborne Simulator (MAS) can be used for
determination of sea ice extent, temperature, albedo, movement, type, and concentration at resolutions
of 250 m - 1 km. After its launch in 1999 on the Earth Observing System (EOS) AM-1 platform, MODIS
will provide, data from 36 spectral bands from visible to thermal infrared (0.45 to 14.4 µ m)
at a spatial resolution of 250 - 1000 m. However, cloud cover and polar darkness preclude continuous
spatial and temporal coverage with those sensors.
The prototype of the MODIS sea ice algorithm is described in this paper, which used MAS to develop
the sea ice mapping methodology. The algorithm will generate maps of sea ice extent as determined
by reflectance characteristics and estimated ice surface temperature (IST) and will include the
estimated IST as part of the data product. The product will be produced globally at 1 km spatial
resolution. In the EOS era, MODIS will be used to generate a standard data product in the Earth
Observation System Data Information System (EOSDIS).
Data Description
MODIS Airborne Simulator
The MAS is a multi-spectral imaging spectrometer that measures reflected solar and emitted thermal
radiation between 0.55 and 14.2 µ m in 50 narrow-band channels, that is flown aboard a NASA ER-2
aircraft. Nineteen of the 50 channels of the MAS have comparable channels on the MODIS
(Figures 1a and 1b). MODIS channels 8-16, designed for ocean color
applications, are shown in Figures 1a and 1b for completeness but may
be of limited use over sea ice due to potential sensor saturation. At a nominal aircraft altitude of
20 km, MAS provides imagery with a swath width of 37 km and a pixel spatial resolution of 50 m at nadir.
The MAS was built to support development of MODIS remote-sensing algorithms. A detailed description of
the MAS, some MAS applications and field campaigns can be found in King et al. (1996) or at the MAS World
Wide Web site (URLs are listed under the Relevant World Wide Web sites heading). A description of the
MODIS instrument can be found in Barnes et al. (1998) or at the MODIS World Wide Web site.
Test Scene
A MAS scene acquired on 8 April 1995 during the Alaska-April95 aircraft campaign was selected for its
good combination of sea ice types, open water and clouds. The scene is located in the Bering Sea with
a segment of the southwest coast of St. Lawrence Island included (Figure 2)
and has a center location of approximately 63.42° north latitude, 172.10° west longitude. The
flight direction is oriented northeast to southwest. A three channel (MAS channels 1,7,10) color
composite of the scene is shown in Figure 3 and is used as a reference
image on which algorithm results are overlaid.
Sea Ice Algorithm
The primary purpose of the MODIS sea ice algorithm is to map sea ice extent. A potential secondary
purpose is the creation of thematic maps showing generalized sea ice classes. The algorithm identifies
sea ice by reflective characteristics and ice surface temperature, generating a four part data product.
A grouped criteria methodology is employed to identify sea ice on a per pixel basis. The rationale for
using the grouped criteria methodology is that the criteria are robust, can be easily modified or changed
and can be dynamically linked with geographic or seasonal parameters affecting identification of sea ice,
and is amenable to enhancements derived from future studies and independent validation. A flow diagram
of the algorithm is depicted in Figure 4. The University of Wisconsin
cloud mask (UWCM) algorithm was chosen to mask clouds because it will produce the standard MODIS cloud
mask data product. The UWCM algorithm uses select channels in the visible, infrared and thermal channels
in various cloud detection tests (single band thresholds, ratios, channel differences and others) with
threshold rules for detection of cloud. Experience gained with the UWCM using the MAS is directly
applicable to cloud masking with MODIS. A brief background discussion of sea ice reflectance characteristics
and ice surface temperature calculation is presented, followed by descriptions of groups of criteria applied
in the algorithm.
Sea Ice Reflective Characteristics
Observations of sea ice reflectance reported in the literature commonly cover the 0.4 - 1.0 µ
m range (e.g. Perovich, 1990; De Abreu et al., 1995) and sometimes extend out to approximately 2.4
µ m (e.g. Grenfell and Perovich 1984; Perovich et al., 1986). Snow-covered, opaque, white sea ice,
thick first-year ice and multiyear ice typically exhibit maximum reflectance between 0.4 - 0.8 µ m,
decreasing into the near-infrared, reaching a reflectance minimum at approximately 1.6 µ m, then rising
again to a peak at approximately 1.9 µ m. Field observations show that sea ice is predominately
snow-covered (Schlosser, 1988) and that this snow is usually present within a few days of formation
(Perovich, 1990). Young sea ice (e.g., grease ice, nilas, young gray) has lower spectral albedos (in the
range of 0.10 - 0.40) than older sea ice over the 0.4 - 0.8 µ m region. At longer wavelengths
(› 1.0 µ m) the spectral albedo of young sea ice decreases to a value similar to older sea
ice (Schlosser, 1988; Perovich, 1990; Allison et al., 1993). Sea ice undergoing ablation and containing
melt ponds has decreasing reflectance from 0.6 to 0.8 µ m, followed by steady uniform low reflectance
to approximately 1.6 µ m (Grenfell and Perovich, 1984; De Abreu et al., 1995). Reflectance curves
(Figure 5) constructed from the mean reflectance of each MAS band from
two spatially separated 5 x 5 pixel samples band of ice, cloud and water in the scene show reflectance
features consistent with those reported in the sea ice observation studies discussed above. Using these
sample curves from the scene and above studies as background, the sea ice reflectance criteria were
developed to identify snow-covered sea ice and new sea ice.
Ice Surface Temperature
Split-window techniques are employed to determine sea-surface temperature (e.g. McMillin and Crosby,
1984; Minnett, 1990) or ice-surface temperature (e.g. Key and Haefliger 1992; Massom and Comiso, 1994;
Yu et al., 1995; Key et al., 1997). Split-window techniques allow for correction of atmospheric effects,
primarily water vapor, and yield reasonably accurate sea surface temperature (SST) or ice surface
temperature (IST) estimates.
Instrument channels used for estimation of surface temperature with the split-window technique are centered
at approximately 11.0 and 12.0 µ m. Several investigators (e.g. Key and Haefliger, 1992; Lindsay and
Rothrock, 1993; Massom and Comiso, 1994; Yu et al., 1995; Key et al., 1997) have employed variations of the
split-window technique for estimation of IST in polar regions using the AVHRR and other instruments. In most
studies, accuracy of the split-window technique has been enhanced by adding predictor variables to account for
effects such as sensor scan angle. Accuracy of the estimate of IST may be increased by regression of the
estimated IST with temperatures modeled from radiative transfer models or with observed surface temperatures
(e.g. Minnett, 1990; Key and Haefliger, 1992; Lindsay and Rothrock, 1993; Key et al., 1997). IST accuracy
of 0.5 - 1.5 K relative to measured or modeled surface temperatures has been reported by investigators using
the split-window technique.
The technique of Key et al. (1997) was adapted for estimating IST using MAS data. MAS channels 45 and 46,
centered approximately at 11.0 and 12.0 µ m, respectively (Figure 1b),
are used to determine IST. IST for MAS data is estimated using the equation
IST = a + bT45 + c(T45 - T46) + d[(T45
- T46) * sec(theta - 1)], (1)
where
IST -- ice surface temperature (K)
a,b,c,d -- are coefficients determined from multilinear regression of brightness temperatures to estimated surface temperatures (4)
T45 -- brightness temperature of MAS channel 45 (11.0 m m)
T46 -- brightness temperature of MAS channel 46 (12.0 m m)
theta -- sensor scan angle (0 - 43° for MAS)
When MODIS becomes operational, MODIS channels 31 and 32, centered at approximately 11.0 and 12.0 µ m, respectively (Figure 1b), are specified to be used in estimating IST with (1).
Data ProcessingRadiance and conversion data required for calculation of reflectance in visible channels and for calculation of brightness temperature in thermal channels were extracted from the MAS data file. MAS channel radiance data stored as scaled-radiance data, were descaled then converted to reflectance or brightness temperature within the algorithm (Figure 4). Equations used to descale and convert the MAS channel data are described in Gumley et al. (1994). Reflectance was calculated as
Rlambda = Ilambda / Io lambda * cos(theta)(2)
where
Rlambda = reflectance at lambda µ m)
Ilambda = radiance measured in the channel (W m-2 sr-1 µ m-1)
Io lambda = solar spectral irradiance (W m-2 µ m-1)
theta = solar zenith angle.
MAS channels 1, 7, 10, 12 and 20, centered at 0.547 µ m, 0.869 µ m, 1.609 µ m, 1.723
µ m, and 2.1 µ m, respectively, were used for detection of sea ice reflectance characteristics.
Thermal channel data are converted from radiance to brightness temperature by the inversion of Planck's
equation (Gumley et al., 1994): [Eq. 3]
where
T = brightness temperature in K
C1 = 2hc2 = 1.1910439 × 10-16 W m-2
C2 = (hc)/k = 1.4387686 &3215 10-2 m Klambda = wavelength in m
R = Planck radiance (W m-2 sr-1 m-1)
Masking
Land: Land was masked using the 1 km resolution United States
Geological Survey (USGS) global land/water mask (USGS, 1997), as used in the UWCM algorithm. This
mask will be incorporated into the MODIS geolocated sea-ice data product.
Cloud: The UWCM algorithm was applied to the MAS scene with the resulting cloud mask used as
input to the sea ice algorithm (Figure 4). The UWCM initializes all
cloud-test flags to cloudy, creating the initial condition of complete cloud cover. Subsequently, the
algorithm clears the flags if no clouds are detected by the cloud tests. The single unobstructed
field-of-view (FOV) flag, essentially a clear or cloudy flag, was initially used as the cloud mask
for the scene. If the unobstructed FOV flag probability was >= 66% the pixel was interpreted
as clear. Only the flagged condition of "certain cloudy" was interpreted as a cloud.
Criteria Groups
Sea ice by reflectance: The function of this criteria group (Figure 4)
is to identify snow-covered sea ice. Beginning with the assumption that sea ice is snow covered and that
snow dominates the reflectance characteristics, the MODIS snow-mapping techniques for land (Hall et al., 1995;
Riggs et al., 1996) are applied to detect sea ice. The main criterion is the normalized difference snow index
(NDSI). The NDSI is used to detect the high reflectance of snow at visible wavelengths, and the low reflectance
at approximately 1.6 µ m (Figure 5).
NDSI = (visible reflectance - near-infrared reflectance) / (visible reflectance - near-infrared reflectance).
NDSI is calculated with MAS bands 1 and 10; NDSI = (1- 10)/(1+10). NDSI will be calculated with MODIS bands
4 and 6, respectively.
The second criterion in this group is high reflectance in a visible wavelength where snow has a high and
water has a very low reflectance. The purpose of this criterion is to separate water, which has high NDSI
values similar to snow (Hall et al., 1995; Riggs et al., 1996), from snow-covered sea ice. If only the NDSI
criterion was used, water would be confused with snow-covered sea ice. Sea ice is identified where NDSI
>= 0.4 and visible reflectance > 0.17 in MAS channel 1 (0.55 µ m).
Sea ice by IST:The function of this criteria group (Figure 4) is
to identify sea ice by the estimated IST. A vicarious approach was taken to estimate surface temperature
as a surrogate for observed temperature in determination of coefficients for IST (1). Assuming that the
polar atmosphere in the cold season under clear sky conditions is nearly transparent (Yu et al., 1995),
the surface temperature may be estimated directly from the observed brightness at 11.0 µ m (Key et
al., 1997).
At 11.0 µ m, atmospheric effects are minimal, thus a greater signal
from the surface is measured, relative to a channel at 12.0 µ m (Massom and Comiso, 1994).
Surface temperature was estimated directly by the general relationship
Ts = Tb/epsilon¼ (4)
where
Ts = surface temperature
Tb = brightness temperature
epsilon = surface emissivity
The channel chosen for this calculation was MAS channel 45, centered at
11.0 µ m. If the difference of surface temperature at 11.0 and 12.0 µ m
(Ts45 - Ts46 ) is small, then the assumption of a clear atmosphere is reasonable.
In this scene, the mean temperature difference (Ts45 - Ts46 ) for the sea water
samples is 0.39 K. Use of the direct surface temperature estimate (4) gives an estimated surface
temperature adjusted to the assumed emissivity of the surface. Sea water emissivity is assumed to be
0.985 (Comiso, 1994). The mean Ts45 for sea water was estimated to be 271.6 K. Sensor
zenith angle is neglected because the equation is applied to essentially a point sampling of data.
Four five-by-five pixel samples of open water off the coast of St. Lawrence Island were extracted for
estimation of surface temperature and multi-linear regression using (1).
A multi-linear regression using the directly estimated surface temperature Ts45 [Eq. (4)]
and sensor brightness temperatures for sea water was run to determine the coefficients used in the split
window technique (1). Values for the coefficients are;
a = -0.0024
b = 1.0038
c = -1.27 × 10-6
d = 1.87 × 10-5
IST was then determined for the scene using these coefficients in (1).
Arctic sea water can have freezing points over the range 271.36 - 271.45 K
(Yu and Rothrock, 1996). For this scene a freezing point of 271.4 K was assumed based on direct estimates
of open water and ice in the scene. A pixel with a temperature less than 271.4 K was identified as sea ice;
any pixel with a higher temperatures was identified as open water.
Combined: The results of the preceding two criteria groups are combined to show where they agree
and disagree in identifying sea ice. Combination maps where sea ice is identified by both criteria groups,
one criteria group but not the other, and by neither criteria group are created using Boolean and relational
expressions.
Discussion and Results
The sea ice extent and ice type maps are in general agreement with
visual interpretation of sea ice in the scene, and previous studies of sea ice types in the same
region (e.g. Massom and Comiso, 1994). An ocean field campaign was not associated with the acquisition
of the MAS data during the April 1995 campaign, thus analysis is confined to descriptive and comparative
analysis with previous studies. Results of cloud masking are discussed followed by sea ice mapping by
reflectance characteristics and estimated IST.
Discrimination of clouds is particularly difficult in the polar regions because of the lack of contrast
between temperatures and reflectance characteristics of certain cloud types and sea ice. The UWCM product
contains information on the probability of cloud, individual cloud test results and the decision path taken
within the algorithm, on a per pixel basis. The UWCM algorithm has eight conceptual domains that apply
different cloud screening tests for several specified conditions related to geographical location, surface
conditions, and viewing conditions (Ackerman et al., 1996). This approach facilitates adapting the algorithm
for global application while seeking to maximize accuracy of cloud detection. Though four domains and their
cloud screening tests were possible depending on the surface in this MAS scene, we are concerned only with
the daytime ocean domain. Table 1 presents the spectral reflectance (Rlambda)
and brightness temperature (BTlambda) cloud screening tests used in domains appropriate for this
scene.
Using only the unobstructed FOV flag, clouds in the scene were masked. In this scene some areas of ocean
were masked as clouds (Figure 6). Clouds were identified over areas of new
ice or open water e.g., in the lower right quadrant, near the St. Lawrence Island coast, and in the left
third of the scene (Figure 6). Cloud misidentification appeared to occur
in the domain of daytime ocean (column one in Table 1), where the ice/snow background
flag was set to off. Investigation of the individual cloud tests done in that processing path revealed that
the most accurate mapping of clouds was obtained with a single cloud test, the 11 - 4 µ m brightness
temperature difference test. When only this test was used as the cloud mask, clouds were mapped with high
accuracy (Figure 7) as determined from visual analysis of the image.
The source of cloud misidentification is the visible reflectance cloud screening test. This is a single
channel reflectance test for discriminating bright clouds over dark surfaces. The channel and reflectance
threshold used in this test depend on the surface being viewed. In the daytime ocean domain
(Table 1) this visible reflectance test used MAS channel 6 (0.66 µ m) with a
reflectance threshold of 7.0%. Ocean pixels flagged as having no ice/snow background were misidentified as
cloud by this test. No other cloud tests applied in this processing domain were found to misidentify cloud.
The misidentification was caused by greater-than-expected reflectances from the ocean in MAS channel 6. The
mean reflectance of the ocean under the misidentified clouds is 8.1%. Since this exceeds the reflectance
threshold, these pixels were identified as cloud. Two possible reasons for the high ocean reflectances are:
1) the ocean surface was not totally ice-free, or 2) the sea surface was significantly wind-roughened.
Plots of reflectance for features in the scene (Figure 5) indicate that
some areas of the ocean had relatively high reflectance, assuming negligible atmospheric contribution to
the signal. Also confounding the test was the fact that some of the new ice types had reflectances near
those for cloud at visible wavelengths, but not in the infrared. The misidentification of cloud in this
situation does not invalidate the use of the UWCM. Rather, it highlights the need for the user to be aware
of and to make use of the individual cloud tests as necessary.
The ice by reflectance criteria group accurately (by visual interpretation) mapped the extent of sea ice,
snow-covered and new sea ice in the scene (Figure 7). These selected
reflectance criteria were based on sea ice reflectances reported in the literature and observed in the MAS
data. Only sea ice extent was determined by this criteria group. With the many channels on the MAS and the
MODIS sensors, it may also be possible to identify broad sea ice classes. Sea ice reflectance curves
(Figure 5) suggest that two generalized classes, namely new sea ice and
snow-covered sea ice, may be differentiated on the basis of reflectances. These classes are interpreted
to be related to the World Meteorological Organization (WMO) stages of development, with new ice corresponding
to stages that are < 10 cm thick and snow-covered ice corresponding to stages > 10 cm thick. New
sea ice types exhibited reflectances intermediate between snow covered sea ice and water over the 0.5 - 0.9
µ m region (Figure 5). New sea ice reflectance characteristics could
cause confusion with clouds in the scene (Figure 5), however the cloud mask
removed this potential confusion. Another difference between snow-covered sea ice and new ice was noted in
the 1.6 - 1.9 µ m region. In this spectral region, snow-covered sea ice has a local peak in reflectance
(Figure 5) while new sea ice has a relatively uniform low reflectance
(Figure 5).
These observed reflectance differences were used to create thematic maps of new, and snow-covered sea ice.
Two methods of classifying ice were implemented. The intermediate reflectance of new ice over the 0.6 -
0.8 µ m region was approximately 60% lower than snow-covered ice and approximately 30% lower than young
ice types (Figure 5). This intermediate reflectance (25% to 45%) was added
as a criterion to the reflectance criteria group to identify new ice. A pixel meeting both the original
criteria and this new ice criterion is identified as new sea ice. If it meets only the original criteria
it is identified as snow-covered sea ice (Figure 8a).
Also, snow-covered and new sea ice can be differentiated using the near-infrared ratio of MAS channel 12
(1.7 µ m) to channel 10 (1.6 µ m). Snow-covered sea ice has ratio values >= 1.1 and
new ice has ratio values of < 1.1. This criterion test (snow-covered sea ice if ratio >=
1.1 and new ice if < 1.1), along with a threshold test in MAS channel 1 (0.55 µ m) can be used
separate water from new ice and was implemented as a separate criterion group. This group resulted in a
reasonably accurate map of snow-covered sea ice (Figure 8b).
Both maps of sea ice classes (Figure 8a and Figure 8b)
agree with visual interpretation of ice type, and are in general agreement with general patterns of sea ice
extent and type for the region as reported in previous studies (e.g. Massom and Comiso, 1994). Absence of
in situ observations limit these results to qualitative discussion. However, they do show that
additional information about sea ice beyond sea ice extent is possible with the MAS and MODIS sensors. For
the MODIS sea ice algorithm only the criteria groups and tests given in Figure 4
will be used in the first version after launch.
Sea ice extent (Figure 9) from the IST criteria group was in agreement with
interpreted sea ice extent in Figure 3. The sea ice extent map created by
flagging any pixel with an IST less than 271.4 K fails to make full use of the IST information. Ice
classification is possible using IST because of the marked thermal contrast between open water, new ice and
first-year and older sea ice types (Massom and Comiso, 1994). Also, the surface temperature of ice is
strongly related, in principle, to ice thickness, allowing inference of ice thickness (Yu and Rothrock,
1996). Based on this information, a hypothetical three-classification sea ice map was created using the
temperature ranges of the classes shown in Table 2. The ice-type map for this
classification scheme is shown in Figure 9. New ice and young ice types
are the dominant ice types identified. Some first-year ice was identified along the coast of St. Lawrence
Island which is a mix of both fast ice and snow-covered land which was not masked by the land mask. Without
surface observations this classification cannot be verified, but it is a reasonable classification for the
region based on descriptions in Massom and Comiso (1994) for the area around St. Lawrence Island for the
same time of year during 1988. The observed consistency of sea ice type and distribution as determined by
the estimated IST method between the studies is encouraging.
Independent validation activities in support of the MODIS sea ice algorithm will be performed to
determine the coefficients, modify the IST equation, and possibly to include dynamically-adaptive
coefficients that can be adjusted for specific geographical regions, seasons, surface temperatures, or
a combination of parameters in order to increase accuracy of the IST.
The motivation for combining the results of the two criteria groups is to enhance analysis and
identification of sea ice, primarily in areas of new-ice formation when the reflectance characteristics of
new sea ice make it difficult to detect and where the surface temperature is near the freezing point of sea
water. Mapped sea ice extent was different in the reflectance and IST criteria groups. There are areas in
the scene, (e.g. along the left third and right middle third of the scene), that appear to be open water in
the color-composite image (Figure 3) that are not identified as ice by the
reflectance criteria group for ice (Figure 7) but are identified as new ice
by the IST (Figure 9) classification. Differences in sea ice extent between
the two techniques are areas where there may be a mix of sea ice and open water or where other factors were
affecting the surface characteristics. Horizontal striping in Figure 9 and
Figure 10 suggest sensor performance effects in the thermal bands. The most
certainty in the sea ice extent occurs where the two identification techniques (reflectance and IST
(Figure 4)) agree (Figure 10).
Summary
Using reflectance and thermal data from MODIS or MAS it is possible to map
sea ice extent and possibly classify sea ice type. The UWCM provided a reasonably accurate masking of clouds
over sea ice when only specific cloud tests were considered in this analysis. Thus use of the cloud mask in
polar regions may require an in-depth analysis of the cloud mask data content to maximize its effectiveness
and accuracy. Albeit limited in scope, this work has been used to develop the MODIS sea ice algorithm.
Algorithm development and analysis continues with MAS data, including comparison of results with SSM/I derived
sea ice extent. It is anticipated that a similar capability for mapping sea ice will be obtained with the
MODIS sea ice algorithm.
The MODIS sea ice algorithm is developed with the same criteria described here for mapping sea ice with
MAS data (Figure 4) and is fully automated. Classification of sea ice
type is not yet included in the MODIS algorithm. The algorithm will be bootstrapped with actual MODIS
data using the coefficients reported here or with better determined coefficients from other studies with
MAS. Initial estimated IST accuracies with MODIS data are anticipated to be poor and the coefficients
adjusted frequently during the first year to improve accuracy. Validation studies are expected to be
used to identify adjustments to the coefficients and to define sets of coefficients for use in specific
regions (e.g. regions within the Arctic, or Antarctica) and possibly seasonal subsets for regions. Sea
ice will be mapped both day and night with MODIS, though at night only the IST will be used. The MODIS
cloud mask will be used for both day and night products. The UWCM does night time detection of clouds
using thermal tests, however, performance of the cloud mask at night over sea ice is still being investigated
and is expected to be lower than its performance during daytime.
The MODIS sea ice data product will be archived in Hierarchical Data Format-EOS (HDF-EOS) composed of
scientific data sets and metadata. Sea ice derived from reflectance characteristics, estimated IST, sea
ice extent by IST and combined sea ice extent (Figure 4) are archived as
scientific data sets in the data product. Night time products will have only IST arrays. Extensive
metadata (information about the data product) supporting EOSDIS Core System (ECS) services and containing
quality assurance information and product summary information is also included with each product. Sea ice
products are to be archived at the National Snow and Ice Data Center (NSIDC) in Boulder, Colorado (Scharfen
et al., 1997). Sea ice products at 1 km spatial resolution are planned for swaths ("scenes") day and night
of MODIS data. Gridded daily products at 1 km and 0.25° spatial resolution in Lambert azimuthal
equal-area polar projection, and composite products of eight day and monthly at 0.25° spatial
resolution are planned.
Acknowledgments
Thanks to N. Larsen (U. Alaska, Fairbanks) for thought-provoking discussions throughout the course of this project, and to A. Klein (Texas A &M University) for assistance with image production and review of the manuscript. Thanks to K. Strabala, R. Frey and L. Gumley for their important roles in coding the MAS cloud algorithm. Thanks also to anonymous reviewers for useful comments and suggestions that much improved the manuscript.
Relevant World Wide Web sites:References
Ackerman, S., Strabala, K., Menzel, P., Frey, R., Moeller, C., Gumley, L., Baum, B., Schaaf, C., Riggs, G., and Welch, R. (1996), Discriminating clear-sky from cloud with MODIS, Algorithm Theoretical Basis Document (MOD35), Version 3.0, 109 pp.
Allison, I., Brandt, R. E., and Warren, S. G. (1993), East Antarctica sea ice: albedo, thickness distribution, and snow cover. J. Geophys. Res. 98(C7):12,417-12,429.
Barnes, W. L., Pagano, T. S., and Salomonson, V. V. (1998), Prelaunch characteristics of the Moderate Resolution Imaging Spectroradiometer (MODIS) on EOS-AM 1. IEEE Trans. Geosci. Remote Sens. 36(4);1088-1100.
Carsey, F. D., Barry, R. G. and Weeks, W. F. (1992), Chapter 1. Introduction, In: Microwave Remote Sensing of Sea Ice (F. D. Carsey, Ed.), Geophysical Monograph 68, AGU, pp. 1-7.
Comiso, J. C. (1994), Surface temperatures in the polar regions from Nimbus 7 temperature humidity infrared radiometer. J. Geophys. Res. 99(C3):5,181-5,200.
De Abreu, R. A., Barber, D. G., Misurak, K., and LeDrew, E. F. (1995), Spectral albedo of snow-covered first-year and multiyear sea ice during spring melt. Ann. Glaciol. 21:337-342.
Grenfell, T. C., and Perovich, D. K. (1984), Spectral albedos of sea ice and incident solar irradiance in the Southern Beaufort Sea. J. Geophys. Res. 89(C3): 3,573-3,580.
Gumley, L. E., Hubanks, P. A., and Masuoka, E. J. (1994), MODIS Technical Report Series, Volume 3, MODIS Airborne Simulator Level 1B Data User's Guide, NASA Tech. Memo. 104594, Vol.3, 37 pp.
Hall, D. K., Riggs, G. A., and Salomonson, V. V. (1995), Development of methods for mapping global snow cover using Moderate Resolution Imaging Spectroradiometer data. Rem. Sens. Environ. 54:127-140.
Key, J. R., and Haefliger, M. (1992), Arctic ice surface temperature retrieval from AVHRR thermal channels. J. Geophys. Res. 97(D5): 5,885-5,893.
Key, J. R., Collins, J. B., Fowler, C., and Stone, R. S. (1997), High-latitude surface temperature estimates from thermal satellite data. Rem. Sens. Environ. 61:302-309.
King, M. D., Menzel, W. P., Grant, P. S., Myers, J. S., Arnold, G. T., Platnick, S. E., Gumley, L. E., Tsay, S., Moeller, C. C., Fitzgerald, M., Brown K. S., and Osterwisch, F. G. (1996), Airborne scanning spectrometer for remote sensing of cloud, aerosol, water vapor, and surface properties. J. Atmos. and Ocean. Technol. 13:777-794.
Lindsay, R., and Rothrock, D. (1993), The calculation of surface temperature and albedo of Arctic sea ice from AVHRR. Ann. Glaciol. 17:174-183.
Massom, R., and Comiso, J. C. (1994), The classification of Arctic Sea ice types and the determination of surface temperature using advanced very high resolution radiometer data. J. Geophys. Res. 99(C3):5201-5218.
McMillin, L. M., and Crosby, D. S. (1984), Theory and validation of the multiple window sea surface temperature technique. J. Geophys. Res. 89:3655-3661.
Minnett, P. J. (1990), The regional optimization of infrared measurements of sea surface temperature from space. J. Geophys. Res. 95:13497-13510
Parkinson, C. L., and Gloersen, P. (1993), Global sea ice coverage, In: Atlas of Satellite Observations Related to Global Change. (R. J. Gurney, J. L. Foster, and C. L. Parkinson, eds.) Cambridge University Press, pp. 371-384.
Perovich, D. K. (1990), The evolution of sea ice optical properties during fall freeze-up. SPIE Vol. 1302, Ocean Optics X. 520-531.
Perovich, D. K., Maykut, G. A., and Grenfell, T. C. (1986), Optical properties of ice and snow in the polar oceans. I: observations. SPIE Proceedings, Vol. 637, Ocean Optics VIII. 232-241.
Riggs, G. A., Hall, D. K., and Salomonson, V. V. (1996), Recent progress in development of the Moderate Resolution Imaging Spectroradiometer snow cover algorithm and product. In 1996 International GeoScience and Remote Sensing Symposium, Remote Sensing for a Sustainable Future, 27-31 May, 1996, Lincoln, NE. Vol 1:139-141.
Scharfen, G. R., Hall, D. K., and Riggs, G. A. (1997), MODIS snow and ice products from the NSIDC DAAC. SPIE Proceedings, Vol. 3117, In Earth Observing Systems II, (W.L. Barnes, ed.) 143-147.
Schlosser, E. (1988), Optical studies of Antarctic sea ice. Cold Regions Sci. Tech. 15:289-293.
United States Geological Survey. (1997), EROS Data Center Distributed Active Archive Center, http://edcwww.cr.usgs.gov/landdaac/1KM/land_sea_mask.html.
Washington, W. M., and Parkinson, C. L. (1986), An Introduction to Three-Dimensional Climate Modeling. University Science Books, Mill Valley, CA, 422 pp.
Yu, Y., and Rothrock, D. A. (1996), Thin ice thickness from satellite thermal imagery. J. Geophys. Res. 101:25,753-25,766.
Yu, Y., Rothrock, D. A., and Lindsay, R. W. (1995), Accuracy of sea ice temperature derived from the advanced very high resolution radiometer. J. Geophys. Res. 100(C3):4,525-4,532.
List of Tables
Table 1. UWCM domains and cloud screening tests used for the MAS scene. The spectral reflectance (R) and brightness temperature (BT) are listed. An × mark indicates that a test was done for that domain.
|
Domains |
||||
|---|---|---|---|---|
|
Cloud Screening Tests |
Daytime |
Daytime |
Snow |
Coastline |
|
BT11 |
× |
|||
|
BT13.9 |
× |
× |
× |
× |
|
R1.38 |
× |
× |
× |
× |
|
BT8 - BT11 and BT11 - BT12 |
× |
× |
× |
|
|
BT3.7 - BT11 |
× |
× |
× |
× |
|
R0.66 |
× |
× |
× |
|
|
R0.87/R0.66 |
× |
× |
× |
|
|
Spatial Variability |
× |
|||
Table 2. Sea ice classes by temperature range.
| Class | Temperature Range |
| open water | IST >= 271.4 K |
| new ice | 271.4 K < IST <= 268.0 K |
| young ice | 268.0 < IST < 262.0 K |
| first-year ice | IST <= 262.0 K |
List of Figures
Figure 1
|
Figure 2
|
Figure 3
|
Figure 4
|
Figure 5
|
Figure 6
|
Figure 7
|
Figure 8a
|
Figure 8b
|
Figure 9
|
Figure 10
|