Content uploaded by Anton Sigitov
Author content
All content in this area was uploaded by Anton Sigitov on Sep 18, 2020
Content may be subject to copyright.
Comparison of satellite imagery based ice drift with wind model
for the Caspian Sea
Yevgeniy Kadranov1, Sigitov Anton1, Sergey Vernyayev1
1 LLP Iceman.kz (Kazakhstan)
ABSTRACT
Many factors influencing the movement of ice such as wind, ice concentration, ice thickness,
roughness, water currents, Coriolis force, bathymetry, artificial and natural obstacles in the
area. Current speeds in the Caspian Sea are relatively small and so the main driving force for
ice movements is wind. Therefore, main goal of this work was to study wind-ice movement
velocities dependence in the region and check how ice concentration and thickness influence
on the movement of ice.
A high number of measurements and observations was made to describe ice drift in the region,
although the data was collected areas and usually not publicly available. In our work, we
have used timely consequent optical and SAR satellite images to observe ice movements and
its displacement over the area. Wind data for the same period and area was taken from wind
models. Ice charts were prepared using visual interpretation of satellite imagery. Ice
information (concentration, stage of development, floe size) were stored as vector data in
SIGRID3 format. The described data has been correlated and analyzed.
The analysis provided in the work can be used for the forecast of short term ice drift on the
operational basis and can be the first step for creation of ice drift forecast model for the
region of North Caspian Sea. The used data, methods and results of the study are described in
this paper.
KEY WORDS: Ice Charting, Ice behavior modeling, Ice drift, Caspian Sea, Optical and SAR
Imagery interpretation, Weather model, GFS
INTRODUCTION
The area of North Caspian Sea is of great interest for Oil & Gas industry with the ongoing
and projected offshore development of relatively complex oil fields. Ice detection and
forecasting is needed for support of winter operations at site and along transit routes of
marine supply lines. One of the major concerns is forecasting ice conditions in the region to
plan logistics and avoid hazardous or time-consuming ice and weather impact on operations
as discussed by Verlaan et al, 2011. As one of the major drivers of ice related hazards to the
offshore installations and navigation is drift, it is useful to understand what the factors control
the ice drift over the region. Developing the relationship between the forecast weather
parameters, local conditions across region and the resulting drift is the way to build a model
to predict ice drift in the region. Objective of this project is to study wind induced ice drift in
the region and define to what extent ice concentration influence on the resulting
displacements based on observations of the relatively mild ice season 2015-2016.
POAC’17
Busan, KOREA
Proceedings of the 24th International Conference on
Port and Ocean Engineering under Arctic Conditions
June 11-16, 2017, Busan, Korea
Figure 1. North-East Caspian Sea
North East Caspian Ice conditions summary
The North-East Caspian Sea normally stays ice covered from late November-Early December
to Late March-Early April. Thermally grown ice thickness typically reaches 40-50 cm along
the northern shoreline during an average winter. Ice drift is mostly wind driven and usually
occurs over deeper parts of the NE Caspian in the central basin or south of the Saddle area
(Figure 1), while ice sheet in the shallower area remains stationary during the major part of
the winter. Ice compaction caused by strong wind events often lead to forming of ice features
such as stamukhi (grounded rubble mounds) and pressure ridges as well as extensive rafted
areas as observed by Crocker et al, 2011. Figure 2 illustrates seasonal variation of ice cover
distribution over the last nine seasons 2007-2016.
Based on the analysis of remote sensing data it was noted that the warmer the winter is, more
mobile areas are observed over the NE Caspian. Therefore, it was decided that the season
2015-2016 should be used for the comparison of ice drift/wind dependency. As the ice season
lasted from late December to early March, it was shorter than usual. Only 221 FDD (as
observed in Atyrau) were collected during this winter – nearly 3 times less than during the
average Caspian ice season. Thickest level ice that has grown this season was about 25 cm as
calculated by formula derived for this region by Jordaan et al, 2011. Ice conditions during the
season 2015-2016 were more dynamic than usual as observed during execution of this project.
This was mainly due to milder weather throughout the year (Figure 2). Most Ice in the NE
Caspian never consolidated and remained mobile expect for the areas along the shore in the
North and East. This created ideal conditions for ice drift analysis based on comparison of
satellite images as same ice floe could be tracked for several weeks (while during other
seasons, after several days of tracking floe would be deformed with compaction against
stationary features or consolidates as part of stationary ice sheet).
Figure 2. This picture compares 9 ice seasons (from 2007-2008 to 2015-2016). FDD data
from NOAA National Centers for Environmental Information.
DATA SOURCES
Satellite imagery
The major source of data for ice drift measurement within the scope of this project wass
satellite images. Choosing between SAR and optical images, SAR is more reliable as it does
not depend on cloud cover. However, due to geographical location of the area of interest SAR
satellites have revisit time of 3-4 days per week and need to be supplemented with any
available optical images.
The season had a fair amount of cloudless days (63% of MODIS images were either clear or
partly cloudy as illustrated on Figure 3), which allowed to use optical MODIS Terra and
Aqua data from NASA worldview portal intensively.
Figure 3. Cloudiness stats for Jan-Mar 2016
Publicly available Sentinel-1 and commercial RISAT-1, TerraSAR-X and Radarsat-2 SAR
images were used to close some of the big gaps caused by cloudy sky conditions. Public SAR
imagery were taken from Copernicus Science hub, while commercial SAR images were
provided by Kongsberg Satellite Services (KSAT).
Ice Drift Data
Ice drift data over the period from 6 January 2015 to 2 March 2016 was gathered by
identifying similar ice floes on subsequent images of the series and tracking their
displacement. The start date of tracking is the day, when ice cover has established over the
major part of the Central Basin of the NE Caspian and South of Saddle.
There were two automated steps of processing satellite images targeted to increase accuracy
of the resulting displacement values. Both steps were performed within QGIS environment
scripting specific plugins. During the first one, distinctive ice floes were identified on a pair
of images and were polygonised as a geo-spatial feature with the following attributes
assigned to each feature:
• Start and End Image date
• Floe ID
• Start and End ice concentration around the floe
• Area and perimeter values
• General Comments ( rotation of floe, unique behavior, deviation from flow direction)
Figure 4. Example of digitizing same ice floe for 3 days (Left). Frequency distribution of drift
observations intervals (Right)
As the second step the other QGIS plugin built vectors
between centroids of floes using date and floe id attributes,
and calculated vector displacement and direction between
adjacent by date floes. Figure 4 (Left) illustrates the
example of floes identification and resulting displacement
vectors represented in graphical form.
In total for the whole season, 494 ice drift vectors were
created and 63 pairs of satellite images were analyzed,
making an average of about 8 drift vectors per analysis.
The most frequent time interval between two images was
about 24 hours, with minimum interval of around 2 hours
(between MODIS Terra and Aqua from the same day) and
Figure 5 Drift vector and its
components
rare maximum of 60 hours (Figure 4 Right).
Drift speed vectors consisting of vector magnitude, direction and two projections on x and y
axes were derived from the displacement data. The magnitude of the average ice drift speed
vector of the floe was found as an ice floe displacement distance divided by the time interval
between the images. Direction of ice drift vector was measured as azimuth between
displacement vector measured from true North. The zonal and meridional components of ice
drift speed vector were found as orthogonal projections of ice drift speed vector on X (East-
West) and Y (North-South) axes correspondingly (Figure 5). The resulting values were stored
in the database.
Ice Conditions Data
Ice conditions description such as total concentrations, partial concentrations, floe sizes and
stages of development was prepared and stored as a geo-spatial database in standard SIGRID-
3 format for the whole region. Ice conditions data has then been assigned to intersecting ice
drift floes with the same date. The resulting ice concentrations were then joined with ice drift
vector for each individual vector as attribute containing categories describing it at the start
and the end of each track (Table 1). As the season was extremely mild the stage of
development more than grey-white ice has never been grown. This has also allowed us to
assume the ice thickness conditions were more or less uniform across the sea and neglect its
effect on drift behavior.
Table 1. Ice drift concentration categories.
Final
Low (1-3)
Medium (4-6)
High (7-10)
Initial
Low (1-3)
LL
LM
LH
Medium (4-6)
ML
MM
MH
High (7-10)
HL
HM
HH
Wind Data
GFS wind analysis model data (10m above Sea Level with spatial
resolution of 30x20 nautical miles for the Caspian Sea) was taken as
grib file from NOAA datahub, and converted into vector geo-spatial
format. The database of wind speed vector components (meridional
and zonal) for the season was derived based on four model runs per
day, and contained two (+0 hrs and +3 hrs) output array data points
from each model run. Considering the data points for resulting
database were taken near the model run it was assumed that the
dataset was close to real.
FINAL DATABASE
The data described above was merged into a single database with
each record containing corresponding information about ice drift data,
wind data and surrounding ice conditions data to perform further
analysis. To do that each drift vector’s centroid was associated with
the nearest GFS grid point. All the wind vectors at that point were
averaged within time interval between two satellite images including
Figure 6 Selected
wind data time
interval for averaging
for each drift interval.
one more wind vector before and after such interval. The following attributes were derived
for each averaged wind vector:
• Averaged zonal (W-E) and meridian(N-S) components of wind vector
• Averaged wind speed (vector magnitude) and direction
• Maximum wind direction deviation from the averaged wind vector using Yamartino
method (Yamartino, 1984).
Figure 7 shows wind data as per oceanographic convention (direction toward which the wind
is blowing) for the period of observed drift intervals. Presentation of wind vectors with
direction ‘toward’ as opposed to ‘from’ simplifies the calculus. The drift and wind values of
compiled database are in form of drift speed distributions by direction. Most of the wind
events have happened towards NW-W and S-SE directions as typical for the region as
concluded from years of observations by Nilsen et al, 2011.
Figure 7 Wind speed and direction (toward) distribution for observed drift intervals (Left) and
Ice drift speed and direction distribution (Right).
Flowchart illustrating all the steps of merging the data is schematically shown on Figure 8.
Figure 8. Data flow for Ice Drift geodatabase compilation
Resulting database revealed that ice drift vectors not always correspond to averaged wind
vectors. Depending on time interval between the images ice floe could experience variable
range of wind speeds with variable wind directions. For example, wind with even the
constant speed can reverse its direction during the time interval, so the real track length of the
floe will be much longer than the net floe displacement resulting in significant decrease of
average drift speed. Figure 9 shows the wind and drift direction difference versus wind speed
illustrating the point above. Most of the used data for wind speeds above 5 knots are within
narrow interval of drift-wind difference values as it should be for wind driven drift as we
expect in the Caspian.
Figure 9. Scatter plot of Drift-Wind directions difference from wind speed
To clean the dataset from the discrepancies caused by the uncertainties of scarce data
observations the whole database was filtered to exclude:
1. All drift vectors with the duration of less than 4 hours. It was found that even during
homogeneous wind event, accuracy of drift vector length calculation can be
significantly affected with image’s resolution (250m for MODIS);
2. All the average wind vectors where wind direction deviation was more 60 degrees;
3. Most of the records where average wind speed was lower than 5 knots. As observed
during the data compilation, such low values are normally associated with erroneous
speed averaging process mentioned above. Low winds events resulting in small
displacements with significant difference between drift and wind directions are also
regulated by other factors such as currents, ice compaction, sea surface tilt during
surge events that become noticeable, but could be neglected with higher winds;
4. Events when the compaction against coastlines or grounded ice features or diverging
processes caused by non-homogeneous wind were observed. In this case the frictional
forces and internal stresses between ice floes starts to play dominating role in floe
behavior
5. Odd cases without reasonable explanation of drift behavior. Sometimes ice floe
displacement was observed with incomparable wind speeds. It potentially could be
caused by discrepancies in wind model output due to time lag or difficult synoptic
situation that could not be handled within model.
As the dataset was cleaned of the cases listed above filtered data has shown clear relationship
between wind and drift directions as can be seen on scatter (Figure 10 Right) as one would
expect it to be in the region with prevailing wind driven drift as observed by Nilsen et al,
2011. The regression line for wind versus drift speed relationship (Figure 10 Left) showed a
relatively big spread of data around the curve indicating that direct calculation of the drift
speed based on one drift/wind ratio is not sufficient for modelling the ice drift in the region.
The distribution of the drift-wind speed ratios frequency shows that the range bin from 2 to 3%
is dominant - 41% (Figure 11 Left) and the coefficients in range 1-4% appear for almost 90%
of time. This has led us to conclusion we need to segregate these ratios based on
concentrations and drift directions as these seemed to be the major factors affecting the
coefficient. It can be seen from the Figure 11(Right) that, in general, the ratio coefficient
tends to increase with lowering concentration.
Figure 10. Scatter plot of data for Drift-wind speed (Left) and Drift-wind direction (Right)
Figure 11. Drift-wind speed ratio frequency distribution (Left). Drift-wind speed ratio
frequency by concentration category (Right).
DRIFT MODEL
General ice drift behavior in regards of its dependency to direction and concentrations
became clear. Further on, equation (1) presented by Leppäranta (2005) and later simplified by
Segboer and Verlaan (2007) was used to describe relationship between drift and wind. As
currents measurements results are not publicly available they were accounted within this
equation as residual variables described below.
,
(1)
where Dx, Dy are orthogonal zonal (West-East) and meridional (North-South) components of
ice drift speed vector
; Wx, Wy – components of synchronized wind speed vector
, Ex, Ey
– components of residual drift speed vector variation
that can’t be explained by influence
of the wind using linear transformations with matrix A with components a11, a12, a21, a22.
A linear regression analysis has been used to determine matrix coefficients a11, a12, a21, a22
and E vector components from Eq. (1) based on records of drift vectors and corresponding
averaged wind data. The results of the regression analysis are presented in Table 2 and Figure
12.
Table 2. Regression analysis outcome for different drift-wind records set
All data
Filtered data
All
LL
MM
HH
Number
of records
493
292
132
51
72
a11
0.0254
0.0257
0.0286
0.0265
0.0219
a22
0.0235
0.0254
0.0277
0.0237
0.0259
a12
0.0009
0.0004
0.0010
0.0038
0.0008
a21
0.0052
0.0070
0.0097
0.0050
0.0055
Ex
0.0152
0.0029
-0.0028
0.0466
0.0086
Ey
-0.0119
0.0038
0.0241
-0.0012
-0.0252
R2 Dx
74.3%
82.8%
88.8%
82.4%
77.8%
R2 Dy
75.2%
83.4%
84.2%
88.4%
77.0%
Figure 12 Regression coefficients comparison for various sets of data.
The set of computed coefficients above was subsequently used to hindcast drift for cases with
known wind and drift conditions based on GFS forecast data archive. As an example, a single
ice floe has been tracked for period of almost 2 weeks (30-Jan-2016 to 12-Feb-2016). The
black arrows in Figure 13 (Left) indicate observed ice displacements over the imagery
interval periods; green arrows are modeled ice displacement over the same periods using
averaged wind and blue arrows are modeled ice displacement over 3-hours wind interval.
Figure 13. Comparison of measured and modeled ice drift for the periods 30/Jan/2016 to
12/Feb/2016 (Left) and 07/Feb/2016 – 17/Feb/2016 (Right)
Figure 14 shows two samples of modelled drift using GFS data and three different
coefficients corresponding to different concentration ratings. These cases show better fit
when corresponding set of coefficients is used in different by concentration conditions.
Figure 14: Results of verification in unconfined conditions of low (left) and high (right)
concentration.
However, once there is an obstacle or confinement at coastline the model does not show any
good fit at all as demonstrated on Figure 13 (Right), when the actual drift has finished in front
of stable ice zone along the coastline and the model continued its progress.
DISCUSSION
The values of a11 and a22 for all set of data are between 0.02 and 0.03. The residual drift
variation (Ex and Ey) values were significant for low wind speeds (especially for MM and HH
concentration categories), but become negligible with increasing wind speeds as wind starts
to dominate over other factors.
The coefficients a12 and a21 are mostly much
smaller than a11 and a22, except for several
cases when a21 almost reaches 0.01 for LL
category, meaning that for Low concentrations
wind zonal component (W-E) tends to
influence drift meridional component (N-S)
turning the drift vector a bit to the left from
wind vector. Apparently positive a12 and a21
coefficients show that the response of ice
depends not only on wind speed, but also on
wind direction. To demonstrate this effect, the
ice drift response ellipse was built (Figure 15).
It shows the response of the ice to
circumferential wind with constant speed of 10
knots. For wind of 10 knots in NE and NW
direction (blue vectors) ice will tend to drift in
the same direction with slight deviation to the left (red vectors). The general idea is that drift
speed in NE or SW directions (big semi-axis), for example, tends to be almost 30% higher
than for NW or SE (small semi-axis) for the same wind speed. These results make sense for
the NE Caspian Sea (Figure 1), that has NE-SW stretched form and fast ice forming along
coastlines creates confinement for mobile ice to easily drift in NE and SW directions.
CONCLUSIONS
The major achievement of this project is that a reliable dataset of ice drift in the region was
compiled for the season 2015-2016. The dataset was used to gain initial understanding of
how ice responds to wind in the conditions of the Caspian basin. Ice moves with 2 to 3% of
wind speed for 41% of tracks and between 1-4% for about 90% of tracks. Correlation of wind
direction to resulting drift direction was challenging due to many factors guiding the process.
The method chosen to correlate wind and drift vectors could be used as a working solution
for variable concentrations if there is no confinement of drift with coastline or landfast ice
and obstacles in form of stationary ice features or artificial structures.
Introduction of Sentinel constellation with sufficient frequency of reattendance in the region
starting last winter (2016-2017) the authors have received opportunity to grow the drift
database and further improve the model. Introduction of additional set of coefficients and
logic considering proximity to obstacles and coastline is the most obvious first step for
improvement. This work is based on public domain data such as the MODIS Terra and Aqua
satellite data and the GFS winds. With higher spatial resolution weather data (for example
ECMWF) as well as more SAR satellite imagery to increase frequency of observations could
improve the results. Coefficients can also be improved a lot if calibrated with field
measurements (e.g. from drift buys) if they were publicly available.
Figure 15 Drift response ellipse to wind
speed circle of 10 knots.
ACKNOWLEDGEMENTS
Sean McDermott has inspired the whole development of the ice project in the history of its
development. Authors are grateful to Paul Verlaan for giving his opinion on the work
performed and advice on future development of this project. KSAT as the leading provider of
SAR imagery in NRT operational mode has been the most significant facilitator of our Ice
Charting operations in the Caspian and has contributed in our research with archive SAR
images that helps us conduct our research and deliver better analysis to community. Carles
Debart has spent significant time on technical discussions with us in regards of using SAR
images and support in acquiring archive data.
REFERENCES
Verlaan P.A.J., Croasdale K. 2011, Ice Issues Relating to the Kashagan Phase II Development,
North Caspian Sea'. Proceedings of the 21st International Conference on Port and Ocean
Engineering under Arctic Conditions, July 10-14, 2011, Montréal, Canada, POAC11-171.
Crocker G., Ritch A., Nilsen R., 2011, Some Observations of Ice Features in the North
Caspian Sea. Proceedings of the 21st International Conference on Port and Ocean
Engineering under Arctic Conditions, July 10-14, 2011, Montréal, Canada, POAC11-118.
Jordaan I., Stuckey P., Bruce J., Croasdale K., Verlaan P., 2011, Probabilistic Modelling of
the Ice Environment in the Northeast Caspian Sea and Associated Structural Loads.
Proceedings of the 21st International Conference on Port and Ocean Engineering under
Arctic Conditions, July 10-14, 2011, Montréal, Canada, POAC11-133.
Leppäranta, M. 2005. The Drift of Sea ice. Praxis Publishing Ltd., Chichester, UK.
Nilsen R., Verlaan P.A.J., 2011, The North Caspian Sea Ice Conditions and how Key Ice Data
is Gathered. Proceedings of the 21st International Conference on Port and Ocean
Engineering under Arctic Conditions, July 10-14, 2011, Montréal, Canada, POAC11-112.
Segboer T.J. and Verlaan P.A.J., 2007, Ice Drift Under the Influence of Winds and Currents
along the Sakhalin Northeast Coast. 2007 Dalian University of Technology Press, Dalian,
ISBN 978-7-5611-3631-7
Yamartino, R. J., 1984. A Comparison of Several "Single-pass" Estimators of the Standard
Deviation of Wind Direction. J. Climate Appl. Meteor., Vol. 23, pp.1362-l366
Raw Data Sources: NASA worldview portal, NOAA National Centers for Environmental
Information, GFS weather model data, Copernicus Science Hub
