Science Service System
You are here : Home : Proposals_Summary

Summary of Proposal HYD0305

TitleMapping River Ice Using Dual Polarization X-Band SAR Data
Investigator Bernier, Monique - INRS, Eau, Terre et Environnement
Team Member
Professor Pottier, Eric - IETR, Remote Sensing
Gauthier, Yves - INRS, Eau, Terre et Environnement
Doctor Allain, Sophie - IETR, Remote Sensing
Mr Mermoz, Stéphane - INRS, Eau, Terre et Environnement
SummaryIn many northern rivers, the development of ice covers leads to important issues : ice jamming and therefore flooding of large areas, power reducing at a hydroelectric generating station, navigation impeding and structures damaging. The FRAZIL project (GEOIDE Network) was initiated to develop a GIS-based system in support to winter river flow modelling. An important aspect of this system is that it should benefit from an optimal use of river ice information coming from satellite radar images in general and of polarimetry in particular. Synthetic Aperture Radar (SAR) imagery is a powerful tool to monitor the hard-to-predict river ice cover, thanks to its night-and-day and all-weather viewing capabilities. SAR data provides more and more useful information every year for sea ice, soil moisture, wind, and land cover mapping. It is established that dual and quad-polarimetric data improves the accuracy of sea-ice classifications. Since 2005, new algorithms have been developed in order to improve the accuracy of river ice type classifications using mono-polarisation and polarimetric SAR data.. These studies led to encouraging results. Another objective is to understand the interaction between the radar signal and the ice cover thickness. A correlation has been found with RADARSAT-1 data in some particular ice conditions. Also, polarimetry could help for this objective. Furthermore, there are problems to get good results with thin ice in C-band. X-band data could solve this problem. The study sites are the Saint-François River, located in Southern Quebec upstream from the town of Drummondville, the Chaudiere River at Levis near Quebec City, and the third one is the Kosoak River at Kuujjuaq near the Ungava Bay in Northern Quebec. The first two are rivers where ice jams occur almost each year. The Saint-Francois has been monitored (field and SAR data collected) by INRS since 2001. The Chaudiere River area is a future calibration site for RADARSAT-2. On the Kosoak River, the ice cover is used during the winter by the Inuits as a vital access to their hunting camp. A monitoring program with RADARSAT-1 is actually in operation. Several exhaustive « In-situ » campaigns are previewed in 2008. The ice cover state will be studied and the thickness will be measured. Furthermore, if RADARSAT-2 is launched in December 2007, images could be acquired in March, 2008. The acquisition of images in the X-band by Terra-SAR would thus lead to a multifrequency study with RADARSAT-1 or RADARSAT-2 data. It is expected that the increased information will directly influence the achieving of objectives. Several supervised and unsupervised classification algorithms will be applied on data. It will permit to confirm previous results and to develop new algorithms. Particularly, a hierarchical classification of ice types, based on ratio parameters, has been developed (Bernier, 2007). The accuracy of this algorithm needs to be validated with new SAR data. The success of the project will be evaluated using the following performance indicators: - A better discrimination of ice types and more accurate classifications: The ground truth will confirm the effectiveness of this study in testing existing algorithms and improving the accuracy of classifications. - An effective algorithm to retrieve the ice cover thickness from SAR data. The final products would be 1) a reliable unsupervised algorithm which discriminate the main ice types, and 2) an algorithm which retrieves the ice cover thickness, both independently of the SAR data particularities. References : Bernier, M., Mermoz, S., Allain, S. and Pottier, E. (2007). Classification of River Ice Types using Quad-Polarization SAR Imagery. Advanced SAR Workshop 2005. (11-13 Sept. 2007).
Final ReportI. INTRODUCTION In many northern rivers of Canada, monitoring the formation and evolution of river ice is essential, as the development of ice covers in cold weather leads to important situations: ice jamming, and then flooding of large areas; reduction of power production at hydro-electric generating stations; navigation hindrance; and structural damage. Satellite based monitoring services offer an ideal solution[1]allowing decision makers to collect information on river ice repeatedly throughout the ice season. Monitoring of river ice through remote sensing has mainly focused on the use of monofrequency (C-band) monopolarized [2][3][4][5] and multipolarized [6] SAR data. Few polarimetric SAR scenes of river ice have been collected to date. The first examples of C-band full polarimetric [7] [8][9] and X-band dual-polarized [10][11]. SAR observations of river ice confirmed the purpose of combining data of different frequencies and polarizations in order to resolve ambiguities in the radar signature of ice types. With regard to the optimum choice of radar frequency, L-band could better separates types of ice [12],while[13]prefere C-band. X-band is also a good candidate[10]. As a function of the frequency, the radar will detect similar characteristics but also dissimilarities[14]. Open water versus ice is perfectly discriminated at C-band [7] and at X-band[10]. The sensitivity of the transmitted radar signal in the X-band facilitates the identification of small-scale surface features of snow and ice cover. Then, X-band data can separate better smooth versus rough ice types than C-band [15]. Moreover, some ice types cause a higher increase in the total response at X-band than C-band due to volume and volume-boundary interactions[15]. Classification can also be achieved using the Bayes theory and the complex Wishart distribution of the covariance matrix elements [16].The merging of decomposition techniques with a Wishart classifier was also proposed to conduct unsupervised classification[17].Other classification schemes such as neural networks[18], hierarchical classifiers[19], and classifiers based on wavelet transform [20] are available.In this project, we evaluate on a quantitative basis the added value of various frequencies and polarizations for mapping river ice conditions. Two classifications methods are used: a SVM and a classification and regression tree (CART). SVM is well suited to handle linearly non separable cases by using the Kernel theory[21][22][24]. The second method is a nonparametric and nonlinear algorithm. The results are compared to those obtained with the Wishart approach[17] and with the single frequency single-pol Icemap approach[4]. II. DATA The test site is the Saint-Francois River (4550N;7222W), located in southern Quebec and upstream from the town of Drummondville. The stream flow is roughly southeast to north-west. The study section is approximately 30km long. Channel width varies from 100m to 850m and the depth reaches 2-4m in general. Three Radarsat-2 full polarimetric images in quad-fine mode were acquired at C-band and two Terra SAR-X in February and March 2009 over the Saint-Francois River. Ground-truth data have been measured the same days as the satellite overpass. Sampling sites location were predetermined based on the latest ice map. III. CLASSIFICATION SCHEMES From ground-truth map of the beginning of February 2009 are selected training areas. Their microwave signatures are extracted and each image pixel are then classified into these ice conditions. This supervised procedure that require detailed independent surface and volume information provides confidence in the results, and attempts at minimizing classification errors. Except for automated Icemap approach which do not require any training samples, 500 samples for each class are selected at the training step. There are more control than training points for robustness characterization. The three river ice classes are: 1) Smooth thermal ice cover, or agglomerated frazil ice, with few and small air inclusions(SI), 2) rough agglomerated frazil ice, or slightly/moderately rough consolidated ice(II), 3)rough consolidated ice(CI). IV. CLASSIFICATION ACCURACY VERSUS FREQUENCY AND POLARIZATION The classification results were evaluated with the help of confusion matrices. At the training step, X-band always performs better than C-band. Using single polarimetric X- versus C-band SAR data, SVM and RegT Mean Accuracy (MA)increase of 0.5% and 2.2% respectively. The multifrequency VVpolarization compromise make SVM MA increasing of 2.8% compared to C-band alone. All ice classes are correctly classified in spite of an overestimation of II. When dual-pol SAR data are used, SVM as well as RegT MA increase of 5% and 3.7% at X-band compared to C-band. This is still due to the fact that CI class is logically better discriminated at X-band (80.4% versus 60.8%). Using both Xand C-band does not improve the results obtained at X-band With the addition of_HHVV,C- and X-band give almost the same results whatever the classification. The improvement induced by multifrequency data is clear and RegT mean accuracy reaches 90.2%. Full polarimetric SAR data can not be compared as X-band data are not available. Nevertheless, full-pol C-band and complex dual-pol X-band data have been combined to classify river ice types.This configuration give the best results which are significantly better than full-pol C-band data but comparable to multifrequency complex dual-pol SAR data. This time, SVM and RegT reach 88.2% and 91.8% MA. If compared with C-band single-pol data, Average Productor Accuracy (APA) increases of 13.6% and 6.1%, Average User Accuracy (AUA) increases of 9.3% and 5.2%, Kappa increases of 24.2% and 10.4% for SVM and RegT respectively. In the end, X-band classify better ice types than C-band when training and test dataset are the same. Multifrequency data show the best results. As well as at C-band, X-band RegT is powerful at the training process. Using single polarimetricX- versus C-band, the results do not improve for RegT contrary to SVM. Multifrequency VV-polarization data considerably improve results. If compared to X-band alone, SVM MA improves of 7.6%, APA reaches 75.2% and AUA 75.5%. RegT MA improves of 16.8%, APA yields 77.6% and AUA 79.4%. When one polarization or _HHVV is added, C-band is better than X-band which decreases,even if the Xband complex dual-pol data results provide good information on SI class. This corroborates our attempts because X-band is more sensitive than C-band to ice physical variations occurring between two acquisitions. Multifrequency complex dual pol data show good results for SVM, except for class CI (58.6%of PA and 84.7% of UA). The best results are reached using C-band full-pol combined with X-band complex dual-pol data Wishart. Wishart reaches 80.5% of APA and 78% of AUA. In fact, the resolution of the images is not too high and the covariance matrix can be well described by the Wishart distribution. V. CONCLUSION X-band classification parameters are less adaptable to other images compared to C-band. In fact, the ice roughness tends to decrease with time and the snow on the ice cover also varies in thickness and spatial distribution. X-band is very sensitive to these variations. Classifications combining two frequencies yield the best results. Multifrequency single-pol data are better in separating ice types than fully polarimetric single frequency data. V1. REFERENCES [1] R. Leconte, P.D. Klassen, Lake and river ice investigations in northern Manitoba using airborne SAR imagery, Artic, vol. 44, no. 1, 153-163, 1991. [2] F.Weber, D. Nixon,J. Hurley, A.A. Khan, R.Picco, Semiautomated classification of river ice types on the Peace river using RADARSAT-1 synthetic aperture radar (SAR) imagery: Canadian J. of Civil Eng.vol. 30, pp. 11-27, 2003. [3] T.M. Puestow, C.J. Randell,K.W.Rollings,A.A.Khan,R.Picco,Near real-time monitoring of river ice in support of flood forecasting in eastern Canada, IEEE TGRSS,vol.4,2268-2271,2004. [4] Y. Gauthier, F. Weber, S. Savary, M. Jasek, L.M. Paquet,M. Bernier, A combined classification scheme to characterize river ice from SAR data,In EARSeL, eProceedings, vol. 5, no. 1,77-88,2006. [5] K.D. Unterschultz, J. Van Der Sanden, F. Hicks, Potential of RADARSAT-1 for the monitoring of river ice: Results of a case study on the Athabasca River at Fort McMurray, Canada, Cold Regions Science and Technology, vol. 55, no. 2, 238-248, 2008. [6] K. Russell, S. Warren, C. Howell, T. Puestow, C. Randell, A. Khan, C. Mahabir, P. Tang, D. Burakov. N. Novik, Improved Satellite-Based River Ice Monitoring Using Dual-Polarized SAR Imagery, 15th Workshop of the CGU HS Committee on River Ice Processesand the Environment, 401-406, St. John’s, Canada, June 15-17, 2009.[7] S. Mermoz, S. Allain, M. Bernier, E. Pottier, I. Gherboudj, Classification of river ice using polarimetric SAR data, Canadian J. of Rem. Sens. October, 2009. [8] S. Mermoz, S. Allain, M. Bernier, E. Pottier, River Ice Mapping from PolSAR Images, IGARSS08, July 6-11, Boston, Ma. CD-ROM. [9] J. van der Sanden, H. Drouin, F.E. Hicks and S. Beltaos, Potential of Radarsat-2 for the monitoring of river freeze-up processes, Proceedings of the 15th Workshop of the CGU HS Committee on River Ice Processes and the Environment, St. John’s, Newfoundland and Labrador, June 15- 17, 2009. [10] S. Mermoz, S. Allain, M. Bernier,E. Pottier, Investigation of Radarsat-2 and TerraSAR-X data for river ice classification, IGARSS09, July 13-17, Cape Town. [11] S. Mermoz, Y. Dribault, S. Allain, M. Bernier, E. Pottier, Y. Gauthier,Investigation of Radarsat-2 and TerraSAR-X data for river ice characterization from remote sensing,15th Workshop of the CGU HS Committee on River Ice Processes and the Environment, St.Johns, June 15-17, 2009. [12] R.A. Melloh,L.W. Gatto, Interpretation of passive and active microwave imagery over snow-covered lakes and rivers near Fairbanks Alaska, Proceedings of the Applications of remote sensing in hydrology Workshop, Saskatoon,259-278, 1990.[13] M.O. Jeffries, K. Morris,W.F. Weeks, Structural and sratigraphic features and ERS-1 synthetic aperture radar backscatter characteristics of ice growing on shallow lakes in NW Alaska, Journal of geophysical research, vol. 99(C11),no.22,459-471,1994. [14] D.K. Hall, Remote sensing applications to hydrology: imaging radar,Hydrological sciences, vol. 41, no.4,609-624, 1996. [15] I. Gherboudj, M. Bernier, R. Leconte, A backscatter modeling forriver ice: analysis and numerical results, IEEE TGARS, vol.48, no.4,2010. [16] J.S. Lee, M.R. Grunes,R. Kwok, Classification of multi-look polarimetric SAR imagery based on complex Wishart distribution, Int.J. Remote Sensing, vol.15,no. 11,2299-2311,1994. [17] J.S. Lee, M.R. Grunes, E. Pottier, L. Ferro-Famil,Unsupervised terrain classification preserving polarimetric scattering characteristics,IEEE TGARS, vol.42, no.4,722-731, 2004.[18] F. Del Frate and D. Solimini, On neural network algorithm for retrieving forest biomass from SAR data, IEEE Trans. Geosci. Remote Sensing,vol.42, no.1, pp.24-34,2004. [19] A. Freeman, J. Villasenor, J.D. Klein, P. Hoogeboom, J. Groot, Onthe use of multi-frequency and polarimetric radar backscatter features for classification of agricultural crops, Int. J. Remote Sens.,vol.15, no.9, 1799-1812,1994. [20] L.J. Du, J.S. Lee, K. Hoppel,S.A. Mango, Segmentation of SAR image using the wavelet transform, Int.J.Imaging Syst. Technol.,vol.4, pp. 319-329, 1992. [21] C.P. Tan, H.T.Ewe,H.T.Chuah, Hybrid entropy decomposition and support vector machine method for agricultural crop classification,In PIERS Online,vol.3,no.7,620-624, 2007. [22] C.J. Burges, A tutorial on support vector machines for pattern recognition,in Data Mining Knowledge Discovery, U. Fayyad, Ed. Norwell,MA: Kluwer, 1998.

Back to list of proposals

© DLR 2004-2016