Previous studies have demonstrated that satellite synthetic aperture radar (SAR) can be used as an accurate scatterometer, yielding wind speed fields with subkilometer resolution. This wind speed generation is only possible, however, if a corresponding accurate wind direction field is available. The potential sources of this wind direction information include satellite scatterometers, numerical weather prediction models, and SAR itself through analysis of the spatial patterns caused by boundary layer wind structures. Each of these wind direction sources has shortcomings that can lead to wind speed errors in the SAR-derived field. Manual and semiautomated methods are presented for identifying and correcting numerical weather prediction model wind direction errors. The utility of this approach is demonstrated for a set of cases in which the first-guess wind direction data did not adequately portray the features seen in the SAR imagery. These situations include poorly resolved mesoscale phenomena and misplaced synoptic-scale fronts and cyclones.
Marine meteorologists increasingly depend on satellite remote sensing to satisfy their need for over-ocean wind vector data. The traditional means by which the near-surface wind vector is remotely sensed is a satellite scatterometer. For example, data from SeaWinds scatterometer aboard the Quick Scatterometer (QuikSCAT) satellite are readily available from numerous online government outlets (e.g., http://manati.orbit.nesdis.noaa.gov/quikscat/ and http://www.ndbc.noaa.gov).
Scatterometry capitalizes on the Bragg scattering relationship between microwave normalized radar cross section (NRCS) of the ocean surface and the wind vector (e.g., Stoffelen and Anderson 1997). The backscattering of microwave radiation at moderate incident angles is primarily accomplished by roughness features having microwave length scales. Thus, NRCS from a C-band (5.6 cm) scatterometer, such as that aboard the European Remote Sensing-2 (ERS-2) satellite, is keenly sensitive to centimeter-scale roughness. Over the ocean, that centimeter-scale roughness is primarily wind generated.
In general terms, microwave NRCS of the ocean surface increases with near-surface wind speed. Wind direction also affects microwave NRCS of the ocean surface. Because centimeter-scale wind-generated waves tend to travel with the wind, NRCS is maximized when the wind direction is opposite to the radar look direction. There is smaller local maximum in NRCS when the wind direction is the same as the radar look direction. The minimum in NRCS occurs when the radar look direction is perpendicular to the wind direction.
Armed with this knowledge, empirical geophysical model functions (GMFs) have been developed for relating microwave NRCS of the ocean surface to wind speed. An example of such is the C-band Geophysical Model Function (CMOD4; Stoffelen and Anderson 1997). Equation (1) is the general form of the CMOD4 GMF:
Here, σ0 is NRCS, U is wind speed, ϕ is the relative angle between the wind direction and the radar look direction, θ is the local radar incident angle, and A, B, C, and γ are parameters that depend on incident angle and wind speed. Figure 1 shows a corresponding relationship between NRCS, wind speed, and ϕ for a 25° incident angle. As can be seen from Fig. 1, the inversion from NRCS to wind speed is not unique because a single value of NRCS corresponds to more than one wind vector. Scatterometers reduce this uncertainty by sensing a given area of ocean surface with multiple antennae. However, the liability of this approach is on the order of 10-km resolution.
Several groups of researchers have demonstrated that satellite synthetic aperture radar (SAR) can be used as an accurate (standard deviation on the order of 1 m s−1 between SAR and scatterometers) high-resolution (order of 0.1–1 km) scatterometer (e.g., Horstmann et al. 2003; Monaldo et al. 2004a). Thus, SAR can be used to produce wind speed fields at a resolution at least one order of magnitude higher than scatterometers. A consequence of this advantage in resolution is that SAR-derived wind speed fields can reveal the signatures of mesoscale and microscale meteorological phenomena that cannot be resolved by scatterometers. See Jackson and Apel (2004) for a review of the various marine uses of SAR data. Beal et al. (2005) and Sikora et al. (2006a) provide reviews of the marine meteorological uses of SAR data.
As described in Monaldo et al. (2004b), a major impedance to the generation of SAR-derived wind speed data is the requirement for a priori knowledge of the wind direction. The problem arises because, unlike scatterometers, SAR has only one antenna. Several wind direction estimation techniques have been applied in order to overcome this limitation. The sources of wind direction most commonly used are the following: scatterometer wind directions (e.g., Monaldo et al. 2004a), linear features within the SAR image that are assumed to be aligned with the wind direction (e.g., Horstmann et al. 2000), and numerical weather prediction (NWP) model wind directions (e.g., Monaldo et al. 2001). Of the three techniques, only the linear feature and NWP model approaches are being applied in a quasi-operational environment. That quasi-operational environment is the Alaska SAR demonstration (Monaldo 2000; see online at http://www.orbit.nesdis.noaa.gov/sod/mecb/sar/).
Each of the wind direction estimation techniques mentioned above has its related problems (Sikora et al. 2006b). In this paper we investigate the impact of first-guess wind direction errors from NWP-modeled synoptic-scale cyclones and fronts on the wind direction–dependent retrieval of wind speed from SAR. We present manual and semiautomated editing and morphing correction algorithms for incorrect first-guess wind direction fields associated with cyclones and frontal wind shift lines. These algorithms will improve the accuracy of SAR-derived wind speed retrieval in conjunction with NWP model wind direction fields. Although we focus on improving upon NWP model wind directions, the methods discussed below are applicable to all gridded wind direction data, including data from scatterometers.
Because wind speed retrieval from SAR backscatter fields is sensitive to the first-guess wind direction field for typical incident angles and wind speeds, it is almost always essential to correct the first-guess wind direction field for any gross errors. When the first-guess wind direction field comes from a NWP model, these errors can include misplaced, missing, or bogus features such as fronts and cyclones (Young et al. 2005). Likewise, when the NWP model lacks the resolution to fully capture any sharp wind shift across fronts, wind direction errors can become large within the frontal zone.
In this section we describe three methods for correcting these types of gross errors in the first-guess gridded wind direction field to obtain a revised-guess wind direction field suitable for use in SAR-based wind speed retrieval. All examples found below are for correcting NWP model wind direction fields. The choice of which method to use depends primarily upon the information available in the first-guess wind direction field. If important features are either missing or erroneously included in the first-guess wind direction field, then manual editing is required. In contrast, if the first-guess wind direction field contains all of the features observed in the SAR-derived wind speed field, then manually guided or semiautomated morphing (i.e., spatial warping) will suffice to correctly position these features. Each of these methods will be discussed in its own section below.
The SAR-derived wind speed data presented below were produced from ScanSAR Wide data (processed at the Alaska SAR facility) from the SAR on board the Canadian Space Agency’s RADARSAT-1. The RADARSAT-1 SAR is C band and right looking with horizontal–horizontal polarization. Our SAR-based wind speed fields are produced using the CMOD4 GMF modified for horizontal–horizontal polarization using a polarization parameter of 0.6. All first-guess wind directions are from forecasts of the U.S. Navy’s Operational Global Atmospheric Prediction System (NOGAPS) NWP model closest in time to the corresponding SAR overpass. Because the CMOD4 GMF underestimates the true wind speed at large values of true wind speed (Donnelly et al. 1999), we have chosen to saturate our SAR-derived wind speeds at 25.0 m s−1. The above methodology closely follows that presented in Monaldo et al. (2004a). All images of SAR-derived wind speed fields provided herein are oriented so that north is directed toward the top of the page.
Because of the limitations of CMOD4, several groups of researchers are investigating the soundness of newer GMFs, such as “CMOD5” (Hersbach 2003). The soundness of newer GMFs is still an area of active research (e.g., Horstmann et al. 2005). Because we are more completely aware of the physical limitations of CMOD4 versus the newer GMFs, we choose to present our wind direction editors based on CMOD4.
a. Manual editing of gridded wind fields
Manual editing of the first-guess wind direction field is potentially time consuming, so efficiency is a chief concern, particularly in time-limited operational forecasting situations. Thus, a wind direction editor for use in SAR-derived wind speed generation should make it easy for the analyst to accept or reject arbitrary sections of the gridded first-guess wind direction field. It should also allow the analyst to add wind direction data points based on the SAR signatures of mesoscale and submesoscale flow features visible in the backscatter field (e.g., Sikora et al. 2006b). Because previously obscured wind direction indicators may become visible as the first-guess wind direction field is corrected and a corresponding revised wind speed field is generated, the editing process can be iterative. Generally, convergence happens in half a dozen or less edit–retrieval cycles.
The editor implemented for this study is shown in Fig. 2. It has a graphical user interface based loosely on the radar editor described in Young et al. (1997), and is implemented in the Matlab computer language (see online at http://www.mathworks.com/). The central feature of the editor is a display of the wind speed field retrieved from the SAR backscatter field. This wind speed field can be updated at the push of a button as the analyst edits the input wind direction field. Because the wind speed retrieval is computationally costly when undertaken for a megapixel image, the analyst usually makes all of the wind direction edits suggested by the current wind speed field before pushing the recompute button.
The analyst starts each edit cycle by using the radio buttons at the left of the window in Fig. 2 to specify which wind direction source to edit: scatterometer observations, NWP model data, or the analyst’s subjectively deduced wind direction estimates. The default is to create or delete subjective wind direction data points based on analyst-observed features of the SAR-derived wind speed field. Other options include turning on or off the influence of data points from scatterometers and NWP model wind direction fields. A slider control allows the analyst to adjust the brush size so that large or small areas can be affected with a single mouse click as needed. Typically, large areas of erroneous first-guess winds are turned off and then single data points are flipped on or off to fine-tune the analysis. The active data type is shown in color and the inactive types in grayscale. For the active data type, the points selected by the analyst for use are shown in an intense hue and those deselected are shown in pastel. Thus, the analyst can readily see and rapidly control the wind direction data going into the wind speed retrieval.
When the analyst triggers a wind speed retrieval calculation, the selected wind direction data points from all sources are combined into a pixel-level field via a hybrid objective analysis scheme featuring bicubic interpolation within the field of wind direction points and nearest-neighbor extrapolation outside the field. This analysis and the CMOD4 retrieval dominate the time required for a recomputation of the wind speed field. Once the analyst is satisfied with the results of the last edit–recomputation cycle, the editor state and the final wind speed and direction fields can be saved at the click of a button.
b. Manual morphing of gridded wind fields
As an alternative to manual editing, the gridded first-guess wind direction field can be spatially transformed in order to correct position errors. This approach can work only when the wind direction pattern is essentially correct in the first-guess field with all features present, but not necessarily in their correct positions. The goal of morphing is then to move each wind direction data point in the first-guess field to the geographic location for which it is actually valid. For example, a surface cyclone might be shifted 100 km to correct for a NWP model position error (e.g., Hollingsworth et al. 1986; Hamill and Snyder 2002). The difference in pass times between scatterometers and SARs could result in a similar displacement. The morphing software would move the input wind direction data near the cyclone center to the appropriate location, leave untouched those distant input wind direction data that were already in their correct locations, and move proportionately those input wind direction data from intervening points. This approach is termed morphing because it uses essentially identical mathematics to that used to distort images in graphics arts software.
Manual morphing of a first-guess wind direction field involves three stages. First, the analyst must recognize the SAR-derived wind speed signatures of misplaced features of the gridded wind direction field: fronts, and singularities such as cyclones, anticyclones, and col points. Second, the analyst must specify a set of paired (before and after transformation) control points. For each pair, wind direction data from the before point will be relocated to the after point, changing its geographic location. Third, the software must calculate the spatial displacement, not just for these control points, but for all of the data points in the first-guess wind direction field.
The first stage, feature position error recognition, is simplified by the existence of a small set of recurrent signatures resulting from the interaction of singularities or near discontinuities in the wind direction field with the response function in the SAR-based wind speed retrieval (e.g., Fig. 3, which is a barrier jet case; Sikora et al. 2006b; Loescher et al. 2006; Colle et al. 2006).
Misplacement of a wind direction singularity such as a cyclone, anticyclone, or col point results in the true wind crossing the first-guess wind at right angles in a band on either side of the false position. If the erroneous wind directions within this band are aligned perpendicular to the SAR look direction and the band is located at typical incident angles (30°–49°), it superposes the maxima in the CMOD4 backscatter-to-wind function on locations that should be using the minima such that the SAR-derived wind speed will be greater than the actual wind speed. Inversely, if the erroneous wind directions with this band are aligned parallel to the SAR look direction, it superposes minima in the CMOD4 backscatter-to-wind function such that the SAR-derived wind speed will be greater than the actual wind speed. The result is a signature featuring four wedges radiating out from the first-guess singularity position, with the pair aligned most nearly parallel to the SAR look direction having markedly lower SAR-derived wind speed than the pair aligned most nearly perpendicular to the SAR look direction. Depending on the true wind direction, one or both of the wedge pairs will stand out from the surrounding SAR-derived wind speed field, giving rise to an hourglass or double hourglass signature as seen in Fig. 4a.
Misplaced frontal wind direction shifts can result in equally obvious signatures in the SAR-derived wind speed field. These signatures are again the result of the first-guess wind direction lying on a different part of the backscatter response function than the actual wind direction. If the actual wind direction is more nearly parallel to the SAR look direction than is the first-guess wind direction, the SAR-derived wind speed will be greater than the actual wind speed. In contrast, if the actual wind direction is more nearly perpendicular to the look direction than is the first-guess wind direction, the SAR-derived wind speed will be less than the actual wind speed. These situations will arise from the misplacement of a front for which the wind direction on one side of the front is in much closer alignment with the radar look direction than that on the other. The result can either be a band of erroneously fast winds, as in the first situation discussed above, or a band of erroneously slow winds, as in the second situation discussed above. Both signatures can occur with any frontal orientation depending on the degree of wind shift across the front and the direction of the frontal position error. Examples are shown in Fig. 5a (a fast band) and Fig. 6a (a slow band). These same signatures can arise when the coarse resolution of the first-guess wind direction field results in the spreading of a narrow frontal wind shift zone across an erroneously broad band of pixels.
The cure for misplaced singularities and frontal wind shift zones is to spatially displace these features to their true location, as deduced from feature signatures in the SAR backscatter field and the error signatures in the SAR-derived wind speed field. For a singularity, this morphing requires one control point pair to displace the singularity and at least three surrounding control points to define the radius of influence (i.e., the spatial region beyond which the wind direction data will not be displaced). For a frontal wind shift zone, at least two pairs of control points are required to displace a frontal zone and at least four pairs to change its width. If the front is curved, more pairs will be required. For each of these feature position errors, the chief challenge is determining where the feature is actually located so that the first-guess wind direction field can be morphed accordingly.
The approach for position errors in wind direction singularities is to morph the point at the center of the hourglass signature to the true position of the singularity. In the case of cyclones and anticyclones, the true position is often visible as the center of circulation defined by curving features in the backscatter field or by a nearly circular area of low backscatter (Young et al. 2005). In the case of col points, the true position often lies along a front. If the true position is not obvious, we recommend displacements to various locations in different directions no further from the original location than the radius of the hourglass signature. Repeat the process starting from the new position with the smallest hourglass signature. Because the fast and slow sectors of the hourglass signature are oriented primarily with respect to the radar look direction, their alignment provides little guidance as to which way to move the center point. However, if one of the two sector types (fast or slow) seems to fit in better with the surrounding wind field, this sector likely has the wind directions correctly defined. Therefore, the true position probably lies more nearly along that axis than the other. Fixed control point pairs should be placed at the periphery of the circulation being displaced, to prevent the surrounding weather features from being distorted.
The approach for position and width errors in frontal zones is similar, but more control point pairs are required because the analyst is working to achieve a one-dimensional strip of maximum displacement (i.e., the frontal position shift) rather than just one point. Despite this, correcting frontal position errors is often easier because the strong winds and rapid change in wind direction across a front typically leads to a sharp discontinuity in SAR backscatter at the actual frontal position (Young et al. 2005) as seen in Figs. 5 and 6. The analyst need only locate this feature on the SAR backscatter image and specify control point pairs on each side of it to reposition and concentrate the wind shift zone along it. The only real challenge lies in coping with the geometric complexity when analyzing the winds near sharply curving or winding fronts.
The analyst-supplied displacement information in the control point pairs is interpolated to the location of each first-guess wind direction point using objective analysis. For the results shown in section 3, the Matlab “griddata” function is used, implementing triangle-based cubic interpolation. Evaluation on analytic test data showed that cubic interpolation did a much better job of preserving vorticity and divergence than did the alternatives: linear or nearest-neighbor interpolation.
c. Semiautomated morphing of gridded wind fields
For the case of a misplaced singularity in the first-guess wind direction field, there is an alternative to the analyst manually determining the feature’s control point displacement. Because a singularity is a point, only one control point need be displaced. Thus, the problem has many fewer degrees of freedom than correcting the position of a front, and is therefore much more amenable to artificial intelligence solutions. Indeed, one of the simplest automated optimization techniques, the Nelder–Mead simplex method [e.g., Lagarias et al. (1998), as implemented in Matlab’s fminsearch function], proves sufficient. The analyst need only provide the algorithm with the singularity’s position in the first-guess wind direction field, and the algorithm then iterates to a solution that minimizes the hourglass error signature.
The Nelder–Mead simplex method works by iteratively moving the trial solution through parameter space, moving down the estimated gradient of an error score. Thus, the method requires a function for evaluating the severity of the position error. Given the hourglass shape of the error signature of a misplaced wind direction singularity, there are two maxima and two minima in wind speed around each ring of pixels centered on the singularity. Thus, the error signature can be approximated by the function
where ϕradial is the angle from the singularity position to the pixel in question and ϕorientation is the alignment of one of the two fast lobes of the signature. The signature’s error score is computed by finding the best-fit orientation for this function at each range ring from 5 to 50 pixels out from the singularity and summing the resulting correlation coefficients. The maximum reflects the typical lobe extent and the minimum the ∼30-pixel circumference required to fit well a double-cosine function. Thus, a well-defined hourglass signature extending 50 pixels out from the singularity receives the maximum error score, while a smaller signature receives a lower error score. This difference allows the Nelder–Mead method to adjust the singularity’s position so as to minimize the size of the error signature, often eliminating it entirely.
The primary advantage of this semiautomated technique is that it requires attention from the analyst only at its initiation and upon completion. In contrast, the manual morphing described above requires the analyst to make all of the decisions. Thus, there is considerable savings in analyst time. The disadvantage of the semiautomated technique is that it typically requires about 5 times as many iterations to arrive at a solution, because it simply works down the spatial gradient of error signature size instead of looking at the image as a whole to anticipate where the singularity should be moved. Because the SAR-based wind retrieval is computationally expensive, this may take as much or more wall clock time as the fully manual technique. Faster computers would favor the semiautomated method, but larger, more detailed images would favor the manual method. Thus, the choice between the two methods will depend on the relative rate of improvement in SAR imagery and the computational resources to analyze them.
Sample image analyses will be used to demonstrate each of the methods described above. The cases shown were selected both to illustrate the potential of the methods and to document the types of situations in which they may fail.
a. Manual editing of gridded wind fields
The improvement in SAR-derived wind speeds that can be achieved with the manual wind direction editor is illustrated using SAR data collected at 0310 UTC 18 February 2000 over the Gulf of Alaska (Fig. 3a). This case features a barrier jet flowing cyclonically along the Alaskan coast (Loescher et al. 2006). Thus, there is a mesoscale boundary running roughly east–west across the center of the image separating the high-speed cyclonic mesoscale flow along the coast from the slower synoptic-scale onshore flow farther out to sea.
Figure 3a shows the SAR-derived wind speed field derived using the raw NWP model wind direction field. Because the NWP model did not fully resolve the coastal barrier jet seen in the SAR data, the first-guess wind direction within the western and center portion of the jet was more nearly across the track than in reality, resulting in an overestimate of the strength of that portion of the barrier jet. Correction of the wind directions to align with the island wakes and other linear features seen in the SAR-derived wind speed field resulted in a marked decrease in the SAR-derived wind speed within the barrier jet (Fig. 3b). Thus, manually editing the first-guess wind direction wind field makes an operationally significant difference in the SAR-derived wind speed field for this case. Success with this tool depends on the analyst being able to use features observed in the SAR image or other information to correct the first-guess wind direction field.
b. Manual morphing of gridded wind fields
Manual morphing of a gridded first-guess wind direction field is most useful when the errors to be corrected involve misplacement of geometrically simple features such as fronts, cyclones, and anticyclones. Use of this method to correct wind feature position errors is demonstrated with an along-look-direction front (Fig. 5) and across-look-direction front (Fig. 6). For cyclones, the manual results are identical to those for the semiautomated morphing system and so will be shown in the next section.
The SAR-derived wind speed field at 1957 UTC 8 January 2004 shows an across-track front in the Sea of Okhotsk (Fig. 5), with the frontal position being identified as in Young et al. (2005). The true front in this case is actually located to the north of that depicted in the NWP model wind direction field. Thus, the wind directions immediately south of the true front are actually more nearly across track (in this case, alongfront from northwest to southeast) than the northeasterly winds depicted by the NWP model. As a result, the SAR-derived wind speed is too high in a band along the south side of the front (Fig. 5a). Morphing the NWP model frontal position to match that of the wind speed discontinuity in the SAR image eliminates this feature, resulting in a more uniform wind speed field to the south of the front (Fig. 5b).
The SAR-derived wind speed field at 1545 UTC 20 December 2001 shows an along-track front in the Gulf of Alaska (Fig. 6) with westerly winds to its west and southerly winds to its east. Because the actual frontal wind shift is more sharply defined (i.e., closer to a zero-order discontinuity) than can be depicted on the NWP model grid, the wind directions immediately west of the front are more along track than in reality while the wind directions immediately east of the front are less along track than in reality. This results in the SAR-derived wind speeds being erroneously fast immediately to the west of the front and erroneously slow immediately to the east of the front (Fig. 6a). Manual morphing allows for the sharpening of the frontal wind direction shift and reduction of these errors (Fig. 6b).
In contrast, the manual morphing procedure cannot be used when the first-guess wind direction field either has a feature that does not actually occur anywhere in the SAR image or if it is missing such a feature (e.g., the barrier jet case shown in Fig. 3). The SAR image at 0308 UTC 16 March 2004 (Fig. 7) shows a similar situation. Here, the SAR indicates a front extending roughly east–west across the southern third of the image while the NWP model wind direction field lacks such a feature. Thus, no combination of repositioning or stretching (i.e., morphing operations) applied to the NWP model wind direction field can correct the banded pattern of wind speed errors seen in this image. Instead, manual editing is required.
c. Semiautomated morphing of gridded wind fields
Morphing to reposition and stretch portions of a gridded first-guess wind direction field can be accomplished automatically for cyclones, anticyclones, and col points because they present an extremely simple geometry with a single morph point connecting the location of the wind direction singularity in the first guess and in reality. Figure 4 shows the SAR-derived wind speed fields before and after semiautomated morphing of the misplaced cyclone observed over the Bering Sea at 1818 UTC 6 December 2000. The Nelder–Mead simplex algorithm described in section 2c required less than 35 iterations to minimize the hourglass-shaped error signature seen in Fig. 4a. The semiautomatic morpher does so by moving the vortex center (i.e., the wind direction singularity) eastward toward the center of the low wind speed area that marks the actual cyclone seclusion (Fig. 4b).
For this SAR image, the simplex algorithm converged to the location shown in Fig. 4b if the operator selected a first-guess position for the cyclone center well within the low wind area (seclusion) around which the frontal wind bands wrap. In contrast, if the operator selected either the hourglass position from Fig. 4a or any position near the inside boundary of the frontal wind bands, the simplex algorithm moved the analyzed cyclone position into the frontal wind band, minimizing the hourglass signature by superimposing it upon the wind band. Thus, the semiautomated system required an educated first guess from an operator in order to perform well.
The quality of the resulting analysis depends not only on the simplex algorithm finding the correct location for the cyclone center, but also on the original model analysis having the correct wind direction pattern around the cyclone center. For example, in the case shown in Fig. 4, the morphing of the cyclone center position brought the wind direction field south and east of the cyclone into agreement with the small-scale features of the SAR backscatter field. Note that the axis of dilatation is moved slightly eastward and rotated slightly so that it aligns closely with the dark, linear frontal signature on the SAR image (Young et al. 2005). Likewise, the wind arrows to the south and east of the cyclone are much more closely aligned with the frontal wind band after the morphing operation, indicating an improvement in the wind direction analysis. In contrast, the wind arrows to the north and west of the cyclone continue to angle across the linear signatures in the SAR image even after the morphing operation. Thus, multipoint morphing, as described in section 3b, would have been required to correct the wind directions in this half of the cyclone.
As with manual morphing, this semiautomated method will fail if a vortex exists in the first-guess wind direction field but does not exist in reality, or vice versa. The SAR-derived wind speed field over the Gulf of Alaska at 0452 UTC 21 January 2004 shown in Fig. 8 illustrates the failure of the automatic morpher when a closed cyclone exists in the NWP model wind direction field but not in reality.
The automorpher converged within 35 iterations but only moved the hourglass error signature (Fig. 8a) northward to the frontal wind speed shift (Fig. 8b). At that point, the error signature was lost in the wind speed gradients associated with the front. The kink in the front at this point suggests that a cyclone may be starting to form in that area, but the failure of either semiautomatic or manual morphing of the cyclone location to eliminate the hourglass error signature suggests that a closed circulation was not present at the time of the SAR overpass.
SAR provides a unique capability for high-resolution remote sensing of wind speed over oceanic regions. Deriving such wind speed fields from SAR backscatter fields requires accurate wind direction fields because of the wind direction dependence of the relationship between SAR backscatter and wind speed. No one source of wind direction fields is entirely adequate for this application. NWP model wind direction fields can have position errors and either missing or extraneous synoptic and mesoscale wind shift features. Scatterometer observations can be contaminated by the mishandling of the ambiguity resolution near frontal and vortex wind shifts. SAR itself can provide wind direction data only in those regions where island wakes, boundary layer rolls, or other meteorological phenomenon yield along-wind streaks in the surface wind speed field. Thus, there are multiple sources of wind direction data, each with their own set of liabilities. The strengths of one source can often, however, be used to compensate for the liabilities of another, allowing appropriately combined data to support SAR-derived wind speed retrievals. There are several potential methods for combining wind direction information from these multiple sources. This paper presents methods for editing gridded (e.g., scatterometer or NWP model) wind direction data to mitigate wind direction errors. All examples presented were for correcting NWP model wind directions.
Manual editing of gridded wind direction data allows for the broadest combination of data types and the maximum input of human insight. The disadvantage of manual editing is that even with a streamlined graphical editor, a single SAR analysis may take 15–30 min of an analyst’s time and half a dozen computationally demanding rederivations of the wind speed from SAR backscatter.
Manual morphing of gridded wind direction data offers the possibility of saving analyst time, and perhaps reducing the number of recomputation cycles, by simply repositioning and reshaping features of the existing wind direction field. This approach is straightforward to implement. The primary disadvantage of manual morphing is that it does not allow the analyst to make full use of all available data sources. If a feature is not present in the gridded wind direction field, the analyst cannot insert it by morphing, even if other wind direction data sources, such as the SAR image itself, indicates that the feature is present. Moreover, while it is generally easy to interpret SAR-derived wind speed error signatures to determine the morph control points, the same cannot be said for their radius of influence.
For cases of misplaced gridded wind direction field singularities such as cyclones, anticyclones, and col points, the existence of a characteristic hourglass-shaped error signature in the SAR-based wind speed field allows for semiautomated application of the morphing approach. This method is faster than manual morphing in that no further input from the human analyst is required after the initial error signature identification. The method is robust in that it yields the same results as single-point manual morphing in those cases where the latter is successful. Its primary disadvantage is the need to assume a priori a radius of influence that may be difficult to gauge. Thus, the best use of semiautomated morphing is as a time-saving first step in a multipoint manual morphing process.
Overall, manual or semiautomated morphing of a NWP model’s wind direction field can serve as one input to the manual data fusion, speeding up that human-intensive task. Given current computational power, an analyst using a single workstation can require as much as an hour to produce a final analysis of wind speed and direction from a SAR backscatter image. The actual analyst time involved is roughly a quarter of this, so operational time constraints could be met with faster workstations or a cluster of computers. For the research case studies presented herein, the analysis time on a single workstation is short enough to make this data fusion technique practical.
This work was supported in part by Grant ATM0240869 from the National Science Foundation and Grants N00014-04-10539, N00014-04-WR-20365, N00014-05-WR-20319, and N00014-06-10046 from the Office of Naval Research. The SAR imagery shown in this paper was provided by the National Oceanic and Atmospheric Administration’s Ocean Remote Sensing Program via the Alaska SAR Demonstration Project.
Corresponding author address: Dr. George S. Young, Department of Meteorology, The Pennsylvania State University, 503 Walker Bldg., University Park, PA 16802. Email: email@example.com