Introduction
The freeze–thaw (FT) signal detected from moderate frequency (37 GHz)
satellite microwave brightness temperature (Tb) observations is
sensitive to changes in the relative abundance of liquid water at the land
surface between frozen and non-frozen conditions. The FT metric defines the
predominant landscape frozen or non-frozen status within a sensor footprint
and is insensitive to potential degradation from solar illumination
variations and atmosphere cloud/aerosol contamination effects. These
properties enable consistent global coverage and daily monitoring from
available satellite microwave sensors (Kim et al., 2011; Zhao et al., 2014).
Global satellite microwave sensors have been operational since the
mid-1970s, while similar Tb observations from overlapping sensor
records have enabled the development of one of the longest satellite
environmental data records, delineating global daily FT dynamics and
associated climate trends over multiple decades (Kim et al., 2011). The FT
Earth system data record (FT-ESDR) was developed to provide a consistent and
continuous long-term daily global FT data record to support climate,
ecological and hydrological applications. The initial FT-ESDR domain
encompassed global vegetated land areas where seasonal frozen temperatures
are estimated to be a significant constraint to annual vegetation growth (Kim
et al., 2011). The FT retrieval is derived on a grid-cell-wise basis using a
temporal change classification of similar daily 37 GHz Tb observations
from Scanning Multichannel Microwave Radiometer (SMMR), Special Sensor
Microwave Imager (SSM/I) and SSM/I Sounder (SSMIS) sensor series, in relation
to frozen and non-frozen Tb reference states (Kim et al., 2012).
The FT-ESDR has been widely used to define frozen temperature constraints to
ecological and hydrological processes, as well as for monitoring seasonal
shifts in land surface energy partitioning where frozen temperatures are a
significant part of the annual cycle (Mao et al., 2015; Zhang et al., 2014;
Zhu et al., 2016). Previous FT-ESDR applications have included documenting
widespread increases in the annual thaw cycle and a general reduction in
frozen temperature constraints to vegetation growth over the Northern
Hemisphere from regional climate warming (Kim et al., 2012; Wang et al.,
2015), analyzing spring thaw- and frost-related impacts on vegetation
phenology and productivity (Buermann et al., 2013; Kim et al., 2014a, b), and
defining a potential growing season climate indicator for the US National
Climate Assessment (Kenney et al., 2014). The FT-ESDR has also been used to
document FT-related impacts to surface hydrology, including
evapotranspiration and the timing and extent of seasonal snowmelt (Zhang et
al., 2011; Kim et al., 2015), and river ice phenology (Park et al., 2016b).
The FT-ESDR-defined frozen season was also used to estimate permafrost extent
and stability over the pan-Arctic (Park et al., 2016a).
In this investigation, we utilize calibrated daily (morning (AM) and evening (PM) overpass)
37 GHz Tb retrievals from SMMR, SSM/I and SSMIS sensor series to
refine and extend the FT-ESDR. New FT-ESDR enhancements include a larger
global domain encompassing all land areas affected by seasonally frozen
temperatures, including vegetated, barren, urban, ice- and snow-cover-dominant, and open-water-dominant areas. We exploit SMMR and continuing SSM/I
operations to develop a longer data record (1979–2014) than prior FT-ESDR
releases (Kim et al., 2011) extending over a shorter 20-year record
(1988–2010). A modified seasonal threshold algorithm (MSTA) is used to
classify daily FT status in relation to a Tb-based FT threshold
calibrated annually for each grid cell using daily surface air temperature
(SAT) records from global model reanalysis. The FT classification accuracy is
evaluated using independent daily SAT measurements from the global weather
station network. Annual quality assessment (QA) maps are derived using
empirical models describing estimated global variations in FT classification
accuracy due to climate and terrain heterogeneity, open water cover and model
calibration uncertainty. The FT-ESDR performance is also evaluated using
other cryosphere data records including in situ lake and river ice
phenology observations and satellite-derived Greenland surface melt dynamics.
The FT-ESDR methods, database description and performance assessment are
provided in the following sections.
Data and methods
Satellite microwave brightness temperature observations
The Nimbus-7 SMMR Pathfinder dataset contains Tb retrievals from
October 1978 to August 1987. The SSM/I and SSMIS sensor records are part of
the Defense Meteorological Satellite Program (DMSP) platform series and
provide continuous global, twice-daily Tb measurements from August
1987 to the present. The SMMR sensor obtained Tb retrievals at
multiple frequencies (6.6, 10.7, 18.0, 21.0, and 37.0 GHz), with vertical
(V-) and horizontal (H-) polarization (pol). The SMMR Tb
measurements were obtained at a constant 50.31∘ incidence angle and
noon, as well as midnight equatorial crossings (i.e., overpass times). The SSM/I and
SSMIS sensors obtain V- and H-pol Tb measurements at 19.3, 37.0 and
85.5 GHz, and 22.2 GHz Tb measurements at V-pol only. The SSM/I
and SSMIS instruments have a constant 53.1∘ measurement incidence
angle and approximate 06:00/18:00 equatorial crossings.
Calibrated, overlapping 37 (V-pol) GHz Tb data records used for
classifying FT status were acquired from SMMR (Knowles et al., 2000), SSM/I
and SSMIS (Armstrong et al., 1994) in a consistent 25 km grid cell
resolution global Equal-Area Scalable Earth Grid (EASE-Grid) v1.0 projection
format (Brodzik and Knowles, 2002). Missing Tb data composed
approximately 34.3 ± 24.3 [inter-annual SD] % of frozen
temperature-affected land areas on an annual basis for the 1979–2014 record.
Missing Tb data attributed to orbital gaps between satellite
overpasses, particularly for the midlatitudes, were filled on a
grid-cell-wise basis by linear interpolation of temporally adjacent,
successful Tb retrievals to generate spatially and temporally
consistent daily (AM and PM overpass) Tb observations as inputs for
the landscape FT classification following previously established methods (Kim
et al., 2011). The SMMR record was matched to the SSM/I record using
pixel-wise adjustment of the SMMR Tb record based on empirical
analyses of overlapping SMMR and SSM/I Tb measurements for 1987 to
ensure cross-sensor consistency (Kim et al., 2012). Large temporal
Tb gaps in January and December 1987 and January 1988 were filled
using empirical relationships developed from ERA-Interim (Dee et al., 2011)
global model reanalysis SAT and satellite Tb data records. The
resulting daily Tb record was used to construct the FT-ESDR over an
extended 36-year period (1979–2014).
For this investigation, the FT-ESDR defines the predominant frozen or
non-frozen condition of the landscape and does not distinguish among
individual vegetation, snow and soil FT elements within a grid cell (Kim et
al., 2011, 2012). The satellite microwave Tb retrievals represent
the mean emitted radiation from the integrated land surface within the grid
cell. The 37 GHz Tb retrievals used in the FT-ESDR have relatively
high atmospheric transmittance and are sensitive to land surface FT
variations (Andre et al., 2015; Holmes et al., 2013), while the V-pol
Tb retrievals are also relatively less sensitive to soil moisture
variation, snow and ice stratification, and scattering than the H-pol
retrievals (Green et al., 2012; Holmes et al., 2009; Owe et al., 2008). The 37 GHz
Tb sensitivity to landscape FT variations from surface soil, snow
cover and vegetation elements within the sensor footprint is related to a
relatively shallower characteristic microwave emitting depth than lower
frequency Tb retrievals (Green et al., 2012), while the 37 GHz
emitting depth is generally greater under drier surface conditions (Ulaby and
Long, 2014). Higher moisture levels in soil and vegetation layers reduce the
effective depth of microwave FT sensitivity, while lower microwave
frequencies have generally greater depth of FT sensitivity (Watanabe et al.,
2011).
Global FT-ESDR domain
We used 0.25∘ resolution ERA-Interim daily minimum SAT
(SATmin) estimates and a simple cold temperature constraint index
(CCI) to define the global FT-ESDR domain as all land areas where seasonally
frozen temperatures influence ecological processes and land surface water
mobility (Kim et al., 2011). The ERA-Interim global reanalysis was selected
for the FT-ESDR domain definition because it contains consistent atmospheric
and surface analyses with similar spatial and temporal resolution as the
satellite Tb records. The 36-year (1979–2014) 0.25∘
ERA-Interim daily SATmin global record was resampled to a 25 km
global EASE-Grid format. The CCI (days year-1) was derived for each
global EASE-Grid cell using a bioclimatic growing season index and
SATmin as developed in Kim et al. (2011). The CCI values were
derived for each year of the 36-year ERA-Interim record on a per grid cell
basis and used to identify cells constrained by seasonal frozen temperatures.
The FT-ESDR domain was defined where the CCI exceeded a minimum threshold
(5 days year-1) from the 36-year average CCI record following
previously developed methods (Kim et al., 2011). The prior FT-ESDR global
domain encompassed only vegetated land areas where seasonally frozen
temperatures are a major constraint to annual vegetation growth; the domain
excluded grid cells where the areal open water cover fraction exceeded
20 % or was dominated by other non-vegetated land cover categories. The
new FT-ESDR domain described in the current investigation is 41.5 %
(∼ 27 million km2) larger than the prior record and encompasses
all frozen temperature-affected land areas (CCI > 5 day year-1),
including vegetated, urban, open water, snow-ice-dominant, and barren
landscapes defined from a static global IGBP land cover classification
(Friedl et al., 2010).
Freeze–thaw algorithms
A seasonal threshold algorithm (STA) was applied to classify daily FT state
dynamics from SSM/I daily 37 GHz (V-pol) Tb time series to
construct the prior FT-ESDR release (Kim et al., 2011). The STA used a
spatial and seasonal scale factor, ΔTb(x,t), defined for
an observation acquired at location (x) and time (t) as
ΔTbp(x,t)=(Tbp(x,t)-FrozRef(x))(ThawRef(x)-FrozRef(x)),
where Tbp(x,t) [K] is the Tb retrieval acquired at
location (x), time (t), and polarization (p); FrozRef(x) [K] and
ThawRef(x) [K] define Tb retrievals under respective frozen and
thawed landscape reference states. The STA threshold was derived annually on
a grid-cell-wise basis using an empirical linear regression relationship
between the satellite ΔTb(x,t) and daily SAT records from
the National Centers for Environmental Prediction and National Center for
Atmospheric Research reanalysis (Kanamitsu et al., 2002).
In the latest FT-ESDR version described in this paper, a new modified
seasonal threshold algorithm (MSTA) is used to classify daily (AM and PM) FT
status from similar calibrated 37 GHz (V-pol) Tb time series from
SMMR, SSM/I and SSMIS records. The MSTA is defined as
FTstatus=Non-frozen,ifTb>thresholdFrozen,ifTb≤threshold,
where the MSTA FT threshold was defined annually for each grid cell using an
empirical linear regression relationship between the satellite Tb
retrievals and daily SAT estimates from the ERA-Interim global reanalysis.
The ERA-Interim SAT record was selected for the MSTA calibration because the
resulting FT record produced the best global accuracy performance relative to
alternative reference states defined from other reanalysis datasets and
satellite (MODIS) land surface temperature (LST) records (not shown). The
annual grid-cell-wise FT thresholds were derived separately for the AM and PM
overpass Tb retrievals using corresponding daily SAT minimum
(SATmin) and maximum (SATmax) values. Larger weighting of SAT
values closer to 0 ∘C was used in selecting the corresponding
Tb-based FT threshold for each grid cell; weighting of the SAT and
Tb regression relationship was derived using a cosine function
within a temperature range extending from -60.0 to 30.0 ∘C and
representing 99 % of the SAT frequency distribution defined from the
36-year ERA-Interim SAT global climatology. The resulting AM and PM FT
classifications were combined into a daily composite (CO) with four discrete
classification levels, including frozen (AM and PM frozen), non-frozen (AM
and PM non-frozen), transitional (AM frozen; PM thawed), and
inverse-transitional (AM thawed; PM frozen) states.
The MSTA approach assumes that the large changes in microwave dielectric
constant of the land surface that occur around the 0 ∘C temperature
threshold are associated with landscape FT transitions that dominate the
corresponding satellite 37 GHz Tb seasonal dynamics, rather than
other potential sources of Tb variability (Kim et al., 2011). An
advantage of the MSTA over the previous STA approach (Kim et al., 2011) is
that the Tb threshold selection does not depend on frozen and
non-frozen reference states derived by averaging Tb measurements
over respective winter and summer periods, and is less sensitive to
Tb data gaps during these reference periods. The grid-cell-wise
Tb thresholds and annual calibration used for the MSTA reduce the
potential influence of spatial and temporal variations in climate and land
surface conditions on FT classification accuracy, and promote greater
consistency in global product performance over the long-term record. The
daily SAT record used for the MSTA calibration approximates FT-related shifts
in near-surface air temperatures using a fixed 0.0 ∘C temperature
threshold. In contrast, the 37 GHz Tb retrievals used to construct
the FT-ESDR are sensitive to land surface dielectric variations stemming from
FT-related shifts in surface soil, snow cover and vegetation elements within
the sensor footprint. The SAT and Tb records represent different
physical phenomena with different spatial and temporal characteristics; these
factors may contribute uncertainty to the MSTA thresholds and classification
accuracy, which are represented in the FT-ESDR accuracy quality assurance
metrics (Sect. 2.4).
FT classification agreement and quality assurance
We used independent in situ daily SATmin and SATmax
measurements from the global weather station network to assess the FT-ESDR
classification agreement following previously developed methods (Kim et al.,
2011, 2012). The WMO weather station daily SAT measurements were obtained
from the National Climate Data Center (NCDC) Global Summary of the Day
(National Weather Services, 1988). The Euclidian distance between each
FT-ESDR EASE-Grid cell centroid and WMO station locations was computed to
select a single representative station closest to the center of a grid cell
when two or more stations were located within the same EASE-Grid cell. Daily
SATmin and SATmax records for the selected stations were used to
define daily frozen (T ≤ 0 ∘C) and non-frozen
(T > 0 ∘C) temperature conditions and compared with
corresponding FT classification results from the overlying EASE-Grid cells
and respective AM and PM overpass periods, assuming that the local timing of
daily SATmin and SATmax occurs near the satellite equatorial
crossing times (Frolking et al., 1999; Kim et al., 2012). The FT
classification agreement was assessed through grid-cell-to-point comparisons
between WMO daily SAT measurements and overlying FT-ESDR results. Previous
studies indicate close correspondence between the SAT and 37 GHz (V-pol) FT
dynamics (Kim et al., 2011, 2012), which characterize the FT state of the
topmost surface layer within a 25 km satellite footprint, including bare
soil, snow, and vegetation elements (Owe and Van De Griend, 2001).
Approximately 4153 ± 632 [inter-annual SD] weather stations were
selected for the 1979–2014 period. Differences between the SAT measurement-based FT value and the corresponding FT-ESDR estimate from the overlying grid
cell were used to define the global daily FT spatial classification accuracy,
expressed as a proportion (%) of the number of stations and overlying
FT-ESDR grid cells with identical FT classification results relative to the
total WMO station cells represented within the FT-ESDR domain on an annual,
monthly and daily basis.
The mean annual FT classification agreement between WMO station SAT
observations and FT-ESDR classification values was evaluated according to the
dominant land cover characteristics within a grid cell, sub-grid-scale
terrain heterogeneity and open water body fraction within a grid cell, and
other factors potentially influencing FT classification agreement. The MODIS
(MCD12C1, collection 051) 5.6 km, 17-class IGBP global land cover product
for the 2001–2012 MODIS record (Friedl et al., 2010) was used to define
dominant land cover categories within each 25 km resolution FT-ESDR grid
cell using a majority approach (de Jong et al., 2013; Dendoncker et al.,
2008). A 300 m water body map for
the 2010 epoch (from 2008 to 2012) produced from the European Space
Agency
(Defourny et al., 2016) was used to derive fractional open water coverage
(Fw, %) within each 25 km FT-ESDR grid cell; the Fw
map only defines static open water body cover and does not account for
temporally dynamic variations in surface inundation. A 1 km resolution
digital elevation model (DEM; Hasting et al., 1999) was used to derive an
elevation gradient (m), defined as the spatial standard deviation (SD) of the
elevation distribution within each 25 km FT-ESDR grid cell. Drop-in-bucket
averaging of finer-resolution (300 m and 1 km) pixels was used to resample
to the 25 km resolution EASE-Grid format.
An FT-ESDR quality assurance (QA) map was determined on an annual basis from
multivariate linear regression analysis of global annual FT classification
agreement with WMO weather-station-based FT results and four independent
geospatial variables. The independent variables used in the QA analysis for
each grid cell included the Fw, elevation gradient, mean annual
duration of transitional FT conditions, and the correlation (r value)
between weighted ERA-Interim reanalysis SAT and satellite Tb time
series used to define the MSTA FT threshold. The resulting QA calculations
defined the mean annual classification accuracy (%) for each grid cell and
year of record, and were linearly rescaled between low (0) and best (1)
relative quality values to define a final dimensionless QA metric. Quality
control (QC) flags were used to identify other potential factors affecting FT
classification agreement, including grid cells and days with missing and
interpolated Tb observations, precipitation events (Ferraro et al.,
1996), extensive open water bodies (Fw > 0.2) or complex
terrain (elevation gradient > 300 m).
Comparison of FT metrics with independent cryosphere data
records
Three FT metrics were derived from the daily CO FT classification series for
each year of record, including frozen season (FS) duration, primary spring
thaw date, and non-frozen season (NFS) duration. The FS duration was defined
from the daily FT-ESDR as the total number of frozen or transitional (AM
frozen, PM non-frozen) days per year (September–August), which is similar to
number of frost days reported by Peterson (2005). The primary spring thaw
date was determined as the first day (DOY) when 12 out of 15 consecutive days
from January through June were classified as non-frozen (Kim et al., 2012).
The NFS was defined as the number of classified non-frozen days for a
calendar year (January–December). The resulting FT metrics were compared
against other independent cryospheric data records, including lake and river
ice duration observations from the Global Lake and River Ice Phenology
Database from 1979 to 2013 (Benson and Magnuson, 2000), reported annual ice
breakup dates for the Tanana River in the interior of Alaska (Nenana Ice
Classic, 2011), and the NASA MEaSUREs Greenland surface melt record (Mote,
2014). The annual duration of lake and river ice was determined for each
observation location as the period between reported seasonal freezing and
thawing, and compared with the collocated FT-ESDR-based FS estimates. The
reported Tanana River ice breakup dates from 1979 to 2003 were compared
against the FT-ESDR-derived mean primary spring thaw dates for the
surrounding basin defined within the overlying 5×5 grid cell
(∼ 15 625 km2) window.
The Greenland surface melt record was available from 1979 to 2012 at the time
of this investigation and was compared against corresponding FT-ESDR-derived
annual NFS results. The Greenland melt record is derived using similar
satellite 37 GHz Tb retrievals as the FT-ESDR, but the
Tb retrievals are used with a microwave emission model to
distinguish either surface melt or no surface melt categorical conditions for
each 25 km grid cell (Mote, 2014). The seasonal progression in proportional
area and annual variation of surface melt and FT-ESDR defined NFS conditions
were compared over the Greenland ice sheet area indicated from the
Land-Ocean-Coastline-Ice database (Knowles, 2004). The Greenland melt record
provides a synergistic geospatial database for FT-ESDR assessment over
northern snow- and ice-dominant land areas where in situ weather station
networks are extremely sparse.
(a) Mean annual frozen season (frozen or transitional status)
and (b) temporal standard deviation of the frozen season over the 36-year
(1979–2014) record and global FT-ESDR domain; white and grey denote
respective open water bodies and land areas outside of the FT-ESDR domain;
grey areas depict grid cells where the mean cold temperature constraint index
(CCI) was less than 5 days year-1 from the 36-year CCI record.
![]()
Results
Global FT-ESDR domain and frozen season characteristics
The resulting FT-ESDR domain represents ∼ 60.5 %
(∼ 93 million km2) of the global land area, compared to
∼ 52.5 % for prior FT-ESDR releases (Kim et al., 2011). The larger
global domain is shown in Fig. 1 and encompasses vegetation-dominant grid
cells, as well as urban, large water body (Fw < 100 %),
complex terrain, permanent snow/ice cover and barren land areas excluded from
prior FT-ESDR releases. The maps in Fig. 1 show the mean annual frozen season
(FS; days) and FS temporal SD derived from the 36-year (1979–2014) FT-ESDR
record. The FS duration is approximately 151 ± 82 [spatial SD] days and
201 ± 167 days for respective Northern Hemisphere (NH) and Southern
Hemisphere (SH) domains. The FT-ESDR captures characteristic FS increases at
higher elevations, including Tibetan Plateau, Andes and Rocky Mountain areas,
though finer spatial-scale FT variations over complex terrain are less
distinct due to the coarse (25 km)-resolution Tb retrievals. The
FT-ESDR also depicts characteristic FS increases at higher latitudes, with
maximum FS durations over polar permanent ice and snow areas of Greenland and
Antarctica. The FS is generally shorter along coastal margins relative to
inland areas, consistent with more moderate maritime climate conditions in
the coastal zone. Annual FS variability is generally greater along the zones
of transition between major air masses and associated climate regimes. Higher
FS variability is also evident for coastal grid cells and warmer climate
zones, which may reflect greater FT classification uncertainty in these areas
as indicated from the FT classification and quality assessments described
below.
FT classification assessment
The mean annual FT spatial classification agreement over the global domain
and 36-year record was 90.3 ± 1.4 [inter-annual SD] % and
84.3 ± 1.7 [inter-annual SD] % for respective PM and AM
overpasses (Fig. 2) as derived from daily grid-cell-to-point comparisons
against in situ SAT measurement-based FT observations from 4153 ± 632
[inter-annual SD] global WMO weather stations. These results showed general
improvement in mean annual classification accuracy of 0.6 % (PM overpass)
and 1.9 % (AM overpass) over prior FT-ESDR versions for the same period
(1979–2012) and vegetated domain covered by the initial product release (Kim
et al., 2011). Areas with improved accuracy include Central Asia, northern
and central Europe, and South America (not shown). The mean annual FT spatial
classification agreement for 1987 and 1988 was lower than other years of
record due to large temporal Tb gaps in December 1987 and January
1988 coinciding with the SMMR to SSM/I sensor transition period. The FT
classification results show positive trends (p<0.001) toward increasing
mean annual classification agreement of 1.4 and 1.1 % decade-1 for
respective AM and PM overpasses over the 1979–2014 record. The increase FT
accuracy coincides with a 0.3 ∘C decade-1 mean SAT warming
trend over the FT-ESDR domain and may be due to a corresponding decline in
frozen conditions (Kim et al., 2012). The apparent FT accuracy trend may also
reflect differences between SMMR and SSM/I portions of record, including
variations in the location and number of weather stations used for validation
and different overpass times between SMMR (noon/midnight) and SSM/I
(∼ 06:00/18:00) Tb records. The lower FT classification
agreement from the AM overpass Tb record is attributed to generally
cooler diurnal temperature conditions and a corresponding larger number of
daily FT variations over the annual cycle in the AM record relative to
characteristic warmer midday temperature conditions and fewer FT variations
in the PM results. Lower AM FT classification agreement may also be due to
wetter surface conditions and temperature inversions during night-time
diurnal SAT minimums (Owe and Van De Griend, 2001).
Mean annual FT spatial classification agreement (%) between
FT-ESDR results and corresponding daily SAT-based FT estimates from the
global WMO weather station network. The FT classification agreement was
determined on a daily basis for individual satellite AM and PM overpasses in
relation to grid-cell-to-point comparisons with independent FT estimates
derived from respective minimum and maximum daily SAT records. The bar graph
denotes the number of global WMO stations used for FT assessment.
![]()
Mean monthly FT classification agreement (%) for AM and PM
overpasses in relation to WMO surface air temperature (SATmax and
SATmin)-based FT metrics for Northern Hemisphere (NH), Southern
Hemisphere (SH) and global domains. Values represent mean monthly spatial
agreement (%) and temporal standard deviation for the 1979–2014 record.
PM overpass
AM overpass
NH
SH
Global
NH
SH
Global
Jan
81.0 ± 1.9
96.3 ± 1.8
81.9 ± 2.0
80.0 ± 1.6
95.8 ± 2.0
80.9 ± 1.7
Feb
81.9 ± 1.8
96.8 ± 1.7
82.7 ± 1.9
80.7 ± 1.6
96.5 ± 1.5
81.6 ± 1.6
Mar
84.3 ± 2.5
96.5 ± 1.6
84.9 ± 2.5
78.9 ± 2.0
95.4 ± 1.1
79.8 ± 2.1
Apr
91.6 ± 1.8
96.1 ± 1.4
91.9 ± 1.9
77.8 ± 2.4
92.5 ± 2.1
78.6 ± 2.5
May
96.0 ± 0.9
96.0 ± 1.4
96.0 ± 1.0
86.3 ± 2.7
87.0 ± 2.8
86.3 ± 2.6
Jun
97.6 ± 0.7
94.8 ± 1.9
97.5 ± 0.7
94.8 ± 1.9
80.1 ± 3.9
94.0 ± 1.7
Jul
98.1 ± 0.5
93.7 ± 2.3
97.9 ± 0.5
97.0 ± 1.2
77.2 ± 3.0
95.9 ± 0.9
Aug
97.5 ± 0.6
94.1 ± 2.0
97.3 ± 0.6
96.0 ± 1.3
80.6 ± 2.9
95.1 ± 1.1
Sep
96.1 ± 0.8
95.3 ± 1.7
96.1 ± 0.8
89.2 ± 2.4
86.2 ± 2.9
89.0 ± 2.3
Oct
91.4 ± 1.9
96.0 ± 1.4
91.6 ± 1.9
77.5 ± 3.2
91.1 ± 1.6
78.3 ± 3.3
Nov
83.5 ± 2.7
96.1 ± 1.7
84.2 ± 2.8
73.7 ± 2.3
93.5 ± 1.5
74.9 ± 2.5
Dec
80.2 ± 1.9
96.0 ± 3.0
81.1 ± 2.0
76.3 ± 2.2
94.7 ± 1.7
77.4 ± 2.3
Spatial distribution of mean annual FT classification agreement
(%) for PM overpass results in relation to grid-cell-to-point comparisons
with 5344 WMO station SATmax-based FT observation locations for 2012.
![]()
Monthly mean FT classification agreements for AM and PM overpasses and the
1979–2014 record are summarized in Table 1 for the global domain and
Northern Hemisphere (NH), and Southern Hemisphere (SH) sub-regions. The FT
classification agreement was assessed using daily SAT measurements from
approximately 3912 ± 556 and 241 ± 102 respective NH and SH
weather stations for the 1979–2014 record. The monthly pattern of mean FT
classification agreement is generally lowest and highest for corresponding NH
winter frozen and summer non-frozen seasons. Similarly, the monthly pattern
for AM overpass FT results show larger agreement in summer relative to winter
in the SH, while the monthly agreement pattern for PM overpass results in the
SH is more consistent throughout the year.
The mean annual FT spatial classification agreement from the PM overpass
results in relation to 5344 global WMO stations for a selected year (2012) is
92.4 ± 9.9 [spatial SD] % (Fig. 3). Lower latitude areas with
characteristic warmer temperatures show generally greater mean annual FT
classification agreement due to a longer non-frozen period. Relatively lower
agreement occurs along coastal margins, and in mountainous regions. Sub-grid-scale heterogeneity in microwave dielectric properties and associated FT
conditions within the coarse (25 km resolution) satellite footprint increase FT
classification uncertainty along coastal margins (Howell et al., 2009;
Kimball et al., 2001). In mountainous regions, complex terrain and
microclimate variability promote heterogeneous frozen and non-frozen
conditions within the coarser resolution satellite footprint, leading to
greater FT classification uncertainty (Du et al., 2015). Arid climate areas
with dry land surface conditions (e.g. Tibetan Plateau) have a deeper
characteristic microwave emitting depth and larger volume scattering relative
to areas with wetter surface conditions (Prigent et al., 2006; Han et al.,
2015); these conditions may promote a lower FT signal-to-noise ratio and lower
classification agreement with WMO SAT measurements (Grody and Basist, 1996;
Han et al., 2015).
Seasonal pattern of proportional (%) agreement and disagreement
derived from grid-cell-to-point comparisons between 5344 WMO surface air
temperature (SATmin and SATmax)-derived FT values and the
corresponding AM and PM FT-ESDR for 2012: (a) SATmin and AM overpass;
(b) SATmax and PM overpass.
![]()
FT-ESDR daily classification error matrices for year 2012 and AM as well as PM
overpasses were assessed relative to corresponding SAT-defined FT dynamics
from the WMO weather stations (Fig. 4). The resulting mean global daily FT
classification agreement is generally lower during seasonal FT transitions
and highest during the predominantly non-frozen summer season. Spatially and
temporally heterogeneous SAT and FT conditions promote lower FT
classification agreement during seasonal transition periods (Du et al., 2015;
Kim et al., 2011). Two cases of FT classification disagreement between the
satellite and in situ measurements include WMO thaw/FT-ESDR freeze and WMO
freeze/FT-ESDR thaw. The WMO freeze/FT-ESDR thaw category implies earlier
FT-ESDR spring thaw and later fall freeze relative to the WMO FT results,
whereas a reverse pattern is observed for the WMO thaw/FT-ESDR freeze
category. The results of the FT classification error matrices suggest a
mixture of both freeze and thaw classification errors during the seasonal
transition periods and winter seasons.
The mean annual FT classification agreement for the selected year (2012) is
summarized in Table 2 for different land cover categories. Grid cells
dominated by open water bodies show the lowest FT classification agreement
consistent with previous studies (Derksen et al., 2005; Du et al., 2015;
Lemmetyinen et al., 2011); however, the FT-ESDR results still show favorable
agreement for water-affected grid cells despite expected larger constraints
from Tb noise effects due to surface dielectric variations and
microwave scattering over open water bodies. Permanent snow- and ice-dominant
grid cells also show generally lower FT agreement than the other land cover
classes, which may be due to variable surface scattering effects on 37 GHz
microwave emissions over snow and ice (Chang et al., 1990; Matzler, 1994);
the apparent FT classification accuracy is still favorable in these areas
but is based on relatively sparse WMO station observations. For most land
cover classes, the PM FT classification agreement is generally greater than
the AM overpass results except for snow- and ice-dominant grid cells where the
AM FT classification agreement is approximately 2 % greater than the PM
overpass results; the greater apparent AM accuracy is attributed to a lower
number of FT variations for AM than PM overpasses over relatively cold snow-
and ice-dominant land areas throughout the year.
Mean annual FT-ESDR PM overpass spatial agreement (%) with WMO
station-based FT metrics grouped into corresponding dominant IGBP land cover
areas for 2012. Values represent mean and 1 standard deviation variability
in relative spatial accuracy, while the number of weather stations used for
FT accuracy assessment within each land cover category is also noted.
Land cover class*
PM overpass
AM overpass
Number of stations
Open water
77.5 ± 18.3
75.0 ± 15.4
536
Evergreen needleleaf forest
88.6 ± 6.5
84.5 ± 6.3
202
Evergreen broadleaf forest
98.6 ± 4.3
88.4 ± 14.8
38
Deciduous needleleaf forest
93.5 ± 5.7
89.3 ± 4.7
32
Deciduous broadleaf forest
96.2 ± 3.1
88.0 ± 4.0
84
Mixed forest
93.1 ± 7.0
86.7 ± 6.7
751
Open shrubland
95.8 ± 5.9
89.9 ± 7.2
345
Woody savanna
96.2 ± 5.9
92.1 ± 5.9
381
Savanna
99.1 ± 2.6
96.1 ± 4.9
19
Grassland
92.9 ± 7.5
87.7 ± 6.9
888
Permanent wetland
91.4 ± 6.2
86.4 ± 4.9
8
Cropland
94.9 ± 5.7
88.8 ± 5.7
1286
Urban
96.5 ± 3.9
89.5 ± 4.6
102
Cropland/natural vegetation mosaic
94.4 ± 4.5
87.2 ± 4.4
559
Permanent snow/ice
81.1 ± 19.6
83.1 ± 21.3
30
Barren or sparsely vegetation
94.3 ± 6.5
91.8 ± 4.5
82
* Closed shrubland (CS) is excluded in this analysis because only a single CS station was identified.
Dominant factors affecting AM and PM overpass FT classification
agreement for individual grid cells, including (a) open water fraction
within a grid cell (Fw, %), (b) terrain elevation gradient (m)
within a cell, (c) correlation (r value) of grid cell linear regression
relationship between satellite Tb retrievals and weighted
ERA-Interim SAT used in determining the MSTA FT threshold, and (d) the annual
FT transitional season duration (days).
![]()
Factors affecting FT classification agreement
The FT-ESDR results determined from the 37 GHz Tb retrievals are
sensitive to the predominant frozen or non-frozen status of the land surface
within the sensor footprint. However, the actual FT pattern may be spatially
complex due to variations in microclimate and land cover conditions that may
not be adequately resolved by the coarse (∼ 25 km resolution)
satellite sensor footprint (Colliander et al., 2012; Du et al., 2015). The
relative agreement between the FT-ESDR and in situ station-based FT results
is therefore expected to be inversely proportional to land surface
heterogeneity, including open water abundance and terrain complexity within
the grid cell.
We investigated potential factors influencing FT classification agreement
between the FT-ESDR and WMO SAT observations, including the open water
fraction (Fw, %) and elevation gradient (m) within 25 km
resolution EASE-Grid cells coinciding with the WMO validation stations. We
also examined other factors potentially affecting FT classification
agreement, including the level of per grid cell correspondence (r value)
between the Tb retrievals and weighted ERA-Interim daily SAT
estimates used in determining the MSTA FT threshold (Eq. 2), and the mean
annual number of classified transitional FT days defined from the AM and PM
overpass Tb retrievals.
FT-ESDR annual quality assurance (QA) map for 2012, aggregated into
low (estimated mean annual spatial classification agreement < 70 %),
moderate (75–85 %), good (85–95 %) and best (> 95 %)
relative quality categories. Land areas outside of the FT-ESDR domain
are denoted by grey shading; grey shadings depict grid cells where the mean
cold temperature constraint index (CCI) was less than 5 days year-1 from
the 36-year CCI record.
![]()
The mean annual FT classification agreement was analyzed for the selected
year (2012) and found to be inversely proportional to Fw
(slope =-0.25 %, p<0.001) as shown in Fig. 5a. These results
confirm generally strong sensitivity of the 37 GHz Tb observations
to the presence of surface water within the satellite footprint. The
Tb sensitivity to Fw is consistent with large differences
in dielectric properties, scattering albedo and microwave emissions between
open water and adjacent land areas (Lemmetyinen et al., 2011; Rees et al.,
2006). The Tb over lake ice may also be elevated relative to
adjacent land areas during the frozen season (Green et al., 2012). These
factors contribute to a general mismatch between in situ SAT observations
that predominantly reflect local land surface conditions and satellite sensor
Tb retrievals representing an integrated regional signal.
The FT classification agreement was inversely proportional to the elevation
gradient (slope =-0.01 %, p<0.001) within a grid cell (Fig. 5b),
but with less apparent impact than Fw. This negative relationship
is consistent with generally greater heterogeneity in landscape dielectric
properties and FT patterns, and associated microclimate variability in
complex terrain, which may not be effectively delineated by the coarse
(25 km) grid cell resolution (Du et al., 2015; Podest et al., 2014).
The FT classification agreement was directly proportional to the mean
correlation between the satellite Tb retrievals and weighted
ERA-Interim SAT used to define the MSTA FT threshold for each grid cell
(slope =34.3 %; p<0.001; Fig. 5c). These results indicate that the
FT classification is strongly sensitive to the reliability of estimated
Tb threshold values, which may be influenced by uncertainty
associated with use of reanalysis temperature data (Alexeev et al., 2012; Han
et al., 2015; Screen et al., 2012). However, there were relatively few
validation stations in cells with larger Fw cover, complex terrain,
or lower Tb and SAT correlations, which may contribute
uncertainties in assessing the impacts of these factors on FT classification
accuracy.
The mean annual FT classification agreement was generally lower for grid
cells with a longer FT transitional season (slope =-0.1 %, p<0.001;
Fig. 5d). These results are consistent with generally greater FT spatial
heterogeneity and reduced classification agreement during seasonal transition
periods (Fig. 4). These results are also consistent with previous studies
documenting increased FT spatial complexity during seasonal transitions in
spring and fall (Du et al., 2015; Naeimi et al., 2012; Rautiainen et al.,
2014).
FT quality assurance assessment
The global QA map provides a discrete, grid-cell-wise indicator of relative
FT-ESDR quality that accounts for potential negative impacts from
Fw cover, terrain complexity, length of FT transitional season, and
MSTA FT threshold uncertainty influencing mean annual classification
accuracy. The resulting annual QA map for selected year 2012 is presented in
Fig. 6 and shows regions of relative high to low quality. The QA values
were stratified into a smaller set of discrete categories ranging from low
(estimated mean annual FT classification agreement < 70 %) to best
(> 95 %) quality. Mean proportions of the four QA categories encompass
54.1 % (best), 36.0 % (good; 85–95 % agreement), 6.6 %
(moderate; 75–85 % agreement), and 3.3 % (low) of the global FT-ESDR
domain. The QA-based multivariate regressions were computed on an annual
basis and explained more than half (56.9 ± 0.04 [inter-annual SD] %)
of the observed global variation in mean annual FT classification
agreement (R2) with the global WMO SAT validation stations over the
1979–2014 record.
The FT-ESDR contains additional quality control (QC) flags that identify
other factors potentially affecting FT classification accuracy. The QC flags
are spatially and temporally dynamic, and assigned on a per grid cell basis
to denote missing satellite Tb records that are subsequently
gap-filled through temporal interpolation of adjacent Tb retrievals
prior to the FT classification. The QC flags also distinguish grid cells with
large open water areas (Fw > 0.20) and extreme elevation
gradients (> 300 m), and days with large precipitation events where FT
classification accuracy is expected to be degraded. The inclusion of open-water-dominant grid cells allows for greater FT coverage, including extensive
boreal and Arctic wetland areas that were excluded from previous FT-ESDR
releases (Kim et al., 2011). Thus, daily FT estimates are still produced for
grid cells with gap-filled Tb values, large Fw cover,
extreme elevation gradients and large precipitation events, but with dynamic
QC flags indicating potentially lower data quality (e.g. Fig. 5).
FT-ESDR metrics comparison against independent cryosphere data
Lake and river ice phenology
The FT-ESDR defined primary spring thaw date metric showed favorable
correspondence (r=0.79; p<0.01) with observed annual variations in
spring ice breakup for the Tanana River (64.56∘ N,
149.09∘ W) in interior Alaska over the 1979–2003 observation record
(Fig. 7); here, in situ observations of river ice breakup timing (DOY) on
the lower Tanana River are compared with the average FT-ESDR-derived primary
spring thaw date within a 5×5 grid cell domain centered over the
site. Years with relatively early (late) river ice breakup in spring
correspond with generally early (late) FT-ESDR defined spring thaw onset.
These results are consistent with previous studies showing that the 37 GHz
Tb retrievals and FT-ESDR results are sensitive to the onset of
seasonal snowmelt within the basin, which increases runoff and the spring
flood pulse that determine river ice breakup timing (Park et al., 2016b;
Rawlins et al., 2005).
Correspondence between reported annual ice breakup dates for the
Tanana River, Alaska (64.56∘ N, 140.09∘ W), and FT-ESDR-derived primary spring thaw onset averaged within a 5×5 grid cell
(∼ 125 km × 125 km) window overlying the basin.
![]()
The FT-ESDR-based mean annual FS metric over lake dominated grid cells was
directly proportional to collocated annual ice duration observations
(r=0.938; p<0.001) from the Global Lake and River Ice Phenology
Database (1979–2013) as summarized in Fig. 8. The observed annual mean ice
duration anomalies are obtained from 183 ± 133.8 [inter-annual SD]
stations and locations representing northern (≥ 36∘ N) land
areas. These results indicate a direct frozen season impact on lake ice
phenology, where years with a relatively longer (shorter) FS coincide with
extended (reduced) ice duration. The FT-ESDR generally captures the observed
annual lake ice variations despite lower apparent FT classification agreement
for grid cells with greater Fw cover (Fig. 5). The FT-ESDR and lake
ice observations both show significant trends (p<0.001) toward shorter
annual NFS and ice duration consistent with regional warming (Bekryaev et
al., 2010), though the trends may
also reflect a temporal decline in the number of reporting lake ice phenology
stations. Generally, earlier FT-ESDR-derived primary spring thaw onset and a
shorter frozen season relative to the in situ lake and river ice observations
indicate earlier seasonal thawing of surrounding land areas, consistent with
snowmelt- and runoff-driven lake and river ice breakup (Park et al., 2016b).
However, other factors influence lake and river ice phenology, including snow
cover depth, ice thickness, water temperature and wind, which may contribute
to differences between the satellite and in situ measurements.
Correspondence between mean annual (1979–2013) lake ice duration
observations from the Global Lake and River Ice Phenology Database (Benson
and Magnuson, 2000), and collocated FT-ESDR-derived annual frozen season
(frozen or transitional status) estimates.
![]()
Seasonal progression in proportional area of surface melt (Mote,
2014) and FT-ESDR-derived non-frozen conditions over the Greenland ice sheet
in 2012 (a). The spatial pattern of the FT-ESDR-derived annual non-frozen
season (days) over the Greenland ice sheet is also shown (b). Annual
variations in annual surface melt area and non-frozen season over the 36-year
FT-ESDR record are also presented (c); these results document a general
increase in the non-frozen season (1.7 days decade-1; p=0.2) that
coincides with increasing surface melt (0.03 million km2 decade-1;
p<0.001) over Greenland and a 0.6 ∘C decade-1
(p<0.01) regional SAT warming trend indicated from ERA-Interim
reanalysis.
![]()
Greenland surface melt season
The seasonal progression in the proportional area of surface melt (Mote,
2014) and FT-ESDR-derived non-frozen conditions over the Greenland ice sheet
is shown for the selected year (2012) in Fig. 9a. Both daily records show
similar seasonal patterns, but with the spring increase in FT-ESDR-derived
non-frozen conditions preceding the onset of surface melt. The FT-ESDR
results also show a 4.3 % larger non-frozen area relative to active melt
areas; these discrepancies may reflect one or more factors, including
different classification algorithms and environmental thresholds, and
different sensitivities of the 37 GHz V-pol and H-pol Tb
retrievals used in the respective FT-ESDR and Greenland melt algorithms (Mote, 2014).
The FT-ESDR-derived mean annual NFS over Greenland
for 2012 is 50.1 ± 29.1 [spatial-SD] days (Fig. 9b), with a shorter NFS
for inland areas. The FT-ESDR also shows an extended NFS over the Greenland
ice sheet in 1987, 1991, 2002 and 2012, relative to the long-term
(1979–2014) satellite record (Fig. 9c). The extended NFS years are
consistent with observed annual surface melt extremes documented from
previous studies (Abdalati and Steffen, 2001; Hakkinen et al., 2014; Nghiem
et al., 2012). The FT-ESDR annual mean NFS variability is directly
proportional to annual mean surface melt area anomalies for Greenland over
the 1979–2012 record (r=0.535; p<0.001; Fig. 9c). Both the FT-ESDR and
Greenland surface melt data show generally increasing trends in NFS
(1.7 days decade-1; p=0.2) and annual melt area
(0.03 million km2 decade-1; p<0.001). These results coincide
with a 0.6 ∘C decade-1 (p<0.01) regional SAT warming trend
indicated from the ERA-Interim reanalysis.
Summary and conclusions
A new FT-ESDR was developed encompassing a larger global domain, longer
(1979–2014) data record and refined classification algorithms relative to
previous FT-ESDR versions (Kim et al., 2011). The FT-ESDR results show
generally favorable agreement with daily FT estimates defined from global WMO
weather station SAT measurements; the FT classification agreement was
generally stronger for vegetated land areas (mean annual spatial accuracy
from 77 to 97 % for AM and PM overpasses) but was also favorable where
other land cover classes were dominant within a grid cell, including urban
(89.5–96.4 %), barren land (91.8–94.3 %), water bodies
(75.0–77.5 %) and snow- and ice-covered areas (81.1–83.1 %). The
FT-ESDR results also showed general consistency with a broad set of other
cryosphere datasets, including global lake and river ice phenology
observations, and Greenland surface melt cycles.
The mean annual FT spatial classification accuracy relative to WMO weather
station SAT measurements was approximately 90.3 % (PM overpass) and
84.3 % (AM overpass) over the global FT-ESDR domain and a 36-year
satellite record. The apparent FT classification agreement was generally
lower for grid cells with larger Fw cover, greater elevation
complexity, and longer FT transitional periods. The FT classification
agreement was also directly proportional to the correlation between the
satellite Tb retrievals and corresponding ERA-Interim global
reanalysis-based SAT values used for grid-cell-wise calibration of the MSTA
FT thresholds. These results were used to define a relative FT-ESDR annual QA
metric that explained more than half of the observed global variability in
mean annual FT classification agreement. However, the grid-cell-to-point
comparisons used for the QA assessment contained relatively few stations in
cells with extreme terrain or Fw conditions, or with large
disagreement between satellite and in situ FT dynamics, which lead to
potential uncertainties in the FT classification assessment. Other factors
potentially affecting FT classification accuracy were represented by dynamic
QC flags assigned on a daily basis for each affected grid cell.
Overall, the expanded global domain and generally favorable FT-ESDR
performance enables new potential science investigations encompassing
wetlands, mountains and snow/ice-dominant land areas that were excluded from
earlier FT-ESDR versions. The longer FT-ESDR also enables more refined
assessments of global environmental changes that can distinguish transient
weather extremes, landscape phenological shifts and periodic climate
anomalies from longer-term climate trends extending over multiple decades.