MONITORING OF SUBARCTIC SHRUB VEGETATION CHARACTERISTICSUSING TERRASAR-X AND RADARSAT-2 DATA

Yannick Duguay1,3, Monique Bernier1,3, Esther Lévesque2,3, Benoît Tremblay4

1 Institut National de la Recherche Scientifique, Québec, Canada2 Université du Québec à Trois-Rivières, Trois-Rivières, Canada

3 Centre d'études nordiques, Québec, Canada4 Ministère du Développement durable, de l'Environnement et de la Lutte contre les changements climatiques

ABSTRACT

Due to its sensitivity to the vertical structure ofvegetation, SAR remote sensing has the potential tobecome a useful tool to map changes in shrubvegetation in subarctic environments. In this paper weevaluate the sensitivity of RADARSAT-2 andTerraSAR-X data to subarctic shrub vegetation.Backscattering was compared with field measurementsof shrub height and regression analyses wereperformed to estimate the sensitivity of SAR signal toshrub height. An unsupervised Wishart classificationwas also performed on the RADARSAT-2 images toevaluate the separability between shrubs and otherelements of the environment. RADARSAT-2 proved tobe the most sensitive to shrub height with R² valuesreaching 0.64 for σ0

HV for fall acquisitions and losinglittle sensitivity during the winter. The unsupervisedclassification provided good separability betweenshrub dominated areas and other land cover types andwas able to distinguish high shrubs from low shrubs.

Index Terms— SAR, Shrub, Subarctic,RADARSAT-2, TerraSAR-X

1. INTRODUCTION

The arctic and subarctic environments are undergoingrapid changes due to warming air temperaturesobserved in the past decades. This affects manyaspects of the subarctic ecosystems, from the thawingof permafrost to the changes in vegetation cover andtheir combined effect on the local populations as wellas the fauna. One process which has received someattention is the expansion of the shrub vegetation atthe expense of the tundra [1]. The presence of shrubscan trigger a feedback loop where shrubs capturewindblown snow, keeping ground temperatures

relatively warm during the winter and accelerating theonset of growth in the spring. This can also have aneffect on the permafrost since the ground temperatureswill remain warmer in the winter. While a few studieshave investigated the sensitivity of SARbackscattering to shrub vegetation parameters [2], [3],none have specifically turned their attention tosubarctic environments. This paper studies thesensitivity SAR remote sensing data to shrubvegetation characteristics, in particular height andground coverage, in subarctic environments. Aclassification is then performed in order to map thevarious types of shrub coverage.

2. METHODS

2.1. Study area

The study area is a 60 km2 region situated around theUmiujaq community (56.55° N, 76.55° W) on theeastern shore of the Hudson Bay in northern Quebec,Canada. It is a discontinuous permafrost zonepositioned at the northern treeline, forming a transitionbetween the boreal forest to the south and the tundra tothe north. The area can be divided into two distinctenvironments: the coastal region to the west and theTasiapik valley to the east. The vegetation in thecoastal region is very sporadic and dominated bytundra vegetation with a few clusters of shrubs. TheTasiapik valley is mainly shrublands with somepatches of lichens as well as clusters of Black spruce(Picea mariana).

2.2. Datasets

A series of RADARSAT-2 Single Look Complex(SLC) Fine Quad-Pol (FQ) scenes (HH, HV, VH, VVpolarisations) as well as Dual-polarised TerraSAR-XSingle Look Slant Range Complex (SSC) Strip Map(SM) scenes (HH, HV polarisations) were acquired

1194978-1-4799-5775-0/14/$31.00 ©2014 IEEE IGARSS 2014

over the study area from October 2011 to April 2012.All the acquisitions were made on descending orbitswith two incidence angle modes, one at low incidencewith θ≈27° and one at high incidence with θ≈38°. Thechoice for the orbit and incidence modes was made inorder to maximize the coverage of the study area withboth sensors while capturing a good range ofincidence angles. The winter data were acquiredduring March and April when the snow cover is at itsmaximum. The fall data are needed as reference snow-free data and have been acquired between late Octoberand early December, at the start of the ground freeze-up and when the snow cover is still shallow and hasnegligible influence on SAR backscattering.

In-situ measurements of shrub characteristics havebeen collected during the summer of 2009. A total of238 circular plots with a 10 meter diameter weresampled. Each species of shrub and tree wereidentified, the percentage of the ground that they coverwithin the plot was measured as well as the height foreach species. Six field campaigns were carried out incoordination with satellite data acquisitions during the2011 and 2012 winter seasons. Snow depth, densityand snow water equivalent, as well as groundtemperatures were measured over various terrain typesat locations corresponding to vegetation samplingsites. Snowpits were dug at selected sites to gatherinformation on particle size and shape in addition tosnow densities from the different layers of thesnowpack.

2.3. SAR Processing

In order to assess the sensitivity of SAR backscatteringto shrub height and percent ground coverage, a

regression analysis is performed. The backscatteredsignal in each polarisation for both RADARSAT-2 andTerraSAR-X data is plotted against the weighted meanof shrub height, which is the height of each shrubspecies weighted by the percentage of ground it coverswithin each plot. The model used for regressionanalysis takes the form:

σ0fpqθ=A+B*log(h)

where σ0fpqθ is the backscattering coefficient (in dB)

at frequency f, polarisation pq and incidence angle θ,A is a constant, B is the slope or sensitivity coefficient(in dB/log of height) and h is the weighted mean ofshrub height. The log transformation was necessary asshown in Figure 1, where it can be seen that themajority of measurements are concentrated below0.5 meter mean height with a dynamic range of over10 dB.

An unsupervised classification is also performed toassess the separability of the various types of shrubcover. The classification is based on the Wishartclassifier applied to the coherency matrix derived fromthe RADARSAT-2 data and using eight predefinedzones in the plane formed by entropy and αpolarimetric parameters [4].

3. RESULTS

3.1. Sensitivity ro shrub height

The first results from the regression analysis show thatthe C-band σ0

HV is the most sensitive to shrub heightwith the B coefficient ranging from 2.3 dB to 2.9 dBand R² values between 0.50 and 0.64 for imagesacquired between October and December. Theincidence angle has little effect on the sensitivity

Figure 1: RADARSAT-2 Backscattering in HV polarisation as a function of vegetation height. SAR data were acquired onOct 19 2011 at 38° incidence. Left: weighted average of vegetation height. Right: vegetation height on log scale.

1195

coefficient, but the data at θ=27° have slightly lowerR² values between. For the same period, copolarizedbacksctattering displays lower sensitivity with a Bcoefficient varying between 0.7 dB and 1.8 dB and R²values ranging from 0.14 to 0.48 for both HH and VVpolarisations. However, incidence angle has moreimpact on the copolarised signal and lower incidence(θ=27°) displays a slightly lower sensitivity (0.7 dB ≤B ≤ 1.4 dB) and significantly lower R² values (0.14 ≤R² ≤ 0.27). During the late winter, in the presence ofsnow, the sensitivity of HV backscattering decreasesvery slightly (2.0 dB ≤ B ≤ 2.2 dB) and the fitdeteriorates slightly (0.35 ≤ R² ≤ 0.45).

In the case of TerraSAR-X data, the sensitivity toshrub height is lower and the signal seems to saturaterapidly compared to RADARSAT-2. The sensitivity ofσ0

HV ranges between 1.3 dB and 1.8 dB in the fall (0.37≤ R² ≤ 0.47) and only decreases slightly in the winterdown to 1.0 dB ≤ B ≤ 1.4 dB. As for σ0

HH there is nosignificant difference between fall and winter with theB coefficient which ranges between 0.8 dB and 1.4 dBwith R² values between 0.10 and 0.40. There are alsono noticeable differences between the two incidenceangles for the TerraSAR-X data, for both co- andcross-polarised data.

Overall, RADARSAT-2 data is more sensitive toshrub height, mainly due to its longer wavelengthenabling better penetration into the canopy comparedto TerraSAR-X. The highest sensitivity is observed

with σ0HV which is mainly influenced by the volume

scattering generated by the randomly orientedbranches of the shrubs. This sensitivity is slightlydeteriorated by the presence of snow during thewinter, but volume scattering by shrubs is stilldominant.

3.2. Wishart classification

The most interesting results with the Wishartclassification were obtained with the RADARSAT-2image acquired on October 19 2011 at incidence angleθ=38° (Figure 2). The map generated from theunsupervised classification scheme was comparedwith high resolution (0.8 m) aerial photographs of thearea in order to assign labels to the classes defined bythe algorithm and to do a basic visual assessment ofthe quality of the classification. The algorithmgenerated 8 classes which were labeled as: bare soils,exposed bedrock, high shrub, low shrub, lichen withsparse shrub, rock boulders, trees and buildings andwater bodies. The “lichen with sparse shrub” class wascovered with 50% or less shrubs, the “low shrub” classwas covered by at least 50% of shrubs with heightbelow 1.5 meters and the “high shrub” class wascovered by at least 50% of shrubs higher than1.5 meters. The double bounce generated by sometrees, especially on river banks and steep slopes in theTasiapik valley area, lead to an aggregation of theseareas with others containing buildings in the village(upper left corner of the image) which resulted in the“Trees and buildings” class. Since humanconstructions are not the focus of this research, theseareas will be masked in further analyses. The areasdominated by single bounce scattering (excludingwater) were classified as bedrock, bare soils or lichenwith sparse shrubs. While there is littlemisclassification between these three classes and theshrub dominated classes, there is some confusionwithin them. Generally, the areas classified as bedrockare overestimated to the expense of bare soils andlichen with sparse shrub classes. In the case of the twoshrub classes there is little confusion between them,indicating a good separability between various shrubheight classes, in addition to being well discriminatedfrom other land cover types. Further classificationmethods using supervised classification techniqueswill be tested to improve the precision and to see if amore fine grained separability can be achievedbetween shrub species or height classes.

Figure 2: Wishart classification of Oct. 19 2011RADARSAT-2 image. The letters on the map represent A)the Umiujaq village, B) the coastal area, C) the Tasiapik

valley. The classes are: ▲ Bedrock (light grey), ▲ Boulders(dark grey), ▲ Water (blue), ▲ Bare soil (yellow),

▲ Lichen with sparse shrub (light blue), ▲ Low shrubs(light green), ▲ High shrubs (dark green), ▲ Trees and

buildings (purple), ▲ No data (black).

1196

4. CONCLUSION

Multifrequency and multipolarised SAR data wereanalyzed to assess the sensitivity of the backscatteringcoefficient to shrub vegetation height and groundcoverage. C-band cross-polarised backscatteringdisplays the most sensitivity to the weighted mean ofshrub height, while TerraSAR-X tends to saturaterapidly. The effect of the vegetation height is stillimportant during the winter and needs to beconsidered for snow characterisation using SAR data.An unsupervised classification using the Wishartclassifier on the polarimetric RADARSAT-2 datayielded interesting results. The next steps will beaimed at improving the vegetation classification usingsupervised classification methods. A betterclassification of the structural characteristics ofvegetation (height, coverage, branch structure) willprovide good indicators for snow accumulation areasand help in monitoring the growth of shrub vegetationin subarctic environments.

5. ACKNOWLEDGEMENTS

The authors would like to acknowledge the financialsupport of the ARCTICNET Excellence Network andthe Center for Northern Studies (CEN). Y. Duguay gota Ph.D. scholarship from NSERC Canada.

6. REFERENCES

[1] M. Sturm, J. Schimel, G. Michaelson, J. M.Welker, S. F. Oberbauer, G. E. Liston, J. Fahnestock,and R. V.E., “Winter biological processes could helpconvert arctic tundra to shrubland,” BioScience, vol.55, no. 1, pp. 17–26, 2005.

[2] H. B. Musick, G. S. Schaber, and C. S. Breed,“AIRSAR Studies of Woody Shrub Density inSemiarid Rangeland: Jornada del Muerto, NewMexico,” Remote Sensing of Environment, vol. 66, no.1, pp. 29–40, Oct. 1998.

[3] T. Svoray, M. Shoshany, P. J. Curran, G. M. Foody,and A. Perevolotsky, “Relationship between green leafbiomass volumetric density and ERS-2 SARbackscatter of four vegetation formations in the semi-arid zone of Israel,” International Journal of RemoteSensing, vol. 22, no. 8, pp. 1601–1607, 2001.

[4] J. S. Lee, M. R. Grünes, T. L. Ainsworth, L. J. Du,D. L. Schuler, and S. R. Cloude, “Unsupervisedclassification using polarimetric decomposition andthe complex Wishart classifier,” Ieee Transactions onGeoscience and Remote Sensing, vol. 37, no. 5 pt 1,pp. 2249–2258, 1999.

1197