Wind Speed Retrieval from Simulated RADARSAT Constellation Mission Compact Polarimetry SAR Data for Marine Application

RADARSAT Constellation Mission (RCM) compact polarimetry (CP) data were simulated using 504 RADARSAT-2 quad-pol SAR images. These images were used to samples CP data in three RCM modes to build a data set with co-located ocean wind vector observations from in situ buoys on the West and East coasts of Canada. Wind speeds up to 18 m/s were included. CP and linear polarization parameters were related to the C-band model (CMOD) geophysical model functions CMOD-IFR2 and CMOD5n. These were evaluated for their wind retrieval potential in each RCM mode. The CP parameter Conformity was investigated to establish a data-quality threshold (>0.2), to ensure high-quality data for model validation. An accuracy analysis shows that the first Stokes vector (SV0) and the right-transmit vertical-receive backscatter (RV) parameters were as good as the VV backscatter with CMOD inversion. SV0 produced wind speed retrieval accuracies between 2.13 m/s and 2.22 m/s, depending on the RCM mode. The RCM Medium Resolution 50 m mode produced the best results. The Low Resolution 100 m and Low Noise modes provided similar results. The efficacy of SV0 and RV imparts confidence in the continuity of robust wind speed retrieval with RCM CP data. Three image-based case studies illustrate the potential for the application of CP parameters and RCM modes in operational wind retrieval systems. The results of this study provide guidance to direct research objectives once RCM is launched. The results also provide guidance for operational RCM data implementation in Canada’s National SAR winds system, which provides near-real-time wind speed estimates to operational marine forecasters and meteorologists within Environment and Climate Change Canada.


Introduction
Wind speed retrieval using spaceborne scatterometer and synthetic aperture radar (SAR) data is a mature field with operational implementation in many countries.The most widely-used algorithms are the C-band model (CMOD) family of geophysical model functions.These models were based on the relationship of the C-band vertical-transmit vertical-receive (VV) backscatter with wind speed, wind direction and incidence angle, derived from scatterometer data [1,2].Early development based on European Remote Sensing (ERS-1) scatterometer data resulted in CMOD-IFR2 [1], with additional development producing CMOD4 and CMOD5, to arrive at CMOD5n [2] and, most recently, CMOD7 [3].SAR imagery applications have employed the CMOD family of models and have also developed SAR-specific models to take advantage of other polarizations.The horizontal-transmit horizontal-receive (HH) SAR backscatter was associated with the CMOD VV backscatter through analysis of two CMOD models (CMOD-IFR2 and CMOD5n) and three RCM modes.We also performed a model comparison between CP and linear CMOD-related parameters.We concluded with a forward look at future data products derived from RCM CP and linear data.

Data
2.1.1.C-band SAR Data RCM data were simulated from RADARSAT-2 Fine Quad (FQ) data using the RCM simulator (v.3.1).A total of 504 FQ SAR images were acquired for the period 2008 to 2012, over Environment and Climate Change Canada (ECCC) meteorological buoys.Three RCM modes were simulated: Low Resolution (LowRes), Low Noise (LowNoise) and Medium Resolution (MR50).The specifications of the selected beam modes are shown in Table 1.The simulation options are specification noise floor (NESZ) and speckle filtering using a Sigma-Lee filter with a window size of 7 × 7 and a target size of 5 × 5; no second filter run was used.Backscatter values were in sigma-naught (σ 0 ).Post-simulation image processing included the derivation of additional CP parameters and image georectification.Twenty CP parameters and five linear polarization parameters were used in this study (Table 2).Image sampling was performed using a 3 × 3 km region of interest (ROI), centred on a meteorological buoy.The mean and standard deviation for each parameter were calculated for each ROI, for each RCM mode simulated.Therefore, the data set comprised 1512 samples (504 × 3 modes).Sample ROIs encompassed ~1915 pixels for the LowNoise and LowRes modes, and ~7680 pixels for the MR50 mode.Parameter values were normalized to decibels (dB) when appropriate for more reliable calculation of the statistics.The ECCC meteorological buoys used in this study are in marine areas on the East and West coasts of Canada (Figure 1); these areas have calibrated and well-maintained buoys.The buoys report hourly wind speed and wind direction measured by two sensors, averaged over an eight-minute interval.The observed wind speed was converted to winds at a 10 m reference height above the ocean surface using the Tropical Ocean and Global Atmosphere Coupled Ocean-Atmosphere Response Experiment (TOGA COARE) bulk flux algorithm [23].Wind direction must be relative to the satellite look direction.This was calculated by adjusting the observed wind direction relative to the satellite track direction +90

SAR Data Quality
Various factors can influence the quality of the data used for algorithm development and testing.Spatio-temporal issues concern the temporal, and thus spatial, difference between buoy data at its reporting time and at the SAR acquisition time.These issues were mitigated somewhat by the use of 3 × 3 km ROIs.Two additional factors were spatial variability within sample ROIs, and low-quality backscatter in the proximity of the noise floor.

Spatial Variability
ROIs that exhibit high variability in backscatter are likely representative of image areas (i.e., 3 × 3 km) containing discontinuous wind slicks or strong wind gradients.Such ROI sample values are probably unrepresentative of the buoy wind speed and/or direction.Therefore, they are likely to introduce a significant source of error and should be removed from further analysis.
An analysis of σ 0 VV standard deviation values shows that a number of samples exhibited relatively high variability compared to the bulk of the samples (Figure 2).An upper threshold for the acceptable spatial variability was set at two standard deviations above the mean: where ̄ is the mean of the σ 0 VV standard deviation values and s is the standard deviation of the standard deviation values.A lower threshold is not needed, as low variability is desired.For example, for the LowNoise mode, the mean of the σ 0 VV standard deviation values was 1.49 dB, and the standard deviation of the standard deviation values was 0.26 dB; thus  1.49 2 ⋅ 0.26 2.01 dB.Samples ≥ τ were omitted: 16 samples for LowNoise (τ = 2.01 dB), 14 for LowRes (τ = 2.04 dB), and 14 for MR50 (τ = 2.17 dB).

Low-Quality Backscatter
Low-quality backscatter exists in situations of low wind speed over water [20].This results from the presence of wind slicks, either at the pixel level or as a portion of a spatially-averaged region of interest.Wind slicks occur when wind forces cannot overcome viscous forces, resulting in specular reflecting surfaces that do not exhibit the commonly-observed Bragg-scattering from water surfaces [24,25].At C-band, this usually occurs at wind speeds <~3 m/s (Figure 3).Low-quality backscatter is caused by the contamination of the radar signal by antenna side lobes and by returns from nearby pixels [20].
Low-quality backscatter negatively affects the phase information of the radar return.This influences any algorithm development that includes such data, often causing non-linear relationships.Low-quality data should, therefore, be treated separately from high-quality data.

SAR Data Quality
Various factors can influence the quality of the data used for algorithm development and testing.Spatio-temporal issues concern the temporal, and thus spatial, difference between buoy data at its reporting time and at the SAR acquisition time.These issues were mitigated somewhat by the use of 3 × 3 km ROIs.Two additional factors were spatial variability within sample ROIs, and low-quality backscatter in the proximity of the noise floor.

Spatial Variability
ROIs that exhibit high variability in backscatter are likely representative of image areas (i.e., 3 × 3 km) containing discontinuous wind slicks or strong wind gradients.Such ROI sample values are probably unrepresentative of the buoy wind speed and/or direction.Therefore, they are likely to introduce a significant source of error and should be removed from further analysis.
An analysis of σ 0 VV standard deviation values shows that a number of samples exhibited relatively high variability compared to the bulk of the samples (Figure 2).An upper threshold for the acceptable spatial variability was set at two standard deviations above the mean: where x is the mean of the σ 0 VV standard deviation values and s is the standard deviation of the standard deviation values.A lower threshold is not needed, as low variability is desired.For example, for the LowNoise mode, the mean of the σ 0 VV standard deviation values was 1.49 dB, and the standard deviation of the standard deviation values was 0.26 dB; thus τ = 1.49+ 2 • 0.26 = 2.01 dB.Samples ≥ τ were omitted: 16 samples for LowNoise (τ = 2.01 dB), 14 for LowRes (τ = 2.04 dB), and 14 for MR50 (τ = 2.17 dB).Low-quality backscatter can be identified using Conformity, ρRVRH, and/or δRVRH [19,20].An analysis of the data shows that δRVRH and Conformity are closely related (Figure 4).A Conformity threshold ≤0 clearly identifies samples with highly divergent phase information, i.e., significant departures from the −90° value expected for open water.However, samples with conformity values as high as ~0.25 also appear to be associated with divergent phase values; these may also affect algorithm development.
Breakpoint analysis was used to identify the Conformity value at which data quality diverges.This was carried out using segmented regression, using all the samples remaining after the spatial variability constraint.The respective breakpoints were at Conformity values of 0.176 (LowNoise), 0.212 (LowRes) and 0.211 (MR50).The higher noise floor of the LowNoise mode resulted in fewer samples exhibiting low data quality, as expected.To ensure high quality data, a Conformity threshold of >0.2 (mean for three modes) was used.This is supported by the analysis of [20].The Conformity threshold removed 19 samples from the LowNoise mose, 32 from the LowRes mode, and 31 from the MR50 mode.

Low-Quality Backscatter
Low-quality backscatter exists in situations of low wind speed over water [20].This results from the presence of wind slicks, either at the pixel level or as a portion of a spatially-averaged region of interest.Wind slicks occur when wind forces cannot overcome viscous forces, resulting in specular reflecting surfaces that do not exhibit the commonly-observed Bragg-scattering from water surfaces [24,25].At C-band, this usually occurs at wind speeds <~3 m/s (Figure 3).Low-quality backscatter is caused by the contamination of the radar signal by antenna side lobes and by returns from nearby pixels [20].Low-quality backscatter can be identified using Conformity, ρRVRH, and/or δRVRH [19,20].An analysis of the data shows that δRVRH and Conformity are closely related (Figure 4).A Conformity threshold ≤0 clearly identifies samples with highly divergent phase information, i.e., significant departures from the −90° value expected for open water.However, samples with conformity values as high as ~0.25 also appear to be associated with divergent phase values; these may also affect algorithm development.
Breakpoint analysis was used to identify the Conformity value at which data quality diverges.This was carried out using segmented regression, using all the samples remaining after the spatial variability constraint.The respective breakpoints were at Conformity values of 0.176 (LowNoise), 0.212 (LowRes) and 0.211 (MR50).The higher noise floor of the LowNoise mode resulted in fewer samples exhibiting low data quality, as expected.To ensure high quality data, a Conformity threshold of >0.2 (mean for three modes) was used.This is supported by the analysis of [20].The Conformity threshold removed 19 samples from the LowNoise mose, 32 from the LowRes mode, and 31 from the MR50 mode.Low-quality backscatter negatively affects the phase information of the radar return.This influences any algorithm development that includes such data, often causing non-linear relationships.Low-quality data should, therefore, be treated separately from high-quality data.
Low-quality backscatter can be identified using Conformity, ρ RVRH , and/or δ RVRH [19,20].An analysis of the data shows that δ RVRH and Conformity are closely related (Figure 4).A Conformity threshold ≤0 clearly identifies samples with highly divergent phase information, i.e., significant departures from the −90 • value expected for open water.However, samples with conformity values as high as ~0.25 also appear to be associated with divergent phase values; these may also affect algorithm development.

Final Data Set
There was a very limited number of high-wind speed samples: only three samples (in each mode) were >17.8 m/s, and these did not have a sufficient incidence angle distribution.Therefore, these samples were removed, and analysis and model development were limited to wind speeds ≤18 m/s.
The final data set contained between 435 and 446 samples, depending on RCM mode (Table 3).Stratified random selection during sampling was used to ensure that the desired incidence angle ranges and wind speed ranges were sufficiently represented.The justification for the incidence angle ranges and wind speed ranges is outlined in [14].
Noise reduction was used in order to compare observed backscatter with CMOD5n-modelled values because CMOD5n was developed with noise subtracted.Both the RCM simulator mode nominal noise floor and the original Rdarasat-2 FQ noise floor were subtracted from the observed backscatter.A further 3 dB was subtracted to account for the noise floor pattern of FQ images, which Breakpoint analysis was used to identify the Conformity value at which data quality diverges.This was carried out using segmented regression, using all the samples remaining after the spatial variability constraint.The respective breakpoints were at Conformity values of 0.176 (LowNoise), 0.212 (LowRes) and 0.211 (MR50).The higher noise floor of the LowNoise mode resulted in fewer samples exhibiting low data quality, as expected.To ensure high quality data, a Conformity threshold of >0.2 (mean for three modes) was used.This is supported by the analysis of [20].The Conformity threshold removed 19 samples from the LowNoise mose, 32 from the LowRes mode, and 31 from the MR50 mode.

Final Data Set
There was a very limited number of high-wind speed samples: only three samples (in each mode) were >17.8 m/s, and these did not have a sufficient incidence angle distribution.Therefore, these samples were removed, and analysis and model development were limited to wind speeds ≤18 m/s.
The final data set contained between 435 and 446 samples, depending on RCM mode (Table 3).Stratified random selection during sampling was used to ensure that the desired incidence angle ranges and wind speed ranges were sufficiently represented.The justification for the incidence angle ranges and wind speed ranges is outlined in [14].
Remote Sens. 2019, 11, 1682 7 of 15 Noise reduction was used in order to compare observed backscatter with CMOD5n-modelled values because CMOD5n was developed with noise subtracted.Both the RCM simulator mode nominal noise floor and the original Rdarasat-2 FQ noise floor were subtracted from the observed backscatter.A further 3 dB was subtracted to account for the noise floor pattern of FQ images, which was lower towards the centre of an image [14].No RCM simulator or RADARSAT-2 FQ noise reduction was used for CMOD-IFR2, as it appeared that no noise subtraction was used during its development; only the 3 dB noise floor pattern value was subtracted.

Wind Speed Retrieval
Wind speed retrieval was performed by inverting the backscatter models.This was accomplished by beginning at two extreme wind speeds (low and high), then incrementing or decrementing the wind speed (by 0.01 m/s) until an observed parameter value was reached.Both the incrementing and decrementing methods must converge at a similar value (within 1 m/s) for a retrieval to occur.The difference between the two methods is usually ≤0.03 m/s.The resulting retrieval is the average of the incrementing and decrementing results.
If the difference between the two methods is >1 m/s, this indicates that the model does not exhibit a monotonic relationship with wind speed.This can result in the retrieval of two significantly different wind speeds, and thus in a lack of convergence.No retrieval occurs in such cases.
Wind speed retrievals < 0 m/s or > 18 m/s occasionally occurred, even though the data set was restricted to values ≤ 18 m/s.This was due the incomplete statistical representation of the model and the stochastic nature of the data.We omitted retrievals < 0 m/s and > 20 m/s.

Wind Speed Accuracy Assessment
Wind speed accuracy assessment was done by comparing the parameter-modelled wind speed with the buoy-measured wind speed.This was assessed using statistical measures: Spearman's correlation, Root-Mean-Square Error (RMSE), slope (and intercept), and overall bias.

Wind Speed Retrieval
CMOD-related parameters have dependencies with incidence angle, wind speed, and wind direction.Therefore, the use of these models for wind speed retrieval necessitates a priori knowledge of wind direction.Buoy measurements are the source of wind direction for the wind speed retrieval tests.
The model comparison focused on the RMSE and slope statistics, because together, they described most of the models' efficacy (Table 5).When averaging all the parameters, the lowest RMSE values occurred in the MR50 mode, and the best slopes were also in the MR50 mode.When averaging each parameter across the three modes, CMOD5n-SV 0 had the lowest RMSE, followed by CMOD5n-σ 0 VV and CMOD5n-σ 0 RV (Figure 6).CMOD5n-σ 0 RH also seemed to perform relatively well at high wind speeds; however, its overall variability was greater.

RCM SAR Wind Case Studies
Three image-based case studies illustrate the applications of RCM CP and linear wind retrieval models.Case 1 illustrates retrievals within an FQ image for a small incidence angle (Figure 7).The small extent of the FQ images (~25 × 25 km) limits wind speeds to a narrow range of values.Therefore, the colour scale is relative to the range of the wind speeds retrieved for the image.Wind direction input to the model comes from the buoy data located within the image (buoy 44141).The wind speed retrievals exhibited a general overestimation at low wind speeds and underestimation at high wind speeds (Figure 6), leading to slope values <1 (Table 5).For the CP parameters, this was most prevalent for σ 0 RV and least prevalent for σ 0 RH , with σ 0 RL , exhibiting a compromise between the two.Overall, CMOD5n-SV 0 had the lowest RMSE values in the MR50 and LowRes modes and was only bested by CMOD5n-σ 0 VV in the LowNoise mode.The skill of CMOD5n-SV 0 was likely the result of its cross-polarized component, which compensated somewhat for the wind speed underestimation of the co-polarized components at high wind speeds.The cross-polarized contribution can likely be improved by using a more nuanced model than the Vachon and Wolfe [5] model used in this study.This may make the CMOD5n-SV 0 retrieval even better.The similarity in slope values between CMOD-IFR2 and CMOD5n was the result of appropriate handling of the noise values: minimal noise reduction in the case of CMOD-IFR2 and significant noise reduction in the case of CMOD5n.Once RCM mode noise values are known, following launch and commissioning, mode-specific noise reduction analysis will be needed to obtain robust retrievals.
All model results were within the accuracy range reported for SAR wind retrieval: between 1.5 and 2.7 m/s [5,6,[27][28][29].Although CMOD5n models usually had better RMSE values than CMOD-IFR2 models, the mean difference across all parameters and modes was only 0.1 m/s.These results provide additional evidence that CMOD-IFR2-and CMOD5n-modelled CP parameters can be used with confidence, instead of CMOD-σ 0 VV models, when only CP data are available.Although the accuracies reported in this study were adequate, they did not achieve RMSE <2 m/s.A number of factors that may have caused the somewhat reduced accuracy, including (1) the temporal mismatch between SAR acquisition and buoy wind speed measurement, (2) the buoy wind direction measurements may not have always been representative of the 3 × 3 km sample area, (3) the addition of noise by the simulator, and subsequent noise reduction, may have added error, and (4) the use of a relatively simple retrieval technique.Higher accuracy can likely be achieved with more sophisticated retrieval schemes; however, the relative accuracy between the linear and CP parameters was of interest in this study.Once actual RCM data are available, greater effort will be devoted to increasing the retrieval accuracy.
The wind speed retrieval at the buoy location (buoy 44141) in Case 1 (Figure 7) was underestimated (11.6 m/s versus a buoy measurement of 15.0 m/s).However, the wind speed gradient in the image was quite strong and the 12-minute temporal mismatch between the buoy measurements and the SAR acquisition can account for this discrepancy.The greater prevalence of higher wind speed retrievals along the near-range edge (left side), and to a lesser degree, the far-range edge, may be representative of the actual wind pattern.However, this may also be due to the noise floor pattern of RADARSAT-2 FQ scenes, which was higher at the near-and far-range edges.Furthermore, on the near-range side, there may also have been remnant filter artifact effects.Further research is needed to isolate the actual cause(s).However, the retrieval in Figure 7 is solely illustrative of a retrieval within the small areal extent (25 × 25 km) of a RADARSAT-2 FQ image, and does not reflect the scale of the operational systems, which are not constrained by such small swath widths.The operational swath width images shown in Figure 8 illustrate retrievals with the RCM MR50 mode (350 km wide); these are able to resolve complex wind fields and provide retrievals in convoluted coastlines over large geographic extents.Nevertheless, noise effects are likely to be a limiting factor and must be carefully considered once actual RCM data become available.

Conclusions
In this study, a set of 504 RADARSAT-2 FQ images was used to simulate RCM image modes, in order to sample CP parameters over meteorological buoy locations on the West and East coasts of Canada.Three RCM modes were simulated: Low Noise, Low Resolution 100 m, and Medium Resolution 50 m.These samples were used to evaluate the efficacy of CMOD-related CP (CMOD-CP) parameters (σ 0 RV , σ 0 RH , σ 0 RL and SV 0 ) for wind speed retrieval, using CMOD-IFR2 and CMOD5n.CMOD-CP accuracy was compared to CMOD results for σ 0 VV and σ 0 HH .

Figure 1 .
Figure 1.ECCC buoy locations, with their ID code, on the West and East coasts of Canada.

Figure 1 .
Figure 1.ECCC buoy locations, with their ID code, on the West and East coasts of Canada.

15 Figure 7 .
Figure 7. Wind speed retrieval with CMOD-IFR2-σ 0 RV for the RCM LowRes mode simulated from a RADARSAT-2 FQ3 image for 23 October 2010.The black point is the location of buoy 44141.The projection is Lambert Conformal Conic, Canada WGS84.

Figure 7 .
Figure 7. Wind speed retrieval with CMOD-IFR2-σ 0 RV for the RCM LowRes mode simulated from a RADARSAT-2 FQ3 image for 23 October 2010.The black point is the location of buoy 44141.The projection is Lambert Conformal Conic, Canada WGS84.

Table 1 .
Specification of RCM modes used for this study; all modes are capable of providing HH, VV, HH+HV, VV+VH and CP polarizations.LowRes and MR50 are also capable of providing HH+VV [21].

Table 2 .
Compact-polarimetry and linear polarization parameters used in this study.

Table 5 .
Wind retrieval accuracies for the C-band models (CMOD).The green cells indicate the best three values for RMSE and slope in each RCM mode.