ABSTRACT DTM generation with Radarsat-2 Data without GCP

Digital terain models (DTMs) were extracted from high-resolution Radarsat-2 stereo high-resolution images. The metadata supplied by MacDonald, Dettwiler and Associates Ltd. were used to replace the ground control data used to compute for the deterministic geometric model. The DTMs are then evaluated with 0.2-m accurate lidar elevation data. Because DTMs included the height of land covers, elevation linear errors with 68 and 90 percent confidence level (LE68 and LE90) were both computed over the full study site and the bare surfaces. Comparisons were also performed with the same rigorous model using accurate differential GPS as ground control points (GCPs). Using the radargrammetric model using eight dGPS GCPs achieved the best results (3.3 m LE68) than using only the metadata (3.9 m LE68). However, this accuracy is compensated by the fact that the user does not have to collect any ground data, which offers a strong advantage in remote and harsh environments.


Introduction
Since Leberl (1978), a large number of researchers around the world have investigated stereo-radargrammetric methods using deterministic modeling and applied to high spatial resolution (HR) of recent spaceborne s#nthetic aperture radar (SAR) sensors (Toutin, 2011).However, the# generall# require ground control points (GCPs) to accuratel# determinate the unknown parameters of radargrammetric models.While the metadata supplied b# MacDonald, Dettwiler and Associates Ltd. (MDA) with Radarsat-2 data are still limited b# the orbit (5 m at 90% for Radarsat-2) and calibration timing uncertainties (Robertson, personal communications, 2009), the# could be used to replace the GCP role for computing the unknown parameters.
The objecties of this research paper are then to apply only these MDA metadata to a radargrammetric model and to evaluate the accurac# of final 3D products.Comparisons when input accurate ground control points (GCPs) to the radargrammetric model are also performed.The radargrammetric model is the deterministic Toutin's model alread# applied to most of SAR data, including Radarsat-2 (Toutin and Chenier, 2009).The Radarsat-2 stereo data are acquired using its high-resolution (HR) mode with steep and shallow viewing angles over Canadian stud# sites.

Study site and data set 2.1 Stud# Site
The stud# site is located north of Quebec City, Quebec, Canada (470N, 71°30'W) and spans different environments: urban and residential, semi-rural and forested.The elevation ranges almost from 10 m in the city in the southeast to around 1000 m in the Canadian Shield in the north.The northern part is a hill# to mountainous topograph# (5°-30° slopes) mainl# covered with forests (deciduous, cornier and mixed) while the south part is a semi-flat topograph# (0°-5° slopes) with urban and residential areas.

Radarsat-2 Stereo Data
The Radarsat-2 SAR data set (Table 1) included two stereo images (20 b# 20 km) acquired September 10 and 14, 2008 with the C-band ultra-fine mode (1 b# 1 look; 1.6 2.4 b# 3 m resolution) in VV polarization from descending orbits with view angles of 30.8°-32° (U2, Figure 1) and 47.5°-48.3°(U25, Figure 2) at the near-far edges, respective!The VV SAR data were processed as angle look complex (SLC) product (16 bits) in the slant range geometr# and orbit oriented with a pixel spacing of 1.3 b# 2.1 m.The metadata were provided in separate files.Reserved" and Courtesy of Canadian Space Agency.

Cartographic data
The reference cartographic data included ground points, m anly road intersections and electrical poles, collected from a differential Global Positioning System (dGPS) survey in November 2008 with 3-D ground accuracy of 10-20 cm.The collected points were used either as GCPs to compute the physical/empirical models or generally as independent check points (ICPs) to quantiiy/vaidate the model accuracy.
In addition, accurate spot elevation data (first echoed return) were obtained on 06 September 2001 from a lidar survey collected by GPR Consultants (www.lasermap.com)(Fowler, 2001).Ten swaths were acquired covering a 5 km by 13 km area, which approximately corresponds to the east part of the stereo images.Since the research objectives were to evaluate stereo-extracted DSMs, the lidar elevation data were not interpolated into a regular-spaced grid to avoid the propagation of interpolation error into the checked elevation and the evaluation.
3 Description of the method The processing steps for DSM generation with HR SAR stereo-images using the deterministic Toutin's model ™ were previously addressed and documented for Radarsat-2 (Toutin, 2010).The m an processing steps are: 1.The acquisition of the SLC SAR images and metadata (orbit information and 3rdorder RFM coefficients).The SLC data were pre-processed into 16-bit amplitude images, filtered with the enhanced Lee filter (Lee, 1980, Lopes etal., 1990).Metadata were used to determine an approximate value for each parameter of TM. 2. The collection on stereo-images of 60 points from dGPS surve# (10-cm accurac#).
The m an error was due to the image pointing (less than one pixel for electrical poles and 1(-2) pixels for road intersections.The collected points spanned the total volume of the terrain relief to avoid extrapolations, both in planimetr# and elevation.The# were mainl# used as independent check points (ICP) for TM validation.3. The computation of TM computed with eight and no GCP (TM 8 and TM 0). 4. The elevation extraction used a hierarchical (7 steps) gre#-level image matching (mean normaliied cross-correlation method with sub-pixel computation of the maximum of the correlation coefficient) applied in the quasi-epipolar geometr# (Ostrowski and Cheng, 2000).The quasi-epipolar DTM was then reprojected with 6 m regular grid spacing into the cartographic projection.
The evaluation of the extracted DSMs with the lidar elevation data was finall# performed over about 5,500,000 points: mean, linear error with 68% and 90% levels of confidence, LE68 and LE90 were computed.

4
Results and discussions 4.1 Accurac# evaluation of stereo-radargrammetric models Toutin and Chenier (2009) alread# demonstrated that TM's precision is around 25 cm: the model will thus not induce an# significant errors in the next evaluation.Table 2 summ aries the results of TM computation for the two tests: the errors (bias and standard deviation, Std in meters) computed on a large number of ICPs (52 or 60) providing independent and unbiased evaluations of the modeling accurac#.The two Tests (TM 8 or TM 0) achieved consistent results.TM 8 also achieved the best Std results being the reference results for TM (Toutin, 2010).These results demonstrate that user-collected GCPs (mainl# with dGPS) still perform better than the metadata in the geometric processing, whatever the recent improvements in the non imaging sensors.On the other hand, the small lost in accurac# for TM 0 is compensated b# the gain of processing the stereo-images with no GCP.

DSM evaluation results
The second results are qualitative evaluations of quasi-epipolar DSMs generated using TM-8 (Figure 3, top) and TM_0 (Figure 4, bottom).Visually, both DSMs well describe the macro-topography and the macro linear trends with mountains and valleys, enhancing the structural geological framework in the northwest-southeast direction.However, some high-frequency topographic variations are perceptible in TM 8 (Figure 3, top) in the flatter areas.
The quantitative evaluation was performed over the coverage of the lidar data, being on the hall eastern part of DSMs.The computed accurac# would be still representative of the overall DSMs because all relief (flat to mountainous; all slopes and azimuths) and all land covers (urban, semirural, bare sols, and forested areas) of the stud# site were embraced in this sub-area.Because the C-band SAR phase scattering center typicall# occurs at 40-60% of the canop# depending on the SAR resolution and the tree characteristics (deciduous, conifer, mited, height, densit#, etc.) (Ken#i et al., 2009), the comparison with lidar elevation data in the vegetated cover can thus generate both s#stematic and random errors.Consequent^, the results computed over the full lidar coverage do not reflect the true DSM accurac# since the dominant source in the error budget is the difference between the compared stereo-SAR and lidar elevation points.not the DSM accurac#.To have the true elevation accurac#, the error evaluation was also performed onty on bare surfaces, where the stereo SAR and lidar points were at the same ground elevation.The bare surfaces were also representatiie of the full terrain relief because the# occur not onty on low lands and slopes but also in the high lands and slopes (m ainl, in the northeast).Consequent^, Table 3 gite the accurac# results (bias, LE68 and LE90 in meters) computed on bare surfaces for both DSMs.The small negative biases are coherent with the modeling results (negative Z-bias over ICPs).It seems normal that the original TM 8 consistent^ achieves the best accurac# versus TM0: 60 and 70 cm difference for LE68 and LE90, respective! because the definition, qualit# and accurac# of dGPS GCPs (10-cm cartographic and sub-pixel pointing accurac#) are better than the metadata accurac# (5 m at 90%).TM 8 should thus enable better contour lines to be derived.

CONCLUSIONS
The objectives of the research were to extract DSMs from Radarsat-2 HR stereo images acquired over a residential/rural M l# area in Quebec, Canada.Taking advantages of the improvements of the metadata the 3D deterministic Toutin's models can then be used without GCPs.In order to verify its adequac# the same processing were performed with 8 GCPs.DSMs were then compared to point-cloud lidar elevation data over bare surfaces onfy.TM 0 with no GCP achieved slightl# worse results than TM 8, both on the stereo-radargrammetric model (8-48 cm difference in the three axes) and on DTM (60 and 70 cm difference for LE68 and LE90, respectivel).This reduction of accurac# is, however, compensated b# the gain of processing stereo-images and generating DSMs with no GCP, which increases the applicabilit# of Radarsat-2 to remote and harsh environments as well as stud# sites without an# cartographic or ground control data.

Table 2 :
Number of ICPs with biases and standard deviations (Std, 1 sigma in metres) ____________________ for each modeling Test.______________

Table 3 :
Accurac# results for DSMs computed over the bare surfaces: bias, LE68 and LE90 in metres.