IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 1
Emulating Sentinel-1 Doppler Radial Ice Drift
Measurements using Envisat ASAR Data
Thomas Kræmer, Harald Johnsen, Camilla Brekke
Abstract—Using data from the Envisat ASAR instrument, this
paper demonstrates how the high precision radial surface velocity
product, which will become available with ESA’s Sentinel-1
satellite, can complement the analysis of sea ice motion. High-
resolution Doppler frequency measurements are used to estimate
the sub-second line-of-sight motion of drifting sea ice in Fram
Strait. We compare the method with buoy measurements as well
as a recent cross-correlation algorithm for tracking ice between
pairs of images. Maximum speeds measured from the time series
were in the order of 20 cm/s. Using our method we measured
instantaneous speeds reaching 40–60 cm/s.
Index Terms—Sea ice, motion estimation, synthetic aperture
radar, Doppler measurements, Sentinel-1
I. INTRODUCTION
The motion and deformation of sea ice has a major impact
on the ice thickness distribution. Diverging ice creates leads
where new ice can grow, and converging ice piles up and
forms ridges. On a pan-Arctic scale, information on sea ice
motion is needed to quantify ice volume exchanges and for
understanding the momentum, mass and energy balance of
the Arctic Ocean (see figure 1 for an overview of large scale
ice drift patterns in the Arctic).
On a local scale, tracking of potentially dangerous ice such
as icebergs, multi-year ice and ridges is required to prevent
damages to ships as well as oil and gas installations in the
Arctic Ocean. The detection of opening leads also allows for
faster navigation through ice infested waters.
Sea ice motion has traditionally been estimated from time
series of satellite images using both optical and microwave
sensors. Algorithms find common patterns in pairs of images
by use of multi-scale cross-correlation between small image
blocks or by matching derived features such as boundary
polygons of floes or leads (see [1] and references therein).
The matched coordinates together with the time difference
between acquisitions provide estimates of the ice displacement
and velocity. Active microwave sensors are by far the most
popular due to their ability to sense independent of daylight
and cloud cover. In particular, synthetic aperture radar (SAR)
provides a nice balance between wide coverage and good
spatial resolution.
In contrast to the high spatial resolution, the temporal
resolution of satellite SAR time series is low relative to
certain weather events. During the 1–3 days separating two
T. Kræmer and C. Brekke are with the Department of Physics and
Technology, UiT – The Arctic University of Norway, 9037 Tromsø, Norway
(e-mail: thomas.kramer@uit.no).
H. Johnsen is with the Northern Research Institute (NORUT).
The work was supported by the Arctic Earth Observation and Surveillance
Technologies (Arctic EO) project 195143 under contract of the Norwegian
Research Council.
acquisitions, wind and currents may have rendered patterns in
the ice untraceable, especially in the highly dynamic marginal
ice zone (MIZ). This decorrelation may cause large gaps in
the estimated flow fields. Furthermore, the estimated flow
fields only provide average velocities. It has been found that
traditional tracking of ice drift from image time series leads
to a consistent underestimation of the true drift speeds with
biases reported to be as large as 10%–20% when tracking floes
from images separated by 1–3 days [2]. The underestimation
is a consequence of the fact that correlation methods connect
matched points by straight lines whereas the real path is likely
more complex. More precisely, the speed is calculated based
on the displacement of the ice rather than the traveled distance.
Higher time resolution is usually obtained by use of buoys
on the ice or upward looking sonar. These instruments are
valuable, but they only provide point measurements. This mo-
tivates the search for methods which will allow us to observe
near real-time ice speeds from satellite over an extended area.
In this paper we use Doppler frequency measurements
from the Envisat Advanced Synthetic Aperture Radar (ASAR)
instrument to estimate the radial surface velocity of the drifting
sea ice over an extended area. Sections II and III define the
relevant quantities and present their estimation and proper
calibration. With the launch of Sentinel-1 these measurements
will become routinely available; our study illustrates the possi-
bilities of using such products in sea ice research. Our results
are compared to the output from a recent cross-correlation
method as well as measurements from a drifting ice buoy
(section IV). We use examples from the Fram Strait, which
is an interesting area due to higher ice speeds compared to
the central Arctic [3].
II. RADIAL SURFACE VELOCITY FROM SAR
By radial surface velocity we mean the ground range
component of the motion of scatterers on the Earth’s surface,
parallel to the antenna pointing direction. During some satellite
passes, the antenna pointing in the range direction aligns with
the motion of the sea ice. Two techniques allow us to directly
measure radial surface velocities from SAR data: (i) along-
track interferometry (ATI) which requires a second receiving
antenna and (ii) lower resolution single-antenna Doppler shift
measurements. Although the resolution of ATI is impressive
under optimal baseline conditions (33 m × 33 m, see [4]),
such products will not be routinely available in the near
future. ATI also requires a high degree of coherence between
scenes, which is not always the case as sea ice is a highly
dynamic medium. During early ice formation, the temporal
decorrelation time will be close to that of open sea surfaces
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 2
Fig. 1. Mean ice drift (blue arrows) and ocean current patterns (red arrows).
The blue arrows in the center show the Beauford gyre and the transpolar drift.
Adapted from [5, fig. 3.1(a)].
which is in the order of 50 ms at C-band, while deformation
monitoring of fast ice can be conducted with a temporal
baseline of hours.
The single-antenna Doppler frequency measurements, how-
ever, have been available with the Envisat ASAR wide swath
mode (WSM) product since 2007 and will continue to be made
available through the Sentinel-1 Level-2 ocean (OCN) product
[6]. Each OCN product contains up to three geophysical
components: the radial surface velocity (RVL), the ocean
surface wind field (OWI) and the ocean swell wave spectra
(OSW) components. In this paper we show the potential of
the RVL product for sea ice drift estimation.
The body of literature concerning the single-antenna
Doppler method for estimating the radial surface velocity for
ocean wind and current retrieval is growing [7]–[9]. However,
use of Doppler velocity measurements for sea ice applications
is still in its infancy. Fujiyoshi et al. (2013) experimented
using ground based 3D scanning X-band Doppler radar to
estimate ice velocities from the coast near Sea of Okhotsk
[10]. To the authors’ knowledge, the only published study on
sea ice drift using single-antenna spaceborne Doppler shift
measurements was by Hansen et al. [11] using the Doppler
grid available with the ASAR WSM product. The resolution
of the ASAR WSM Doppler grid is relatively low (∼4 km ×
8 km in range and azimuth respectively) because the Doppler
centroids are estimated directly from the raw SAR data which
requires substantial averaging to obtain the required precision.
In this paper we use the estimation strategy of the Sentinel-
1 RVL product where Doppler frequencies are estimated from
single-look complex (SLC) data. For the Sentinel-1 interfero-
metric wide-swath (IW) mode this provides a higher resolution
grid with ∼1 km×1 km cells with an effective resolution of
∼2 km×2 km mid-swath [12].
III. ESTIMATION METHODOLOGY
This section provides an overview of the process for obtain-
ing calibrated radial surface velocities from observed Doppler
frequency measurements (see figure 2). The output is a nor-
malized radar cross section (NRCS) image, a bias-corrected
Doppler frequency image which can be converted to an RVL
image and the predicted standard deviation of the Doppler
estimates (STD). The quantities discussed are expressed in
standard SI units where f denotes a linear frequency in Hz,
t denotes time in seconds, c denotes a speed in meters per
second and λ denotes a wavelength in meters. For exact details
of the Doppler estimation process, we refer the reader to the
RVL algorithm specification document [13].
A. Doppler centroid estimation
The Doppler centroid, f
Dc
, defined as the radar return
frequency shift at the antenna beam center, is related to the
relative motion between the satellite platform and the rotating
Earth
f
Dc
= −2
v
rel
λ
(1)
where v
rel
is the effective relative velocity between the SAR
instrument and the Earth surface and λ is the carrier wave-
length of the SAR system [14]. The sign convention is such
that the Doppler frequency is positive for scattering elements
approaching the radar and negative for elements moving away
from the radar. Precise estimation and calibration of f
Dc
forms
the basis of estimating radial surface velocities.
The starting point for the Doppler centroid estimation is raw
(unfocused) Level-0 Envisat ASAR image mode (stripmap)
data. Each scene was focused to an SLC image using the
full bandwidth of the data, i.e. without applying any window
functions. The Doppler centroid is then estimated from SLC
data using the methodology of Bamler [15] extended to
compensate for side band effects [13]. By side band effects
we mean aliasing of energy from the side-lobes of the antenna
pattern into the main lobe due to strong intensity gradients
in the azimuth direction. This often happens close to land,
where the topography may cause bright returns, but any large
change in backscatter intensity may affect the estimation.
Hansen et al. [16] corrected for this effect by post-processing
of the estimated Doppler frequencies based on a linear fit
between the Doppler frequencies and the backscatter intensity.
In our method, this correction is done directly as part of
the Doppler estimator, using the method presented in [12],
[13]. The Doppler centroid estimation is performed block-wise
using 992 azimuth lines and 192 range pixels with overlap
steps of 1/4 side length (in both directions), producing a ∼1
km×1 km grid with a resolution of ∼2 km×2 km mid-swath.
The stable orbit and attitude of the Envisat satellite allows us
to accurately predict the contribution to the Doppler centroid
due to the satellite–Earth geometry. Chapron et al. [9] showed
that Doppler centroid measurements do not agree perfectly
with the predicted Doppler centroid and that the anomaly
can be related to a geophysical movement of scatterers on
the Earth’s surface. The quantity of interest is therefore this
residual motion. We define the Doppler centroid anomaly as
f
Dca
= f
Dc
− f
geom
(2)
where f
Dc
is the observed Doppler centroid and f
geom
is the
predicted geometric Doppler shift due to the moving Earth
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 3
Raw data
Full-bandwidth
processed SLC
Doppler
estimation
(Section III-A)
Bias correction
(Section III-B)
Absolute
Doppler
calibration
(Section III-C)
Absolute
radial velocity
(Section III-D)
Fig. 2. A schematic overview of the estimation and calibration process for the RVL product.
[14]. To calculate f
geom
we need the orbit, attitude and relative
velocity between sensor and a target on the surface at the beam
center, which is readily found by solving the Range–Doppler
equations using precise state vectors [14], [17]. A geophysical
Doppler shift is then related to the Doppler centroid anomaly
by
f
phys
= f
Dca
− f
bias
(3)
where f
bias
is the total error due to uncertainties in the orbit,
attitude, antenna pattern and topography. After correcting
known biases, f
phys
can be converted to a radial surface
velocity, as described in section III-D. The following section
discusses systematic biases contributing to f
bias
and how to
correct them.
B. Bias corrections
In this section we discuss the dominant biases in the Doppler
anomaly that should be corrected before interpretation of the
results. For calibration purposes it is customary to use images
containing large areas of homogeneous backscatter. Figure 3a
shows the estimated f
Dca
for a VV polarization scene acquired
over the Borneo rainforest 24 July 2004 in the IS5 swath.
Over a rainforest scene we expect the Doppler anomaly to
be zero over homogeneous areas, but in figure 3a we can
easily identify two biases that need to be corrected to properly
calibrate the Doppler measurements: a slow varying trend
in the range direction and a rapid periodic variation in the
azimuth direction. We assume that the biases are independent
and can be corrected for separately.
1) Range bias: In the range direction, the pattern due to
electronic mispointing of the antenna is clearly visible. The
bias, denoted f
range
bias
(x), varies with range cell x and is a
result of the degradation of the transmit/receive modules of
the ASAR antenna array over time [18]. This is a slowly
time-changing bias which varies with polarization and swath.
An estimate of the bias can be obtained by analysis of the
Doppler anomaly calculated over a scene with homogeneous
backscatter. Typically this is done using rainforest scenes and
we use the Borneo scene as an example. However, for the
sea ice images we did not have rain forest scenes which
were close in time. We therefore estimated the profile over
a homogeneous part of the ice where the Doppler standard
deviation was less than 5 Hz. An estimate of the electronic
mispointing, which varies across the swath in the order of
45–53 Hz for IS5, is shown in figure 3b. The estimated range
profile can then be subtracted from each row to correct the
bias.
2) Periodic azimuth bias: In the azimuth direction, a pe-
riodic signal modulates the SLC from which we estimate the
Doppler frequencies. The amplitude of this signal is small
and barely visible in the intensity image (∼0.1 dB), but still
large enough to severely bias the Doppler centroid anomalies.
Figure 3c shows a subset of the azimuth profile obtained by
averaging over all range cells in the Doppler frequency grid
(figure 3a) with a standard deviation less than 5 Hz. The profile
is referred to as f
Dca
. The source of the pattern is a small
difference in gain between calibration cycles in the raw data.
There is a gap in acquisition every 1023rd azimuth line, where
internal calibration measurements are performed. It turns out
that the frequency of this periodic pattern for ASAR stripmap
data is exactly half the frequency of the calibration pulses and
is thus a function of the pulse repetition frequency (PRF) of
the radar. The PRF varies with swath and the modulation has
been found to have a period of 1.24 s in IS2 and 0.98 s in
IS5.
If we take the Fourier transform F{f
Dca
} of the profile in
figure 3c and plot the magnitude spectrum (see figure 3d)
we see a clear peak at 0.98 Hz with one harmonic. For
longer scenes, up to three harmonics have been observed.
Note that this frequency is not a Doppler frequency, but refers
to the variation of the Doppler frequency with azimuth time
as shown in figure 3c. We can estimate the parameters of a
sinusoidal signal with N harmonics from the magnitude and
phase coefficients of F{f
Dca
} [19, p. 256]. Since we know
the base frequency, we search a narrow band centered on
the expected frequencies. The peak within each small band
of the magnitude spectrum |F{f
Dca
}| gives an estimate of the
angular frequencies ω
i
in rad/s and amplitudes A
i
. The phases
φ
i
are obtained from the corresponding position in the phase
spectrum
6
F{f
Dca
}. We then create a correction signal
f
azimuth
bias
(y) =
N
X
i=1
A
i
cos(ω
i
t
y
+ φ
i
) (4)
where N is the number of harmonics and t
y
is the azimuth
time corresponding to azimuth cell y in the Doppler grid.
Figure 3c shows the estimated sinusoidal signal plotted
alongside the original profile. For the scenes in this paper
the periodic signal had a base amplitude A
1
in the order of
2–3 Hz. Note that the correction signal is zero-mean, but
has been plotted with an offset for comparison. The zero-
mean sinusoidal signal is then subtracted from each column
to remove the periodic bias.
C. Absolute Doppler calibration
After compensating for the geometric Doppler and known
range and azimuth biases we expect scatterers which are
not moving relative to the rotating Earth to have a Doppler
centroid anomaly of zero. Due to imperfect estimation of
the mentioned biases as well as inaccuracies in the predicted
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 4
0 20 40 60 80 100
Range cell
0
100
200
300
400
500
600
Azimuth cell
36 40 44 48 52 56 60 64
f
Dca
[Hz]
(a)
35.5 36.0 36.5 37.0 37.5 38.0 38.5 39.0 39.5
Incident angle [deg]
45
46
47
48
49
50
51
52
53
Mispointing f
range
bias
[Hz]
(b)
0 5 10 15 20
Azimuth time t
y
[s]
48
49
50
51
52
53
54
55
Doppler profile f
Dca
[Hz]
Estimated signal f
azimuth
bias
[Hz]
(c)
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
Frequency [Hz]
−40
−30
−20
−10
0
10
20
30
|F{f
Dca
}| [dB]
(d)
Fig. 3. Doppler frequency anomalies estimated over a rainforest scene showing the discussed biases. (a) Estimated Doppler anomaly f
Dca
. (b) Estimated
electronic mispointing in the IS-5 swath as function of incident angle. (c) Mean azimuth profile of f
Dca
for lines 0-200 showing periodic azimuth bias as
well as the estimated signal. Note that the estimated signal is zero-mean, but has been given an offset for easier comparison. (d) Magnitude of the Fourier
transform |F(f
Dca
)| of the profile in (c) showing two distinct peaks.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 5
f
geom
, a residual global offset f
offset
bias
is sometimes observed (in
the order of 1–5 Hz in our case). We therefore estimate this
residual by averaging over land areas in the image with a land
height less than 200 m (as in [16]) and a standard deviation
less than 5 Hz. The offset is then subtracted to produce our
final estimate of f
phys
. Thus, the total bias correction signal is
given by
f
bias
(x, y) = f
range
bias
(x) + f
azimuth
bias
(y) + f
offset
bias
(5)
where we have included the range cell x and azimuth cell y
to make the independence of the range and azimuth biases
explicit.
D. Absolute radial velocity
After correcting for known biases we can convert the
geophysical Doppler values to geophysical radial surface ve-
locities by solving equation (1) for the relative velocity and
projecting to ground range
v
rel
= −
λf
phys
2 sin θ
i
(6)
where θ
i
is the angle of incidence. The corresponding standard
deviation in m/s is then given by
hv
rel
i =
λ
2 sin θ
i
hf
D
i (7)
where hf
D
i is the estimated Doppler standard deviation in
Hz. The RVL grid in the Sentinel-1 product is produced from
Doppler values which have been corrected for the geometric
Doppler as well as antenna mispointing. However, the user
will be responsible for applying any other corrections such as
calibrating the measurements to land.
For easy reference, we provide a summary of the processing
steps here:
• Focus the raw data to a full-bandwidth procesed SLC
image (i.e. without using window functions).
• For each cell in the Doppler grid, obtain the estimated
Doppler frequency f
D
(eq. (30) in [13]).
• Calculate the geometric Doppler for each grid cell (see
e.g. [14, Chapter 12]) and subtract it from f
D
to produce
f
Da
(equation (2)).
• Get antenna mispointing f
range
bias
(x) by estimating it from
data or from auxiliary data and subtract it from f
Da
.
• Further, estimate and subtract the periodic bias in azimuth
f
azimuth
bias
(y), then estimate and subtract any residual offset
in the Doppler over land to produce f
phys
.
• Convert the calibrated Doppler to a radial surface velocity
using equation (6).
E. Uncertainties
We have not done a detailed investigation on the magnitude
of each uncertainty. However, this section discusses sources of
uncertainty which should be kept in mind when analyzing the
data. We assume that the wavelength is known perfectly. Thus,
the uncertainties contributing to v
rel
are those contributing
to f
phys
and the incidence angle θ
i
(see equation (6)). As
discussed above, these are f
Dc
, f
geom
and f
bias
.
Fig. 4. Subset of an ice scene showing the size of a single Doppler estimation
cell (red box) containing a mixture of ice and water. The size of the box is
992 azimuth lines and 192 range pixels (∼4 km×3 km). Since the estimation
is done on a regular grid, the probability is high that a cell may contain both
ice and water in the same cell which will increase the standard deviation of
the Doppler estimates.
1) Uncertainties in f
Dc
: There are two things that can
influence the bias and standard deviation of the Doppler
estimates f
Dc
: (i) If there are strong scatterers (e.g frost flowers
are known to cause strong backscatter for C-band radars [20])
close to the estimation cell, we may get aliasing into the
estimation area. If side-band effects are not corrected for, this
will introduce a bias and increase the standard deviation, but
our approach does correct for this. (ii) There may be a mix of
intensities in the SLC image and/or a mix of motions inside
the physical estimation area or both. Figure 4 shows a subset
of the full resolution image and a Doppler estimation cell
where such mixing occurs. If there is a mix of intensities
the Doppler standard deviation will increase, but we still
obtain an unbiased estimate of the average Doppler centroid.
There could potentially also be multiple motions inside the
estimation cell. In this case we still get an unbiased estimate
for the average motion inside the cell, but it may not be very
informative.
2) Uncertainties in f
geom
: The accuracy of f
geom
and the
angle of incidence θ
i
depends on the precision of the orbit state
vectors and the solution of the Range–Doppler equations [14].
We have used the Doppler Orbitography and Radio-positioning
Integrated by Satellite (DORIS) precise orbit state vectors
which have an accuracy of 30 cm along each axis and the
Range–Doppler equations are solved iteratively to centimeter
precision [21].
3) Uncertainties in f
bias
: The total bias, f
bias
, is composed
of measurements of the electronic mispointing of the antenna,
the periodic azimuth pattern and any residual offset. In our
study with ASAR data, the largest unknown is the antenna
mispointing as we estimated this over the ice sheet within
each image. For Sentinel-1, the mispointing will be monitored
continuously making the correction more reliable.
The periodic azimuth variation seems to be specific to the
stripmap mode for both ASAR and Sentinel-1. Sentinel-1 will
be using the extended wide-swath (EW) mode to study polar
sea ice and the wave (WV) mode over ocean, neither of which
seem to be affected by this phenomenon.
![Fig. 1. Mean ice drift (blue arrows) and ocean current patterns (red arrows). The blue arrows in the center show the Beauford gyre and the transpolar drift. Adapted from [5, fig. 3.1(a)].](/figures/fig-1-mean-ice-drift-blue-arrows-and-ocean-current-patterns-26qw54eu.webp)
