ASSESSMENT OF RADARSAT-2 HR STEREO DATA OVER CANADIAN NORTHERN AND ARCTIC STUDY SITES

Digital surface models (DSMs) extracted from high-resolution Radarsat-2 (R2) stereo images using a new hybrid radargrammetric modeling developed at the Canada Centre for Remote Sensing are evaluated over two Canadian northern and arctic study sites. Because the new hybrid model uses the full metadata of R2, it does not require any ground control point. The first study site in the north of Quebec is used for the scientific validation where accurate checked data (dGPS, lidas) is available. The second study site in the Arctic (steep relief and glaciated surfaces) is challenging for the operational evaluation of topographic mapping capabilities of R2. For the first study site, the bias and elevation linear errors with 68 percent confidence level (LE68) of R2 stero-extracted DSM compared to lidar data were computed over bare surfaces: LE90 of 3.9 m and no bias were achieved. For the second study site the comparison was performed between the R2 DEM and ICESat data. A negative 18-m bias was computed and certainly results suggests a bias in the stereo-model of R2 and thus in the metadata used in the model computation because there is few temporal variation in the data acquisition (R2 and ICESat)/ LE68 of 28 m was obtained. However, the differential melting and thinning depending of the glaciers elevations and planimetric surging of glacier tongues with less accumulation of debris and moraines, a lower LE68 of around 20 m could be expected. In addition to evaluate the potential of R2 over ice bodies, which generally have low slope relief and because the errors are strongly correlated with slopes, other statistical results of elevation differences were also computed: LE68 of 15 m was obtained over ice fields with 0–5° slopes while a little more than 20-m over less than 30° slopes was achieved.


l. INTRODUCTION l.l Canadian Context
Canadian Arctic suffers from poorly-known relief.In addition, the surface state of g laciated reg ions is rapidly evolving due to snowfall, snow transport by wind and/or surface melt, remote sensing data (Synthetic Aperture Radar (SAR) or optical) used to retrieve the topography must be acquired with the shortest possible time interval to marimize their coherence or correlation.This is also important because the flow of the g lacier (up to a few meters per day) during this time interval can bias the topog raphic measurement.
Stereo radargrammetry using synthetic aperture radar (SAR) data can thus be an appropriate solution, even with multi-date across-track capability, due to different SAR advantages specific to ice regions.First, the backscatter of SAR sensors is more dependent of the rugosity or the dielectric; component, which enable more radiometric contrast over ice fields with supraglacial debris, rock gllciers, moraines, etc.Second, the SAR sensors are operated in all-weather conditions and not dependent of the solar illumination conditions, which thus cancdled the large shadowed areas with optical data in hith latitudes.Third, the convergence of heliosynchronous orbits to North/South poles combined with a large rang e of look angles (20°-60°) gives thus a strong advantag e to drastically reduce the temporal variations to few days between the multi-date stereoimages acquired in the highest latitudes.Fourth, the new * Corresponding author.satellite SAR sensors have now hith-resolution (HR) capabilities (sub-to few meters), and are dedicated toward operational applications.
1. Due to the remote and harsh environments of the Canadim Arctic, the 3D geometric; processing of SAR images should require no ground control points (GCPs) collected by users for the operational applications.A new hybrid radargrammetric model recently developed for Radarsat-2 (R2) at the Canada Centre for Remote Sensing was thus used for the stereo-modeling and the generation of ditital surface models (DSM) (Toutin and Omari, 2011).In order to evaluate the performance of the process in a well-controlled environment as well in an operational environment, two Canadian northern study sites were used: the first one for the scientific; validation and the second one for the operational evaluation in Arctic.

Canadian Northern Study Sites
The first study site is located north of Quebec City, Quebec, Canada and spans different environments: urban and residential, semi-rural and forested (Fiture 2).The elevation ranges almost from 10 m in the city in the southeast to around 1000 m in the Canadim Shield in the north.The northern part is a hilly to mountainous topography (5°-30° dopes) manly covered with forests (deciduous, conifer and mixed) while the south part is a semi-f^t topography (0°-5° dopes) with urban and residentiil areas.
The second study site is located in the Baffin Island, Nunavut at approximately 70° 50' N and 71° 30' W (Fiture 1).There is no vegetation cover, except small plants.More than 80% is covered by ice fields with cirque glaciers (permanent icecovered mountains), outlets and valley glaciers and glaciers tongues surrounded by spectacular fiords with 70°-90° cliffs of 500-800 m heitht.Bare surface mountains also with steep dopes surround the ice fields and glaciers.The valley glacier in the south-east is about 1-km wide with up-to-4° dopes surrounded by 600-m heitht bare surffce and cirque glaciers.The elevation ranges from sea level to 1840 m and the dopes vary from 0° to 90° at ford cliffs, illustrating a very challengi ng environment (in terms of land cover and relief).
November 2008 with 3-D ground accuracy of 10-20 cm.The collected points were used either as independent check points (ICPs) to quantify/validate the new hybrid model accuracy.In addition, 10-cm accurate cloud-point data (first echoed return) were obtained froma lidar survey collected by GPR Consultants.For the second study site, the R2 stereo images were acquired in 2009 from descending orbit with the ultra-fine mode (U mode with 3-m resolution) in HH polarization: U12 and U26 (20 by 20 km; 1.6 by 2.5-3.0 m pixel spacing) on September 28 with 38.83°-39.84°right look angles and October 09 with 48.12°-48.93°look angles, respectively (Figure 3).The radiometry of these stereo-data are however, dominated by the geometric; issues due to high relief with no vegetation cover: more severe layover in U12 over the east-oriented dopes and more occluded areas over the west-oriented sopes in U26.In fact, the southnorth curved land-water boundary of the left islnd represents the cliff layover over the ocean, and not the "smoother" coast line of a gllciil-eroded ford, and part of the low lands dong this coastline cliff thus "disappeared".On the other hand, it is almost impossble to discriminate the true water-land boundary for the opposite coastline cliffs due to the SAR shadow/occluded areas.We can notice on Fiture 3 that all coastline cliffs and most cirque glaciers as having strong sopes (60°-90°) while the ice fields with their outlet glaciers have in general lower sopes (0 20°).However, steep 20°-90° dopes also occurred in some ice fields.
The lidar ICESat data (ascending and descending tracks) over the 27F13 map sheet was extracted from GLA14 product (L2

DESCRIPTION OF THE PROCESSING STEPS
The processing steps for DSM generation with HR SAR stereoimages were previously addressed and documented (Toutin, 2010).The new hybrid Toutin's model developed for the radargrammetric processing of R-2 at CCRS does not require any GCP collected by the user (Toutin and Omari, 2011): it only uses the information in the meta-data of the images for computing its parameters.The hybrid model has been proven to be 25-cm precise (Toutin and Chenier, 2009), and the accuracy (1 o) of the results in stereoscopy was better than one pixel with one-pixel biases in the three axes.The m an processing steps are: 1. Acquisition and pre-processing of the SLC SAR images and metadata; 2. Collection of 60 ICPs from the dGPS survey; 3. Computation of hybrid models and their validation with ICPs (systematic and random errors); 4. Elevation extraction used a 7-step hierarchical greylevel image matching performed on the quasi-epipolar stereo-images and geocoding of this epipolar DSM (Ostrowski and C h en , 2000); 5. Evaluation (systematic and random errors) of the geocoded DSMs with the lidar elevation data.

RESULTS AND DISCUSSIONS
Results are first retted to the systematic and random errors (1) of the hybrid radargrammetric; models computed over ICPs and (2) of the stereo-extracted DSMs computed over the lidar elevation data.

Radargrammetric model evaluation
Because there was no control data in the second study site, Table 1 only summarizes the results of the radargrammetric; modeling computation for the first study site and data set previously described: the errors (bits and standard deviation, Std in meters) computed over 60 ICPs providing independent and unbiased evaluations of the modeling accuracy.Biases of one pixel (or half SAR resolution) or slightly worse for Ydirection are obtained.Similarly, Std results in the order of one pixel are better for X-direction.It is certainly due to the better knowledge of the range direction than the aZmuth direction corresponding to the satellite displacement.Both results in Zdirection are also very good versus the SAR resolution and the same-side stereo-geometry.These results are comparable (10% difference), but a little worse in the Y-direction, to the orginal radargrammetric; model computed with user-collected GCPs (Toutin and Chenier, 2009).On the other hand, the small lost in accuracy for the hybrd model is compensated by the gain of processing the stereo-images without GCP.

*JL
Fiture 3. R2 DSM ofthe first study site The quantitative evaluation was performed over the coverage of the lidar data, being on the half eastern part of DSM (Fiture 3).The computed difference between R2 DSM and lidar data (Table 2) would be still representative of the overall DSM: a l relief (flat to mountainous; a l dopes and aZmuths) and a l land covers (urban, semirural, bare sols, and forested areas) of the study site were embraced.

Surfaces
Bias (m) LE68(m) A l surfaces 6.6 7.5 Bare surfaces 0.1 3.9 Table 2. Differences between lidar data and R2 DSM over different surface types: Bias and LE68 in metres However, the results computed over the full lidar coverage (Table 2, 1st line) do not refect the true DSM accuracy since the dominant source in the error budget comes from: (1) the footprint and penetration in the vegetated cover are different for both sensors (SAR and lidar); and (2) the compared stereo-SAR and lidar points are not at the same elevation in the vegetated cover (70% of lidar coverage).These errors are thus refected in the 6.6-m bias and 7.5-m LE68.To have the true elevation accuracy, the error evaluation was performed onfy on bare surfaces (Table 2, 2nd line) where the stereo SAR and lidar points were at the same ground elevation.Almost no bias and 3.9-m LE68 is thus obtained.The bare surfaces were also representative of the full terrain relief because they occur not onfy on low lands and dopes but also in the hith lands and dopes (mainfy, in the northeast).

3.2.2
Second study site: The R2 DEM is displayed in Fiiure 5 with the ice field and glaciers boundaries (in red) and supraglacial debris and morianes boundaries (in blue).The DEM looks relativefy smooth over the ice fields, even with the backscatter homogeneity in ice covered areas, mainfy due to the choice of the matching parameterization.In addition, the planned R-2 acquisition at the end of the melt season increased the roughness and thus the radiometric contrast.During this period, the ice covered by supraglacial debris and dust offered its maximum degree of surface texture.It is the m an reason of the few percent mismatched areas.Conversefy, the DEM looks very strange along all coast cliffs displaying over 50° dopes and large geometric and radiometric differences between the two images.The combination of these geometric and radiometric distortions, which onfy occurred in such a challenging Arctic study site, would impede any image matching.
The quantitative evaluation was performed with ICESat lidar.The heitht measurements of land surface will be the prime interest for DEM quality assessments (Zwalfy et al., 2002).
While the total number of ICESat data from ascending and descending orbits is limited to thousand points (Fiiure 5, blue and red footprints), it will be more limited because ICESat accuracy strongfy degrades with dopes.Because the primary goal of ICESat is to measure inter-annual and long-term variations in the polar ice-sheet elevation and volume of Greenland and Antarctica, there were no absolute validation results over more than 5° dopes.Consequent! as a function of the expected accuracy for R2 DSM, we onfy conddered for R2 DSM evaluation the ICESat data on dopes less than 30° (Fiture 5, blue footprints).The two biases suggest a systematic error in the stereo-model of R-2 and thus in the metadata used for this stereo-model computation.The differential bias could be due to elevation change (thinning, surging) over ice fields because there is few years difference in the data acquisition.The LE68s are relativefy sm lar but worse for a l surfaces due to more steep dopes outdde the ice field boundaries.
As mentioned before for the differential melting and thinning depending of the glaciers elevations and planimetric surging of glacier tongues with less accumulation of debris and moraines (Schwitter and Raymond, 1993), LE68 of R-2 DEM should thus be lower, around 20 m.In addition, most of ice bodies have low slope relief, other statistical results of elevation differences between ICESat and R-2 DEM show that LE68 is strongly corrected with dopes: LE68 of 15 m was then computed over ice fields with 0-5° slopes while a little more than 20-m over less than 30° dopes was achieved.

" CONCLUSIONS
A new hybrid radargrammetiic model, which does not require any user-collected GCP, was evaluated with R2 stereo-data for DEM generation over two study sites in the north of Canada.
The first site having accurate control data enabled DSM accuracy of 3.9 m (LE68) with no bits over bare sols to be obtained.
The second site, a challenging environment with glacitted surfaces, fords and steep relief, was used to evaluate the mapping potential of the method in the Canaditn Arctic without control data.In this remote and harsh environment, DSM LE68 of 28 m with targe bits (-18 m) was achieved over ice field.This major parr of this bits is certainly due to a systematic eror in the metadata and partially due the ice thinning.
While other methods using optical and SAR systems could achieved similar and even better results, this application demonstrated the capability of R2 to generate DSM with better than 20 m LE68 without collecting control data over ice bodies depending ofthe terain slopes (0-30°) or around 15 m over ice sheets (dopes less than 5°).This new method increases the applicability of R2 to remote and harsh environments.The sith t loss in accuracy when usng dGPS is then compensated by the gain of no control data. Zwally

Figure 1 .
Figure 1.Top: Perspective view from south to north ofthe second study site, generated with Google Earth usng images from TerraMetrics and WorldView.© 2010 Google and Images © 2010 TerraMetrics and DigitalGlobe.Bottom: Ortho-rectified Landsat-7 with 1:50,000 map sheet grid 1.3 Data SetFor the first study site, the R2 SAR data set included two stereo images (20 by 20 km) acquired September 10 and 14, 2008 with the C-band ultra-fine (U) mode (1 by 1 look; 1.6-2.4 by 3 m resolution) in VV polarization from descending orbits with view angles of 30.8°-32° (U2 Fiture 1) and 47.5°-48.3°(U25) at the near-far edges, respectively.The reference cartographic data included ground points, manly road intersections and electrical poles, collected from a dfferentiil Global Positioning System (dGPS) survey in

Table 2 .
Differences between ICESat blue footprints and R2 DSM over different surface types: Bias and LE68 in metres Fiiure 4. R2 DSM of the second study site with ice field and glacier boundaries (in red) and supraglacial debris and moraine boundaries (in blue) overlaid.