A survey of algorithms developed for satellite snow and sea ice detection George Bonev A literature review submitted to the Graduate Faculty in Computer Science in partial fulfillment of the requirements for the degree of Doctor of Philosophy, The City University of New York Committee Members: Dr. Irina Gladkova (Advisor) Dr. Michael Grossberg Dr. Peter Romanov July 2, 2015
39
Embed
A survey of algorithms developed for satellite snow and sea ice … · 2015-10-26 · Radiometer Suite (VIIRS), the Advanced Microwave Scanning Radiometer-2 (AMSR-2), the Special
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
A survey of algorithms developed
for satellite snow and sea ice
detection
George Bonev
A literature review submitted to
the Graduate Faculty in Computer Science in partial
fulfillment of the requirements for the degree of Doctor of
Table 3.2: NASA Team Sea Ice Algorithm Tie Points and the coefficients usedin the claculation of SSMI Sea Ice Concentration [Comiso et al., 1997].
The two sets of SSM/I tie points (one for the Northern and Southern Hemi-
spheres) represent a “global” set that is designed for mapping global sea ice
concentration on a large scale. Improved accuracy can be achieved through
the use of regionally calibrated tie points.
20
3.4.2 Bootstrap Algorithm
The main assumption used in the Bootstrap algorithm is that within each
satellite field-of-view the surface is covered by either sea ice or ice-free ocean.
The brightness temperature recorded by the satellite’s passive microwave sen-
sor is thus a contribution of radiation from ice convered and from ice free
surfaces. Using this assumption, the brightness temperature TB is expressed
by the following linear equation:
TB = T1C1 + T0C0 (3.11)
where T0 and T1 are reference brightness temperatures of open water and sea
ice, while C0 and C1 are their respective concentraion percentrages. Using the
fact that C0 + C1 = 1, eq. 3.11 can be solved for C1 to obtain:
C1 =TB − T0T1 − T0
(3.12)
Sea ice in this case can be any or a combination of the various ice types. TB,
T0 and T1 all include contributions from the intervening atmosphere. Appro-
priate values of T0 and T1 should be used in equation 3.12. These values are
determined from a set of two channels (1 and 2) in the following manner.
In the HV37 and V1937 scatter plots shown in figure 3.2a,b, a data point
represented by (T1B, T2B) for channels 1 and 2 has corresponding reference
brightness temperatures for ice given by (T11, T21) and for open water given
by (T10, T20). These three points are labeled B, I and O respectively. Using
this formulation, a data point at location I along the line AD is assumed to
represent 100% ice concentration of the same type as at point B. Point I is
21
Figure 3.2: (a) Scatter plot of 37H versus 37V SSMI monthly January 1992data for entire Arctic region. (b) Scatter plot of 19V versus 37V SSMI monthlySeptember 1992 data for entire Antarctic region. [Comiso et al., 1997]
determined from the intercept of lines AD and OB. Point O is close to the
lowest brightness temperatures for open water and near the intersection of OA
and OW , where OW is the line representing water data. Data points that fall
along the line OI thus represent different concentrations of ice. As dictated
by eq. 3.12, the ice concentration varies linearly along this line. Solving the
set of two equations for lines AD and OB for the two unknowns, the values
at the intercept are given by:
T11 =(T1A − T1O − T2ASAD + T2OSOB)SOB
(SOB − SAD)+ T1O − SOBT2O, (3.13)
T21 =(T1A − T1O − T2ASAD + T2OSOB)
(SOB − SAD), (3.14)
where SAD and SOB are the slopes of lines AD and OB, respectively, (T1A, T2A)
represents any point along line AD, and (T1O, T2O) represents the open water
reference brightness temperature.
For any data point B the ice concentration can be derived from eq. 3.12
22
or from the ratio of distances OB and OI given by:
C =
√(T1B − T1O)2 + (T2B − T2O)2
(T21 − T2O)2 + (T11 − T1O)2. (3.15)
In equations 3.13 and 3.14, T2A and T1A represent the brightness temperatures
at point A along the line, and it is usually convenient to choose, T2A = 0 and
the offset as shown in table 3.3 for T1A.
Winter Summerchannel N. H. S. H. N. H. S. H.19V 179◦K 179◦K 181◦K 179◦K37V 202◦K 202◦K 203◦K 202◦K37H 130◦K 130◦K
Table 3.3: Pre-computed open water tie point offsets (T1A in equations 3.13and 3.14) based on 1992 SSM/I data.[Comiso et al., 1997]N.H.: Northern Hemisphere, S.H.: Southern Hemisphere
For the Northern Hemisphere, the HV37 scheme is applied to data above the
dotted reference line AD− 5 K in figure 3.2a (corresponding to ice concentra-
tions > 90%) and the V1937 scheme is applied to the rest of the data. This is
due to the fact that V1937 provides better discrimination between open water
and ice in the seasonal regions and near the marginal ice zone, than HV37. In
the Antarctic region, there is less variation in the ice signature, so only V1937
is used.
3.5 Machine Learning Based Algorithms
Machine learning presents an alternative to the statistical and physical meth-
ods for estimating snow and sea ice extent. Artificial Neural Networks (ANNs)
are a machine learning technique often used for learning relationships between
input and output variables. ANNs define an information processing model
23
that stores empirical knowledge through a learning process and subsequently
makes it available for future use. ANNs have been utilized in various remote
sensing applications. They are superior to some other approaches due to the
fact that they don’t require the assumption of a pixel being a linear mixture
of signals.
There are only a few studies that utilize machine learning, and more
specifically ANNs, for satellite snow and sea ice detection. Two algorithms
from this space are reviewed in this section. One implements a fractional
snow cover mapping approach via artificial neural network analysis of MODIS
surface reflectance data [Dobreva and Klein, 2011]. The other uses a com-
bination of feed-forward neural networks and data from the 1.6 µm middle
infrared channel to classify satellite data into sea ice, cloud, water and leads
[McIntire and Simpson, 2002].
3.5.1 Fractional snow cover via ANN and MODIS re-
flectance
In this study a multilayer feedforward Aritificial Neural Network with one
hidden layer is used. Since the network is feedforward and not recurrent, it
does not include any feedback loops, so inputs to a neuron are not influenced
by its output. In this multilayer feedforward configuration the input layer of
source neurons (surface reflectance, NDSI, NDVI, and land cover of a pixel)
project to a hidden layer of neurons that project directly to the output layer,
which corresponds to the snow fraction of the pixel. The network utilizes the
hyporbolic tangent function as its transfer function. The network is trained
using the backpropagation algorithm. The training data is derived from
higher resolution imager data obtained from the Landsat instrument. Snow
24
fractions from the higher resolution images were aggregated to compute the
expected snow fraction values at MODIS resolution.
The algorithm performed well with accuracy around 90%. This level of
accuracy was in most cases equal or better than that of the operational
MODIS snow fraction product [Dobreva and Klein, 2011].
3.5.2 Arctic sea ice, cloud, water and lead classification
For this method a combination of reflectance data obtained from the Chinese
Fengyun-1C satellite (with a sepcial emphasis on the daytime 1.6-µm middle
infrared data that is extremely useful for discriminating between ice and
clouds) and a multistage feedforward Neural Network are used to achieve
improved daytime classification of clouds, sea ice, leads, and open water that
is needed for polar studies.
The algorithm uses three stages to separate Arctic sea ice from cloud,
water, and leads. Each neural netwrok stage, computes an image-specific
nomalized [0, 1] dynamic threshold for a specific wavelength band. Each
normalized dynamic threshold is then compare with also normalized image
data for classification at that stage.
Preprocessed (i.e. noise removal, navigation, subsection) input data en-
ters stage 1 of the algorithm, which associates the majority of illuminated
water cloud with high values of 1.6-µm observations. Stage 2 detects residual
clouds using a low 11-µm signature in the unclassified data pass to it from
stage 1, based on the assumption that the detected residual cloud is an
ice cloud. Stage 3 examines the remaining unclassified data: water/leads
25
have low values of albedo at visible wavelenths, while sea ice has high val-
ues. Cloud shadow in the scene can potentially compromise stage 3, thus
a post processing cloud shadow removal method is utlized to mitigat this issue.
Each stage of the algorithm contains an Artifical Neural Network with
one hidden layer. These networks are trained individually using the back-
propagation algorithm. The training data used consists of the layer inputs,
i.e. sensor image statics for each of the three sensor frequencies used (0.63µm,
1.6µm, and 11µm), and predetrmined optimal thresholds, obtained from a
human expert with several years of experience.
The algorithm performance was validated against Indepndent Sea Ice Analysis
[McIntire and Simpson, 2002] of the National Weather service for random
periods between April and August 2001. The overall accuracy of classification
was found to be 98% for the 218,700 testing data points.
3.6 Multi-Instrument Integrated Algorithms
Given the limitations of instrument types, presented in Chapter 2, algorithms
that attempt to mitigate these limitations by utilizing data from different
sensors provide a promising avenue for improving our satellite snow and
sea ice detection abilities. Two methods for combining data from different
instruments are reviewed.
The first method described in [Foster et al., 2011] derives a global snow
extent map using data from MODIS (visible and infrared), AMSR-E (passive
microwave) and QSCAT (active microwave). Since snow cover extent is
26
identified better at the visible and infrared wavelengths, MODIS observations
are used as the default. The passive and active microwave derived snow
cover is used only in areas where MODIS observations are limited by clouds
or darkness. The “blending” is done at the binary mask level therefore the
output based on simple boolean logic between the maps generated by the
three types of instruments and external masks that specify cloud, night, and
weather affected pixels.
More recently [Liang et al., 2015] developed a snow depth retrieval algo-
rithm that integrates microwave brightness temperatures and visible and
infrared reflectance. In this study, the snow depth retrieval is regarded as
a regression problem that is solved by a Support Vector Machine. In this
formulation, the independent variables are the microwave brightness temper-
ature and the visible/infrared surface reflectance. The dependent variable is
the snow depth to be retrieved. Initially, a set of sample data points are used
to train the SVM and generate a regression function that maps the remotely
sensed data to snow depth. This set of sample data points was derived by
collocating satellite visible and infrared observations with ground station
measurements of snow depth in the regions covered by the satellite pixels.
Once the regression function is computed, snow depth can be easily retrieved
by applying it on new testing data.
This algorithm was validated against a number of different operational snow
depth retrieval products [Liang et al., 2015]. It was shown to be significantly
more accurate than all existing approaches at the time of publication.
27
Chapter 4
Discussion and Conclusions
This paper has presented a number of different approaches to satellite snow
and sea ice detection. With the exception of the human curated IMS product,
the operational algorithms based on both the visible and infrared, and the
passive microwave portions of the electromagnetic spectrum suffer from the
limitations of the nature of instruments they were designed for. The suite of
MODIS products does not work under cloudy and dark conditions and the
suite of passive microwave products suffer from a coarse spatial resolution
and artifacts due to weather effects. Although, the IMS product seems to
provide a reasonable alternative to these limitations, it has some of its own.
Manual snow and sea ice mapping drawn by humans, is a subjective, labor
intensive and time consuming procedure. Thus, automated snow and sea
ice detection algorithms that utilize all available sources of information and
generate output at the highest spatial resolution are to be desired.
From the small number of attempts at applying machine learning tech-
niques to this task, the results seem encouraging. Moreover, coupling these
28
techniques with a multi-sensor integrated approach seems to be the most
promising path to more accurate high quality snow and sea ice detection.
29
Bibliography
[Barnes et al., 1998] Barnes, W. L., Pagano, T. S., and Salomonson, V. V.(1998). Prelaunch characteristics of the moderate resolution imaging spec-troradiometer (modis) on eos-am1. Geoscience and Remote Sensing, IEEETransactions on, 36(4):1088–1100.
[Barry, 2002] Barry, R. G. (2002). The role of snow and ice in the globalclimate system: a review. Polar Geography, 26(3):235–246.
[Brown et al., 2014] Brown, L. C., Howell, S. E., Mortin, J., and Derksen, C.(2014). Evaluation of the interactive multisensor snow and ice mapping sys-tem (ims) for monitoring sea ice phenology. Remote Sensing of Environment,147:65–78.
[Brown and Mote, 2009] Brown, R. D. and Mote, P. W. (2009). The responseof northern hemisphere snow cover to a changing climate*. Journal of Cli-mate, 22(8):2124–2145.
[Cavalieri and Parkinson, 2012] Cavalieri, D. and Parkinson, C. (2012). Arcticsea ice variability and trends, 1979–2010. Cryosphere, 6(4):881–889.
[Comiso and Cho, 2013] Comiso, J. and Cho, K. (2013). Chapter 6: De-scription of gcom-w1 amsr2 sea ice concentration algorithm (descriptionsof gcom-w1 amsr2 level 1r and level 2 algorithm). Japan Aerospace Explo-ration Agency Earth Observation Research Center.
[Comiso et al., 1997] Comiso, J. C., Cavalieri, D. J., Parkinson, C. L., andGloersen, P. (1997). Passive microwave algorithms for sea ice concentration:A comparison of two techniques. Remote sensing of Environment, 60(3):357–384.
[Dobreva and Klein, 2011] Dobreva, I. D. and Klein, A. G. (2011). Fractionalsnow cover mapping through artificial neural network analysis of modis sur-face reflectance. Remote Sensing of Environment, 115(12):3355–3366.
[Eastwood et al., 2014] Eastwood, S., Breivik, L.-A., Andersen, S., Godøy,Ø., Lind, M., Porcires, M., and Schyberg, H. (2014). Sea ice product user’smanual. Norwegian and Danish Meteorological Institutes.
30
[Foster et al., 2011] Foster, J. L., Hall, D. K., Eylander, J. B., Riggs, G. A.,Nghiem, S. V., Tedesco, M., Kim, E., Montesano, P. M., Kelly, R. E., Casey,K. A., et al. (2011). A blended global snow product using visible, passivemicrowave and scatterometer satellite data. International journal of remotesensing, 32(5):1371–1395.
[Frei and Gong, 2005] Frei, A. and Gong, G. (2005). Decadal to century scaletrends in north american snow extent in coupled atmosphere-ocean generalcirculation models. Geophysical Research Letters, 32(18).
[Frei et al., 2012] Frei, A., Tedesco, M., Lee, S., Foster, J., Hall, D. K., Kelly,R., and Robinson, D. A. (2012). A review of global satellite-derived snowproducts. Advances in Space Research, 50(8):1007–1029.
[Grumbine, 1996] Grumbine, R. (1996). Automated passive microwave sea iceconcentration analysis at ncep. US Department of Commerce, NationalOcean And Atmospheric Administration, National Weather Service, Na-tional Centers for Environmental Prediction, Technical Note, OMB con-tribution, 120.
[Hall et al., 2005] Hall, D. K., Kelly, R. E., Foster, J. L., and Chang, A. T.(2005). Estimation of snow extent and snow properties. Encyclopedia ofHydrological Sciences.
[Hall and Riggs, 2007] Hall, D. K. and Riggs, G. A. (2007). Accuracy assess-ment of the modis snow products. Hydrological Processes, 21(12):1534–1547.
[Hall et al., 1995] Hall, D. K., Riggs, G. A., and Salomonson, V. V. (1995).Development of methods for mapping global snow cover using moderateresolution imaging spectroradiometer data. Remote sensing of Environment,54(2):127–140.
[Hall et al., 2001] Hall, D. K., Riggs, G. A., Salomonson, V. V., Barton, J.,Casey, K., Chien, J., DiGirolamo, N., Klein, A., Powell, H., and Tait, A.(2001). Algorithm theoretical basis document (atbd) for the modis snowand sea ice-mapping algorithms. NASA GSFC.
[Hall et al., 2002] Hall, D. K., Riggs, G. A., Salomonson, V. V., DiGirolamo,N. E., and Bayr, K. J. (2002). Modis snow-cover products. Remote sensingof Environment, 83(1):181–194.
[Helfrich et al., 2007] Helfrich, S. R., McNamara, D., Ramsay, B. H., Baldwin,T., and Kasheta, T. (2007). Enhancements to, and forthcoming develop-ments in the interactive multisensor snow and ice mapping system (ims).Hydrological Processes, 21(12):1576–1586.
31
[Ivanova et al., 2014] Ivanova, N., Johannessen, O. M., Pedersen, L. T., andTonboe, R. T. (2014). Retrieval of arctic sea ice parameters by satel-lite passive microwave sensors: A comparison of eleven sea ice concentra-tion algorithms. Geoscience and Remote Sensing, IEEE Transactions on,52(11):7233–7246.
[Kaleschke et al., 2001] Kaleschke, L., Lupkes, C., Vihma, T., Haarpaintner,J., Bochert, A., Hartmann, J., and Heygster, G. (2001). Ssm/i sea ice re-mote sensing for mesoscale ocean-atmosphere interaction analysis. CanadianJournal of Remote Sensing, 27(5):526–537.
[Key et al., 1997] Key, J. R., Collins, J. B., Fowler, C., and Stone, R. S. (1997).High-latitude surface temperature estimates from thermal satellite data.Remote Sensing of Environment, 61(2):302–309.
[Liang et al., 2015] Liang, J., Liu, X., Huang, K., Li, X., Shi, X., Chen, Y.,and Li, J. (2015). Improved snow depth retrieval by integrating microwavebrightness temperature and visible/infrared reflectance. Remote Sensing ofEnvironment, 156:500–509.
[Luo et al., 2008] Luo, Y., Trishchenko, A. P., and Khlopenkov, K. V. (2008).Developing clear-sky, cloud and cloud shadow mask for producing clear-sky composites at 250-meter spatial resolution for the seven modis landbands over canada and north america. Remote Sensing of Environment,112(12):4167–4185.
[McIntire and Simpson, 2002] McIntire, T. J. and Simpson, J. J. (2002). Arc-tic sea ice, cloud, water, and lead classification using neural networks and1.6-µm data. Geoscience and Remote Sensing, IEEE Transactions on,40(9):1956–1972.
[Monitoring et al., 2012] Monitoring, A. et al. (2012). Changes in arctic snow,water, ice and permafrost. swipa 2011 overview report. Arctic Climate Issues2011.
[Murphy et al., 2006] Murphy, R., Ardanuy, P., Deluccia, F. J., Clement, J.,and Schueler, C. F. (2006). The visible infrared imaging radiometer suite.In Earth science satellite remote sensing, pages 199–223. Springer.
[Musial et al., 2014] Musial, J., Husler, F., Sutterlin, M., Neuhaus, C., et al.(2014). Probabilistic approach to cloud and snow detection on advancedvery high resolution radiometer (avhrr) imagery. Atmospheric MeasurementTechniques, 7(3):799.
[Nagler et al., 2014] Nagler, T., Fernandes, R., Derksen, C., Metsamaki, S.,and Loujus, K. (2014). Digital coding of snow products for snowpex. Snow-PEX - The Satellite Snow Product Intercomparison and Evaluation Exper-iment.
32
[Oki et al., 2010] Oki, T., Imaoka, K., and Kachi, M. (2010). Amsr instru-ments on gcom-w1/2: Concepts and applications. In Geoscience and remotesensing symposium (IGARSS), 2010 IEEE International, pages 1363–1366.IEEE.
[Painter et al., 2009] Painter, T. H., Rittger, K., McKenzie, C., Slaughter, P.,Davis, R. E., and Dozier, J. (2009). Retrieval of subpixel snow coveredarea, grain size, and albedo from modis. Remote Sensing of Environment,113(4):868–879.
[Pizzolato et al., 2014] Pizzolato, L., Howell, S. E., Derksen, C., Dawson, J.,and Copland, L. (2014). Changing sea ice conditions and marine transporta-tion activity in canadian arctic waters between 1990 and 2012. Climaticchange, 123(2):161–173.
[Ramsay, 1998] Ramsay, B. H. (1998). The interactive multisensor snow andice mapping system. Hydrological Processes, 12(10):1537–1546.
[Ramsay, 2000] Ramsay, B. H. (2000). Prospects for the interactive multisen-sor snow and ice mapping system (ims). In Proceedings of the 57th EasternSnow Conference, pages 161–170.
[Riggs et al., 2006] Riggs, G. A., Hall, D. K., and Salomonson, V. V. (2006).Modis sea ice products user guide to collection 5. NASA Goddard SpaceFlight Center, Greenbelt, MD.
[Rivas, 2007] Rivas, M. B. (2007). Sea ice extent from satellite microwavesensors. Triennial Scientific Report 2007-2009, page 9.
[Romanov, 2014] Romanov, P. (2014). Global 4km multisensor automatesnow/ice map (gmasi) algorithm theoretical basis document. NOAA NES-DIS Center for Satellite Applications and Research.
[Romanov et al., 2000] Romanov, P., Gutman, G., and Csiszar, I. (2000). Au-tomated monitoring of snow cover over north america with multispectralsatellite data. Journal of Applied Meteorology, 39(11):1866–1880.
[Stroeve et al., 2015] Stroeve, J., Blanchard-Wrigglesworth, E., Guemas, V.,Howell, S., Massonnet, F., and Tietsche, S. (2015). Improving predictionsof arctic sea ice extent. EOS, 96.
[Trishchenko et al., 2006] Trishchenko, A. P., Luo, Y., and Khlopenkov, K. V.(2006). A method for downscaling modis land channels to 250-m spatialresolution using adaptive regression and normalization. In Remote Sensing,pages 636607–636607. International Society for Optics and Photonics.
[Vihma, 2014] Vihma, T. (2014). Effects of arctic sea ice decline on weatherand climate: A review. Surveys in Geophysics, 35(5):1175–1214.
33
[Zhang, 2005] Zhang, T. (2005). Influence of the seasonal snow cover on theground thermal regime: An overview. Reviews of Geophysics, 43(4).