Simulating Cloud Motion Vectors From Global Circulation Model Data For Use in OSSEs: A Preliminary, But Useful, Algorithm For Application to Current NASA/NOAA OSSE Projects.



October 18, 2001






Christopher O’Handley

G.D. Emmitt

Steven Greco


Simpson Weather Associates

809 E. Jefferson St.

Charlottesville, VA 22902











1. Introduction


a) Overview


The purpose of this report is to describe the method we have developed for producing realistic cloud motion wind vectors (CMV) for use in OSSEs. Currently, such CMV are obtained by simply using the locations of all observed CMV for the time of interest. The numerical model wind at each location/height is assigned to the CMV at that location, and a gaussian random error is added to each report. This approach has two major shortcomings. First, it may result in CMV in areas where the model has no clouds. Second, there are no correlated errors (such as the slow bias discussed below) in the resulting CMV.


To assure realistic impact assessments using OSSEs, our objective is to produce simulated cloud motion vectors (CMV) that exhibit both random and correlated errors similar to those occurring in practice. We will describe the procedures currently used to produce operational CMV, their spatial distributions, and the magnitudes and sources of errors. We then describe our approach for simulating CMV and their errors using numerical prediction data. We close with a description of future work that is needed to evaluate the resulting winds and further develop the techniques. Note that this paper deals only with CMV from visible and infrared imagery; future work will address the need to simulate CMV from water vapor imagery.


b) Operational cloud motion wind algorithms


The following description, based primarily on Nieman et al (1997), corresponds most closely to the procedures used to derive GOES CMV at NESDIS.


Generally, a set of 3 infrared (IR) images, at 30 minute intervals, is used to derive CMV. The first step is an image registration analysis between the three images, using the middle image as the starting point. Although the images are navigated and registered prior to receipt by NESDIS, registration errors greater than 2 km (equal to a 1 m/s speed error over 30 minutes) are encountered on a regular basis. The registration correction involves an examination of as many landmarks as possible on the images; the individual shifts are then averaged to produce a single correction in x and y for the image set. Only corrections which exceed 2 km are actually applied, which suggests that the residual registration error is less than 2 km in x and y. However, because of variations in the number of usable landmarks, there are times when the residual error exceeds 2 km. Also, since a single correction is applied across the whole image, there is no way to account for rotational shifts or other nonlinear types of error.


The second step is target identification. A grid of approximately 60x60 km is used to search the first image of a triplet for sufficiently non-homogeneous cloud patterns. The algorithm uses spatial-coherence analysis to identify the highest (brightest) cloud in the scene, and find the region of strongest brightness gradient. The cloud is identified as a target if the gradient exceeds a certain magnitude. While any type of cloud can be tracked, scenes with multiple cloud decks are eliminated because these make cloud height difficult to determine.


Once a target has been identified, the cloud top height is determined. Several methods are available. The simplest, which is used as a back-up, is the infrared window approach, in which the coldest 20% of the sample is assumed to represent the cloud top temperature, and height is determined by comparison with a temperature profile from a numerical forecast. More advanced techniques, such as the CO2 slicing and H2O intercept methods (Nieman et al, 1993), are used when possible, and are particularly beneficial in cases where the target cloud is semi-transparent (and thus susceptible to contamination by surface radiation). Both methods use information from multiple channels and require both cloudy and clear-sky radiances. Though more accurate, these techniques also rely on information from numerical forecasts to estimate radiances. The current GOES satellites have sensors which allow use of the H2O intercept, whereas 1993-era GOES sensors required use of CO2 intercept. Recently, techniques have also been developed that enable estimation of cloud base heights for low-level (below 700 mb) cloud tracers, as it is known that such clouds generally move with the wind speed close to cloud base.


The fourth step is the actual tracking of target clouds from image to image. The position of the target on image 2 is estimated using the wind predicted by a numerical forecast model. A correlation procedure is then used to search the second image; if the maximum correlation exceeds a pre-defined threshold, the target is identified and the resulting 30 minute shift defines the first CMV. This step is repeated using images 2 and 3, which yields a second CMV.  If the difference between the two vectors is sufficiently small (generally about 5 m/s), an average CMV is computed from the two vectors, and the resulting vector is passed on to the quality control system.


While the above steps are fairly similar among the various CMV producers (Butterworth, 1997a), the quality control procedure which follows differs considerably (Holmlund et al, 2001). The most advanced, completely automated procedures are implemented at EUMETSAT and NESDIS. The NESDIS technique, known as the auto-editor, involves two steps of a fully three-dimensional recursive filter objective analysis. In the first step, a background analysis is produced using a numerical forecast along with all available non-CMV observations. This is followed by a second step which incorporates the CMV. During this process, a penalty function is used to re-assign the heights of the CMV to the level of best fit if necessary (the adjustment is limited to 150 mb), and winds which differ greatly from the resulting analysis are flagged. At EUMETSAT, the individual CMV vectors are put through a series of consistency tests that primarily compare them to other CMV reports, though one test involves a model forecast field. Both methods also yield quality control flags for each CMV.


c) Known errors in CMV


Recent comparison studies between CMV and co-located radiosondes indicate that rms error values of CMV increase with height, ranging from about 3-4 m/s below 700 mb, to 4-6 m/s between 400 and 700 mb, and 6-8 m/s at higher levels (Tomassini et al, 1999). It should be noted that these are not absolute accuracies, as radiosondes themselves have errors ranging from roughly 2 m/s (low levels) to 4 m/s (high levels). More troubling is a persistent slow bias in CMV, particularly at higher wind speeds. Some studies have found the slow bias to be as large as 5-6 m/s in uncorrected CMV. This error is so important (jetstream wind speeds are already underforecasted by numerical models) that NESDIS actually implements a “slow bias correction” as part of their auto-editor. All CMV whose speed exceeds 10 m/s are augmented with 7% of the forecast wind speed at that location. The corrected CMV that result from the auto-edit procedure still exhibit a slow bias of approximately 1-2 m/s, depending on windspeed.


While the magnitudes of CMV errors have been well defined, the exact causes of the errors are somewhat uncertain. It has been suggested that the slow bias may be due to non-representative cloud motions, i.e. that the tracer clouds, generally cirrus, do not move with the winds near jet level. This has not been conclusively shown, although it is well known that other clouds do not move with the wind at cloud top (the height at which the CMV is assigned). For instance, tropical cumulus have been shown to move with the wind near cloud base, and there is often a 1-2 m/s difference between the winds at cloud base and cloud top (Peslen et al, 1986). It would seem that other explanations could account for the slow bias. In particular, the presence of sub-visual cirrus clouds above the target clouds could contaminate the height assignment, making the target cloud appear higher. Where the  windspeed increases upward, the effect would be to place a slower CMV higher up in the atmosphere, in a region with faster windspeed – thus introducing a slow bias. This will be examined in more detail below.


More generally, the height assignment process is susceptible to numerous errors. In the case of semi-transparent cloud targets, contamination by upwelling surface radiation can corrupt the height assignment. The CO2 and H2O techniques help circumvent this problem, but at the expense of introducing information from numerical forecast models. In certain situations, this presumably could lead to large errors in height assignment. Other sources of error include image registration shifts, inaccuracies in the correlation algorithms used to identify and track cloud tracers, and cloud development and dissipation.


d) CMV data coverage


At present, CMV are produced from the following 5 satellites: GOES-W (135 W), GOES-E (75 W), METEOSAT (0 W), GMS (140 E), and INSAT (80 E). Generally, winds are produced within about 45-60° from the satellite sub-point. This results in nearly global coverage within about 45° of the equator. However, the exact types of winds (infrared/visible/water vapor), as well as the techniques and quality control procedures, differ among the producers. For comparison, in 1993 there were 4 satellites producing CMV (the above minus INSAT), which resulted in a gap of 30-45° in longitude in the Indian Ocean region.




e) CMV use in data assimilation


Because of the differences in types of winds produced, methods used, and resulting accuracies, the major weather prediction centers have put in place severe restrictions on the use of CMV in their data assimilation systems. For example, at ECMWF (Butterworth 1997b), winds from INSAT are not used, nor are any visible winds above 700 mb. Also rejected are IR winds over land that are north of 20° N, and IR winds over land below 500 mb that are south of 20° N.  Furthermore, all high-density winds are thinned to one observation per (approximately) 1.25° box at each model level; the observations are chosen on the basis of the quality control flags. Note that other types of restrictions are applied to water vapor and sounder winds, which are not discussed here. At NCEP, INSAT winds are not used, GMS winds are used everywhere, and GOES and Meteosat winds are used everywhere south of 20° N, and only over the ocean elsewhere. Note that CMV use in data assimilation is constantly being evaluated and modified, so this description may not be 100% accurate.


2. Data


For this study, simulated  cloud track winds were produced using data from an ECMWF T213 spectral model 30-day forecast, which serves as a  “nature run” for the NPOESS OSSE project. The data was available on a one half degree latitude/longitude grid at 31 pressure levels from 1000 to 100 mb, every 6 hours, from 06 UTC 2/5/1993 to 00 UTC 3/6/1993. Parameters that were used include cloud fraction, cloud liquid water content, temperature, and u and v wind components. Low-level cloud fractions (below 800 mb) were corrected  to eliminate the global under-estimation of low cloud in the nature run (Masutani et al, 1999).


3. Methods and Results


a) Cloud-motion winds: target selection


At each model grid point within 60° of a satellite sub-point, a “satellite cloud view” (SCV) is calculated by applying cloud overlap assumptions to the model cloud fractions. For contiguous cloud layers (those in successive model layers), maximum overlap is assumed, i.e. the cloud layers are considered vertically stacked. As an example, assume that the model cloud fraction is 40% in the higher of 2 layers, and 50% in the next layer. The SCV of the first layer is 40%. Applying maximum overlap, 40% of the cloud in the second layer is considered obscured, leaving an additional SCV of 10% in that layer. If the second layer has less cloud than the higher layer, none of it can be seen. For non-contiguous cloud layers, i.e. cloudy layers separated by at least one clear layer, random overlap is applied.  Using the above example, the SCV in the lower layer could range from 10% to 50%. If the higher layer has 40% and the lower layer 30%, then the SCV in the lower layer could range from 0-30%, i.e. all of the cloud might be seen. The random overlap essentially states that non-contiguous cloud layers are considered to be completely unrelated dynamically.


The SCV profile is then examined from the top down to determine if a suitable cloud tracer exists. Any layer with a SCV cloud fraction of between 0.05 and 0.25 is assumed to be a potential tracer, and the first such level encountered is chosen as the tracer for that location. It is assumed that if the SCV is less than 0.05, there is insufficient cloud for tracking. On the other hand, if SCV exceeds 0.25, the cloud is assumed to be more stratiform in nature, with insufficient edges or brightness variations for tracking to be successful.  This approach is based on the general targeting algorithms described earlier, which rely on inhomogeneities in the cloud field. The exact SCV range is somewhat arbitrary, and has been adjusted to give “reasonable” distributions of CMV. Further work is needed to evaluate the CMV distributions that result. In particular, these values have been adjusted to give CMV totals that are close to 1993 levels; current operational CMV totals are much higher.


Once the target has been located, the T213 model u and v wind components are recorded as the true wind associated with the CMV at that location. An example of the distribution of CMV produced using this method is shown in Figure 1. Panels a and b contain the column totals of opaque and cirrus cloud (SCV), which show the distribution of clouds and weather systems. Panels c-f show the locations of the CMV in 4 different layers, and Table I lists the number of CMV for this case, statified by height and cloud fraction. The total number of CMV is about 4 times larger than the number of true CMV for the same time period, and the vertical distribution is too heavily skewed to the low layers.  It is assumed that inclusion of cirrus targets would increase the number of high level CMV while reducing those at lower levels; this remains for future work. However, the spatial distribution of the simulated CMV is reasonably similar to observed distributions of CMV in 1993, especially if the simulated vectors are randomly thinned to the same total number. This is demonstrated in Figure 2 (note that these figures are 12 hours apart).


b) Modeling of Slow Bias


As indicated earlier, comparative studies between CMV and radiosonde winds have revealed the presence of a slow bias which increases with wind speed. It has been suggested that these types of errors exhibit spatial correlations which could have a strong negative impact on numerical weather prediction (Butterworth, 1997b; Holmlund et al, 2001). Unfortunately, no studies have clearly documented such a spatial correlation, though it seems reasonable to suspect that the bias patterns might be correlated along the path of jet streams.


The cause of the slow bias is not known. However, it has been attributed  to non-representative cloud motions near the jet stream or, in the specific case of water vapor winds, to improper height assignment or edge effects on cloud tracking (Daniels et al, 2000). Our attempt to simulate a slow bias is based on the idea that a slow bias may be the result of contamination of the cloud height estimate by thin or sub-visual cirrus clouds. If present, such clouds would make the target cloud appear to be colder than it really is, and thus higher in the atmosphere. Since winds generally increase upward throughout most of the troposphere, this would effectively introduce a slow bias by making a lower, slower-moving cloud appear to be at a higher height, where windspeeds are faster.


Our approach is as follows. For each profile where a cloud target is found, we determine if thin cirrus cloud is likely to be present above the target cloud. If such cirrus is predicted, then the pressure difference dp between the highest cirrus and the target cloud is calculated. dp is multiplied by a gaussian random deviate (between 0-1), and the resulting value dpadj is used to adjust the pressure of the target cloud upward to a new apparent pressure (papp). The model wind speed at this new level is the equivalent of the true, or radiosonde speed. However, the wind speed at the original level is still applied to the CMV. This introduces a “cold bias” error which may or may not also be a negative wind speed error, depending on the vertical profile of wind speed.


As it turns out, this procedure does produce a slow bias. Table II shows the resulting bias values for all CMV above 700 mb, for the case shown in Figure 1. The speed bias is tabulated two ways, with respect to the true or radiosonde speed (BIAS1) and then with respect to the CMV speed (BIAS2). BIAS1, which is equivalent to the way the values are usually reported in the literature, shows a strong slow bias that increases dramatically with increasing wind speed. The bias values for speeds above 30 m/s greatly exceed any published statistics, which may indicate that we are allowing vertical adjustments that are too large. However, it should be noted that the sample size is relatively small. BIAS2 shows a much more constant slow bias across all speed ranges. This differs from the recent findings of Lalaurette and Garcia-Mendez (2000), who plotted the slow bias for winds above 400 mb versus CMV speed. They found a linear increase in the slow bias from 0 m/s at 15 m/s to –5 m/s at 45 m/s. Unfortunately, we have not seen other observational studies in which the slow bias is tabulated based on CMV speed for comparison. Our overall simulated slow bias of –2.15 m/s is reasonably close to recently published values (Tomassini et al, 1999)


The other issue regarding the slow bias is horizontal correlation of the errors. Figure 3 shows those CMV that have a slow bias greater than –1 m/s (panel a) and those that have a fast bias greater than +1 m/s (panel b). Clearly this procedure does introduce concentrated areas of slow bias, in regions where cirrus clouds are found (compare to Figure 1). Unfortunately, there are no published studies showing spatial distributions of bias errors in operational CMV for comparison. However, observational studies have shown that the slow bias is predominantly a mid-latitude phenomenon, and this is reflected in the spatial distributions shown in Figure 3.


c) Image Registration Errors


As noted earlier, NESDIS applies a registration correction to each image triplet if the calculated registration errors exceed 2 km. However, the accuracy of the correction depends on the number and distribution of landmarks, which can vary considerably. Thus, errors exceeding 2 km may be fairly common. In addition, rotational and non-linear errors cannot be eliminated using the current approach. Since a 2 km shift in 30 minutes introduces a wind speed error of 1.1 m/s at nadir (and more near the edge of the image), this is obviously a significant source of correlated error that must be included in simulated CMV.


The main difficulty in simulating these errors is the spherical nature of the earth. A simple west-east registration shift will introduce only zonal wind errors, though the magnitude of the error will increase toward the edges of the image. However, a north-south shift will introduce both longitudinal and zonal wind components, with the ratio of the components changing depending on latitude and longitude. To quantify the distribution and magnitude of wind errors across the CMV domain, map projection equations are used. We have used the equations for the equatorial azimuthal orthographic projection:


                                           x = cos(lat)sin(lon)                                                                 (1)    

                                           y = sin(lat)                                                                              (2)


The resulting projection, shown in Figure 4, is not exactly the satellite projection, but it is very close.


The procedure to estimate wind errors introduced by a registration shift is as follows. First, an x,y shift (dx, dy) is randomly assigned for the image as a whole. This value is assumed to be valid at nadir. Then, at the location of each cloud target, the following is done. First, equations (1) and (2) are used to calculate the unshifted x and y positions of the point based on the true latitude and longitude; this location is recorded as xtrue and ytrue. Next, the shifted longitude position xshift is determined from xtrue + dx, and a new longitude is back-calculated from equations (1) and (2). This is repeated for latitude, resulting in a new (or apparent) latitude and longitude for the location. Finally, the differences between the true and apparent latitudes and longitudes are used to determine the distance shifts in km, which yields a wind speed which corresponds to the image registration error.


Figure 5 shows the estimated wind errors resulting from registration shifts of 2 km over 30 minutes. A due eastward shift (panel a) causes westerly winds across the domain, with speeds of 1-2 m/s. The wind speed is constant in latitude, but increases toward the western and eastern edges of the domain. A northward shift (panel b) causes a symmetric pattern, with southerly winds predominating near the center and increasing zonal components toward the corners. Total wind speeds can exceed  3 m/s at the corners of the map. A combined northward/eastward shift (panel c) causes a more complicated pattern, with wind speeds greater than 2 m/s over significant parts of the domain and maxima exceeding 5 m/s in the SW and NE corners


4. Discussion/Future Work


We have produced simulated CMV whose spatial distribution appears broadly similar to operational CMV from 1993. However, the vertical distribution is too highly skewed to lower altitudes, and the total number of vectors is too large by a factor of 4. Several items need to be explored to fix these problems. First, we need to examine the cloud overlap calculation, particularly the method used to identify thin (sub-visual) cirrus clouds.  After this, we will need to add cirrus targets into the tracking algorithm. Operational CMV producers do track cirrus clouds, and these make up a significant source of CMV. It is presumed that using cirrus targets would give us a more realistic vertical distribution of CMV; it might also change the slow bias results significantly. We will need to modify the algorithm to simulate CMV of the density that are currently produced, which greatly exceeds the levels of 1993. This will involve testing different cloud fraction ranges for the target cut-off, and examining the spatial patterns of CMV that result.


We need to more carefully analyze the results of our slow bias simulation. Observational studies frequently present the bias statistics as a function of latitude and height; we have only looked at the entire sample above 700 mb. We will also look at the spatial patterns to determine how error magnitudes are correlated; this should also involve an examination of the direction errors (observations suggest that there is little directional component to the slow bias). This process should help us determine if we are producing a slow bias for the wrong reason.


Many researchers have indicated that CMV are frequently contaminated by surface radiation, which would tend to produce a warm, or low, bias in the cloud heights. Seemingly, this would tend to produce a fast bias and in fact at slower wind speeds, observed CMV do tend to show a slightly positive bias (Lalaurette and Garcis-Mendez, 2000). We need to incorporate a surface radiation component to our height adjustment, probably only in cases where the target cloud is “thin”. Potentially, a single, thin target cloud could be affected by both a surface-induced warm bias and a cirrus-induced cold bias, of varying magnitudes. It will be interesting to see how these effects combine, and whether a slow bias is still present in our simulated CMV.


Finally, our registration error simulator could be improved by incorporating the map projection equations for the true satellite projection. It is believed that this would lead to slightly larger induced winds, especially near the edges of the domain. We may also want to add the ability to create rotational image shifts, though this may not be necessary as the current approach serves the purpose of introducing wind errors that are correlated across the satellite domain.


5) References


Butterworth, P., 1997a: NWP on-line scientific note no, 2: Investigation into satellite winds – I. Production.


Butterworth, P., 1997b: NWP on-line scientific note no. 3: Investigation into satellite winds – II. Treatment at UKMO.


Daniels, J., C. Velden, W. Bresky and A. Irving, 2000: Status and development of operational GOES wind products.  Proc. Fifth Intl. Winds Workshop, Lorne, Australia.


Holmlund, K., C.S. Velden, and M. Rohn, 2001: Enhanced automated quality control applied to high-density satellite-derived winds. Mon. Wea. Rev., 129, 517-529.


Lalaurette, F. and A. Garcia-Mendez, 2000: Monitoring satellite winds at ECMWF. Proc. Fifth Intl. Winds Workshop, Lorne, Australia


Masutani, M, K.A. Campana and S-K Yang, 1999: Note on cloud cover of the ECMWF nature run used for OSSE/NPOESS project. Office note 427. NCEP.


Nieman, S.J., J. Schmetz and W.P Menzel, 1993: A comparison of several techniques to assign heights to cloud tracers. J. Appl. Met., 32, 1559-1568.


Nieman, S.J., W.P. Menzel, C.M. Hayden, D. Gray, S.T. Wanzog, C.S. Velden, and J. Daniels, 1997: Fully automated cloud-drift winds in NESDIS operations. Bull. Amer. Met. Soc., 78, 1121-1133.


Peslen, C.A., S.E. Koch and L.W. Uccellini, 1986: The effect of the arbitraty level assignment of satellite cloud motion wind vectors on wind analyses in the pre-thunderstorm environment. J. Clim. Appl. Met., 25, 615-632.


Tomassini, M., G. Kelly and R. Saunders, 1999: Use and impact of satellite atmospheric winds on ECMWF analyses and forecasts. Mon. Wea. Rev., 127, 971-986.




Table I: Number of cloud track winds for each atmospheric layer, and range of cloud fraction for those winds.












































Table II: Speed bias (m/s) for all simulated CMW above 700 mb. BIAS1 is the bias as tabulated by radiosonde (true) speed, with NUM1 the number of observations for that category. BIAS2 and NUM2 are the same, but tabulated against the CMW speed.


Speed range





0-10 m/s





10-20 m/s





20-30 m/s





30-40 m/s





40-50 m/s
















































Figure 1: Simulated cloud track winds for 1200 UTC 2/6/1993. (a) Column total cirrus (b) column total opaque cloud (c) Low-level CMV (below 850 mb)














Figure 2: Comparison between actual cloud track winds from 1993 and simulated winds. (a) Actual CMV from 0000 UTC February 6, 1993 (b) simulated CMV from Figure 1, randomly thinned to approximately the same density as the observed winds.


















Figure 3: Locations of CMV above 700 mb that exhibit a bias greater than 1 m/s. (a) slow bias (b) fast bias


Figure 4: Map projection used to simulate registration errors. Latitude and longitude shown at 10° intervals.









Figure 5: Wind vectors resulting from image registration shifts of 2 km in 30 minutes. (a) eastward shift (b) northward shift (c) northward and eastward shift.