Abstract

The remotely sensed upper-tropospheric water vapor wind information has been of increasing interest for operational meteorology. A new tracer selection based on a local image anomaly and tracking procedure, itself based on Nash–Sutcliffe model efficiency, is demonstrated here for the estimation of upper-tropospheric water vapor winds both for cloudy and cloud-free regions from water vapor images. The pressure height of the selected water vapor tracers is calculated empirically using a height assignment technique based on a genetic algorithm. The new technique shows encouraging results when compared with Meteosat-5 water vapor winds over the Indian Ocean region. The water vapor winds produced by the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) from Meteosat-5 and the present algorithm are compared with collocated radiosonde observations according to Coordination Group for Meteorological Satellites guidelines. The proposed algorithm shows better accuracy in terms of mean vector difference, rms vector difference, standard deviation, speed bias, number of collocations, and mean speed and mean direction differences. Also it is found that the sensitivity of the spatial consistency check in the quality indicator is not so significant for the improvement of statistics.

1. Introduction

The upper-atmospheric wind extracted from the water vapor (WV) images of any geostationary satellite sensors is a very important parameter for operational weather prediction. In the early 1990s, WV images were used only for visual interpretation by synoptic forecasters. However, later studies (Laurent 1993; Holmlund 1993; Tokuno 1996; Velden et al. 1997) in this area have shown much maturity with the ability of automated wind extraction from the WV channels in series of Geostationary Operational Environmental Satellite (GOES), Meteosat satellites, and Geostationary Meteorological Satellites (GMS). The use of geostationary WV imagery has allowed the determination of both upper-level moisture content and the wind fields that correspond to WV layers. One of the main advantages of WV images is their ability to provide atmospheric winds in cloud-free regions as well. A common procedure for the automated extraction of winds from a sequence of WV images is the technique of pattern matching (Merrill 1989). Two very frequently used techniques to match a pattern between a pair of images are the maximum cross correlation (MCC) and minimum of sum of squared difference (MSSD, also known as Euclidean distance method); both methods seek to find the best agreement between an initial target scene and a matching area in the corresponding image pair.

At the National Environmental Satellite Data and Information Service (NESDIS), WV targets are identified by evaluating the bidirectional gradients surrounding each pixel in the target array and selecting the maximum value (Velden et al. 1997). For tracking, a short-range (6–12 h) forecast is used to estimate where to begin the search process for computing the sum of squared differences for all possible scenes in the search window. The scene corresponding to the smallest sum is then selected as the match. This process is repeated forward and backward in time before a final average vector is computed from the two individual estimates.

At the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT), WV winds are extracted operationally (Holmlund 1995) on an equidistant grid (baseline 32 × 32 pixels) with a target size equivalent to the grid size from first-generation geostationary satellites like Meteosat-5/Meteosat-7 imagers over the Indian Ocean region. The algorithm consists of several steps: in the first step, the tracers in WV images are chosen by considering the medium- and high-level clusters, which have the coldest (lowest) WV mean count. The determination of the height of WV image tracers is based on the potentially semitransparency-corrected WV mean count. Note that an inverse atmospheric absorption correction is not carried out in WV images, which might be a deficiency.

The MCC technique is used operationally for tracking WV winds at EUMETSAT (Rattenborg and Holmlund 1996). However, in a research and development mode, the following three basic matching methods had been investigated: 1) cross correlation in the spatial domain, 2) cross correlation in the Fourier domain, and 3) sum of squared distances or Euclidean distances (Dew and Holmlund 2000). The relative quality of the wind vectors produced by the two cross-correlation methods using real WV imagery (Dew and Holmlund 1998) indicates a strong correlation between the mixed radix fast Fourier transform (FFT) and the spatial domain correlation. It also showed that the performance benefit of mixed radix FFT relative to the spatial domain cross-correlation method is significant. The CPU load time is approximately 60% less for a 16/72 target/search area combination and is 20% less for a 32/96 target/search area. The study by Dew (2004) had shown that MSSD tracking is the most effective procedure as compared with both the cross-correlation techniques for the WV (6.1 μm) channel in cloud-free areas. The MSSD technique produces a relatively larger number of high-quality wind vectors in comparison with cross correlation, even in the low-contrast regions of the images.

The Australian Bureau of Meteorology generates WV winds four times per day using three half-hourly Multifunctional Transport Satellite 1R (MTSAT-1R) images. The tracers are selected for potential target areas of 30 × 30 WV imagery by examining maximum and minimum pixel values and brightness temperature gradient maxima criteria, and tracking is done automatically using a model forecast to initiate the search for selected targets on sequential images (Le Marshall et al. 1999, 2002). A lag correlation technique is used to estimate the vector displacement. At the National Satellite Meteorological Center (NSMC) of the China Meteorological Administration, tracers are simply selected at each 1° latitude and longitude grid location both in the infrared (IR) and WV in the central image of three consecutive images. The operational processing area extends from 50°N to 50°S and ±50° in an east–west direction from the subsatellite point, which leads to 101 × 101 candidate tracers for one image triplet per channel (Jianmin et al. 2006). The tracer tracking of the NSMC scheme adopts a stepwise search with a constraint on consistency in time. The search area consists of 96 × 96 pixels around the center of the matching template. Because the maximum correlation method is very computer time intensive, it is not performed at all locations in the search area. Instead a stepwise search procedure consisting of three steps is used. In the first step, correlation calculations are performed for every fourth line and element location. In the successive steps, a refined correlation surface is calculated around the six largest local maxima of the previous step, first at every second location and finally at every possible location. At the end of the correlation calculations, the pixel position of the maxima and the largest secondary peak are kept. This search strategy reduces the computer time to less than one-third of the time needed for the full cross correlation and also gives the same results as correlation tracking 99.8% of the time. The stepwise search procedure makes the NSMC atmospheric motion vector (AMV) derivation scheme independent from NWP forecasts and thus provides truly independent data from the NWP wind field. In dark areas on a WV image, cross correlation often failed in tracing. For those areas in WV images, Euclidean distance is used for tracing (Dew and Holmlund 2000). The absolute maximum and secondary peaks (if present) at two successive image pairs are used to identify potential displacement vectors. The vectors derived from the first image pair are compared with the vectors derived from the second image pair. The most consistent vector pair is selected as the final displacement of the tracer.

The Meteorological Satellite Center (MSC) of the Japan Meteorological Agency has started the retrieval and dissemination of WV winds since 1995 from the GMS series of satellites, which are now replaced with the MTSAT-1R. At this time, operational AMV of MTSAT-1R is derived four times per day at 0000, 0600, 1200, and 1800 UTC (6 hourly) from three consecutive infrared, visible, and/or WV images taken at half-hourly intervals in tracking the movement of clouds in the area from 50°N to 50°S and from 90°E to 170°W. For selecting tracers, an equivalent blackbody temperature (EBBT) histogram analysis is performed for an area of 32 × 32 pixels centered at a candidate point in the WV images. The traditional cross-correlation technique for tracking tracers between two half-hourly images uses a 32 × 32 template window and 64 × 64 window in the searching area of the subsequent image (Tokuno 1996).

The main components for the estimation of WV winds are 1) tracer selection and tracking, 2) quality control, and 3) height assignment. This study focuses on a relook at the techniques of derivation of upper-tropospheric atmospheric winds from the observations from geostationary satellites. EUMETSAT geostationary satellite Meteosat-5 is stationed over the Indian Ocean (63°E) with a horizontal resolution of 5 km both in scan and pixel lines since 1998. In this study an attempt has been made to find an alternate approach to the tracer selection and tracking procedure for the extraction of winds from the WV channel of Meteosat-5. One-month (October 2006) of Meteosat-5 WV image triplets (centered at 0000, 0730, and 1200 UTC, respectively) and corresponding EUMETSAT derived WV winds are also acquired for this study. Further application and verification of this technique will be carried out using observations from the Indian geostationary meteorological satellites KALPANA and Indian National Satellite (INSAT)-3D.

2. Tracer selection and tracking

Water vapor tracers are generally identified using the local bidirectional gradients in a template of specified size and compared with empirically determined thresholds to identify the features with sufficient variability (Velden et al. 1997); those past the threshold value are identified as tracer, and for cloud-free environments the pixel with maximum bidirectional gradient is chosen as tracer. However, in this study tracers are selected by computing a local image anomaly in a 32 × 32 template window, both in cloudy and cloud-free regions. The local image anomaly is calculated using the following formula:

 
formula

where I(i, j) represents the gray value for the (i, j) pixel of a template window and the bar represents the mean of gray values within that template. To compare the performance of gradient-based and anomaly-based tracer selection criteria, we selected a section of Meteosat-5 WV image (0000 UTC 1 October 2006) with reasonable image features. Another image from the same image section, but artificially shifted by 5 pixels and 5 scan lines, was then used for tracking the features based on the above two methods. A simple cross-correlation method was used to track the features. Although both gradient-based (Fig. 1a) as well as anomaly-based tracers (Fig. 1b) captured the expected shift of (−5, −5), it is clearly visible that anomaly-based tracers are superior as compared with gradient-based tracers, because the former method produced a smooth feature field. In comparison, the gradient-based features appear noisy. This difference can help in reducing the tracking errors.

Fig. 1.

Maximum correlation of Meteosat-5 WV tracers calculated (for a sample template from 1 Oct 2006 valid at 0000 UTC) using (a) bidirectional gradient-based and (b) anomaly-based techniques. Contours levels are −0.3, −0.1, 0.1, 0.3, 0.5, 0.7, and 0.9.

Fig. 1.

Maximum correlation of Meteosat-5 WV tracers calculated (for a sample template from 1 Oct 2006 valid at 0000 UTC) using (a) bidirectional gradient-based and (b) anomaly-based techniques. Contours levels are −0.3, −0.1, 0.1, 0.3, 0.5, 0.7, and 0.9.

The cross-correlation technique is used operationally for tracking the tracer between two WV images in most operational centers; however, the cross-correlation technique has a major disadvantage of producing multiple correlation maxima in a template and also of being computationally expensive. A section of Meteosat-5 WV images for 1 October 2006 valid at 0000 and 0030 UTC, respectively, is shown in Figs. 2a,b. The corresponding maximum MCC coefficient for tracking WV tracers between Fig. 2a and Fig. 2b is shown in Fig. 2c. It is clearly visible from Fig. 2c that the cross-correlation technique has produced two maxima, shown by the black shaded region.

Fig. 2.

Meteosat-5 WV images at (a) 0000 and (b) 0030 UTC. The maximum (c) cross-correlation coefficient and (d) Nash–Sutcliffe efficiency coefficient, calculated during tracking between the image in (a) and the image in (b). Contour levels are −0.2, 0, 0.2, 0.4, 0.6, 0.8, and 0.9.

Fig. 2.

Meteosat-5 WV images at (a) 0000 and (b) 0030 UTC. The maximum (c) cross-correlation coefficient and (d) Nash–Sutcliffe efficiency coefficient, calculated during tracking between the image in (a) and the image in (b). Contour levels are −0.2, 0, 0.2, 0.4, 0.6, 0.8, and 0.9.

However, in this study the degrees of matching between two successive images are calculated by the Nash–Sutcliffe model efficiency (Nash and Sutcliffe 1970) coefficient E. The Nash–Sutcliffe model efficiency has been reported in scientific literature for model simulations of discharge and water quality constituents such as sediment, nitrogen, and phosphorous loadings to find the match between modeled discharge and the observed data (Moriasi et al. 2007). Here an attempt has been made to investigate this technique for tracking tracers between two successive images. It is defined as

 
formula

where It and Is are the variance of the gray values for the template window and search window and It is the average of variance of the template window. Here, n = 32 × 32 is the size of the template window and corresponding template in the searching area. The size of the searching area in the subsequent image is taken as 96 × 96. The coefficient E is normalized to values between −∞ and +1. An efficiency E = 1 corresponds to a perfect match, E = 0 means that the search window is as accurate as the mean of the template window, and E < 0 implies the lack of matching between template and search window. The closer the model efficiency is to 1, the more accurate the matching between the windows is. The bar indicates the average of the gray values of the pixels. The maximum Nash–Sutcliffe efficiency coefficient for tracking WV winds between Fig. 2a and Fig. 2b is shown in Fig. 2d. It is clearly visible from Fig. 2d that the Nash–Sutcliffe coefficient has only one maxima (shown by shaded region) with sharp gradient for a randomly chosen tracer window. The Nash–Sutcliffe efficiency method also takes less computer time to reach the maximum point as compared with the cross-correlation method. The same procedure is repeated for the other two subsequent images to derive second displacement vectors. After consistency checks, the average is calculated, which represents WV winds. One of the main advantages of this matching technique is that it reduces the possibility of multiple maxima, because the parameter E has a higher sensitivity to differences between two features than does MCC. In other words, the value of fitness function E falls more rapidly (in comparison with MCC) with the lack of matching between the search and the target window, thus minimizing the possibility of producing erroneous movements of tracers. In one more example, we used two identical image sections and slowly degraded the second image section using random noise of gradually increasing magnitude. Figure 3 shows the values of E and MCC for varying noise level in the second image section. It is clear from this figure that, even with increasing differences between the image features (due to noise), the MCC does not degrade sharply, thus leading to the possibility of multiple maxima. On the other hand, fitness function E falls sharply for the same magnitude of noise. This is clearly an advantage—in particular, when the difference between two features is small, which is often the case in the dry regions of a WV image.

Fig. 3.

The values of fitness function E (dashed) and MCC (solid) for varying noise levels.

Fig. 3.

The values of fitness function E (dashed) and MCC (solid) for varying noise levels.

3. Quality control

The quality check is generally based on vector acceleration checks and simple threshold techniques that compare the derived vectors with their surrounding vectors or with collocated forecast fields. All vectors that show an acceleration, directional deviation, or discrepancy with other observations larger than a predefined value are rejected. In this study, we have employed an automatic quality-control procedure used at EUMETSAT (Holmlund 1998). The scheme derives a quality indicator (QI) for each individual vector based on the properties of the vector itself and its consistency with other vectors. The scheme consists of four different tests, which are normalized by a tanh function that returns a value between 0 and 1. A weighted average of these individual quality indicators is then used for screening of poor-quality vectors from final output. If S is the mean “speed” of a vector computed from two pairs of images, then different quality functions are computed as follows. The direction consistency check DCC is defined as

 
formula

The speed consistency check SPCC is

 
formula

The vector consistency check VCC is defined as

 
formula

Last, the spatial consistency check SCC is found by

 
formula

In the above formulation, Δθ, ΔS, and ΔV represent the difference of direction (in degrees), the difference of speed, and the length of the difference vector between the first and second satellite wind component. Here, ΔVm is the length of the difference vector between the satellite wind component and its best neighbor. The best neighbor is determined by the smallest vector difference. Quantities AN, BN, CN, and DN are constants (EUMETSAT 2005). The final quality indicator of a wind vector is given as

 
formula

All the vectors with QI < 0.6 are rejected. Initially a set of verification statistics is generated by considering w1 = 1, w2 = 1, w3 = 1, and w4 = 0, that is, without considering any spatial consistency check. Similar statistics is also generated by using w1 = 1, w2 = 1, w3 = 1, and w4 = 2 as suggested in EUMETSAT (2005).

Figure 4 shows quality-controlled wind vectors (using the identical quality-control criteria) produced by two tracer selection methods: 1) bidirectional gradient and 2) the new method proposed in the paper (described in section 2). It shows that the new method in general produces a greater number of vectors with higher quality and consistency.

Fig. 4.

A sample figure for 24 Oct 2006 valid at 0730 UTC showing the difference of the density of quality-controlled vectors (using the identical quality-control criteria) produced by two tracer selection methods: (a) bidirectional gradient and (b) the new method proposed in this study.

Fig. 4.

A sample figure for 24 Oct 2006 valid at 0730 UTC showing the difference of the density of quality-controlled vectors (using the identical quality-control criteria) produced by two tracer selection methods: (a) bidirectional gradient and (b) the new method proposed in this study.

4. Height assignment

The height assignment of WV winds is a long-standing problem. In cloud-free regions, the radiometric signal from a pure WV structure is a result of emittance over a finite layer. It is further complicated by radiance contributions from multiple moist layers (Weldon and Holmes 1991). The challenge is to assign a height that best represents the motion of the moisture feature. The most common height assignment technique for WV tracers is to take the effective brightness temperature and assess the height at which displacement of the tracers is attributed (Velden et al. 1997). In this method, the brightness temperature of the target box is averaged and matched with a collocated model guess temperature profile and the level of optimum fit is then used to assign the initial pressure height. Last, this pressure height is corrected with 3D objective analysis using a recursive filter (Hayden and Purser 1995). This method is currently operational at NESDIS. At EUMETSAT, the clear-sky WV winds from first-generation Meteosat satellites were available at a resolution of 160 km using the single-level height assignment based on the cluster EBBT method. Another method based on the WV contribution function calculated from the radiative transfer model was also used to calculate the WV tracer’s height (Rattenborg 2000).

However, for this study we have developed an empirically derived height assignment technique based on the genetic algorithm (GA). The GA is one of the best empirical techniques to determine the best relationship between the independent and dependent parameters. The GA considers an initial population of potential solutions, which is later subjected to an evolutionary process by choosing equations that best fit the data. The newly generated population is subjected to mutations that change fractions of information. The evolutionary steps are repeated with the new generation. The process ends after a number of generations a priori determined by the user. More details about GA applications are given by Szpiro (1997), Alvarez et al. (2001), and Singh et al. (2006). A very short description of the GA is as follows. Let p() be a smooth mapping function that explains the relationship between a desired variable x and a set of independent variables {a, b, c, d, e, . . .}, so that

 
formula

First, for an amplitude function x, a set of candidate equations for p() is randomly generated. An equation is stored as a set of characters that define the independent variables a, b, c, d, e, and so on, in the above equation, and four arithmetic operators (+, −, ×, and /). A criterion that measures how well the equation strings perform on a training set of the data is its fitness to the data, defined as the sum of the squared differences between data and the parameter derived from the equation string. The equations with best fits are then selected to exchange parts of the character strings between them while the equations with less fits are discarded. Last, a small percentage of the equations strings, single operators, and variables are mutated at random. The process is repeated a large number of times to improve the fitness of the evolving equations. The fitness strength of the best-scoring equation is defined as

 
formula

where Δ2 = Σ(xcx0)2, xc is parameter value estimated by the best scoring equation, x0 is the corresponding “true” value, and 〈x0〉 is the mean of the true values of x. Szpiro (1997) has shown the robustness of the GA to forecast the behavior of a one-dimensional chaotic dynamical system. Later, a number of studies have been reported using the GA for the prediction of space–time variability of sea surface temperature (Alvarez et al. 2000), estimation of surface heat fluxes (Singh et al. 2006), and monthly mean air–sea differences (Singh et al. 2005) from satellite observations. In the current study, an attempt has been made to use this empirical approach to determine the height of the WV tracer both for cloudy and noncloudy pixels.

The development of the retrieval algorithm for the estimation of tracer height involves a number of steps. In the first step, a number of independent variables from the imagers (brightness temperatures of the coldest pixel and warmest pixel, cosine of latitude, and zenith angle information of the center of template window, etc.) are considered in a large set of possible parameters. In the second step, we choose randomly a large number of Meteosat-5 images and corresponding WV winds (derived by EUMETSAT) from a 1-month (October 2006) period as the training/validation dataset. Approximately 120 000 valid WV wind vectors were available in the above dataset, but only 20% of the data were further selected randomly for the purpose of training, and the remaining data were used for validation. A small ratio of training and validation data size is expected to ensure the robustness of the retrieved functions and also prevents the possibility of overfitting. Separate optimized functions were generated for cloudy and noncloudy scenes (templates). A 32 × 32 template window was considered to be “cloudy” if the average brightness temperature of the 25 coldest pixels was less than 220 K. The genetic algorithm is an automatic method that determines the best-fitting relationship between dependent and independent fields using a random search and optimization criteria. In this case, the optimized GA solution retains only the following independent parameters that are needed for height assignment of a tracer: 1) average brightness temperature of the 25 coldest pixels, 2) average brightness temperature of the 25 warmest pixels, and 3) cosine of latitude at the center of the template window. However, the form of the function changes from cloudy to noncloudy tracers. One advantage of the GA method is that the complex and often nonlinear relations can be obtained in functional forms that are easier to use than lookup tables. The GA-based empirical function developed for noncloudy and cloudy pixels after training and optimization is as follows:

 
formula
 
formula

Here x1 and x2 represent the mean radiance of the five coldest and warmest pixels in the template window and x3 represents the cosine of latitude for the center of the window. The values of the coefficients (a, b, c, d, . . . , etc.) are given in Table 1. Here H1 and H2 are the two functions defined by GA for derivation of tracer height for noncloudy and cloudy pixels. These mapping functions are then used for estimation of WV tracer height information. The tracer heights derived by the above two equations are in hectopascals. The current GA-based approach is an ad hoc method and tries to mimic statistically the operational height assignment method used in Meteosat-5, which has its own limitations. Because the above GA-based mapping functions are generated using the training of Meteosat-5 very high resolution radiometer (VHRR), the utilization of this technique with the other satellites needs further training, because the mapping will be sensitive to the satellite sensor response function.

Table 1.

Values of coefficients.

Values of coefficients.
Values of coefficients.

A typical example of WV winds derived from Meteosat-5 VHRR for 31 October 2006 valid at 0000 UTC using the present technique is shown in Fig. 5. It shows that the present technique is able to produce the wind with uniform coverage, large-scale and synoptic-scale features are captured well, and the vertical distribution of information is between the 100- and 500-hPa portion of the troposphere.

Fig. 5.

A typical example of WV winds for 31 Oct 2006 valid at 0000 UTC derived from Meteosat-5 VHRR using the algorithm presented in this paper.

Fig. 5.

A typical example of WV winds for 31 Oct 2006 valid at 0000 UTC derived from Meteosat-5 VHRR using the algorithm presented in this paper.

5. Verification

The validation of the present height retrieval algorithm is given in Fig. 6, where the height derived by the present algorithm is compared with the height derived by the EUMETSAT algorithm. The mean absolute error in height is approximately 27 hPa, and the combined correlation of cloudy and noncloudy pixels is about 85%. The quantitative evaluation of derived WV winds is calculated according to the Coordination Group for Meteorological Satellites (CGMS) guidelines, in which derived WV winds are validated with collocated radiosonde winds. According to CGMS guidelines, the vector difference VD between an individual wind i and the collocated rawinsonde wind r used for verification is given by

 
formula

The speed bias (BIAS) is calculated as

 
formula

Last, the mean vector difference MVD is reported as

 
formula

and the standard deviation SD about the mean vector difference traditionally reported is

 
formula

The root-mean-square error RMSVD traditionally reported is the square root of the sum of the squares of the mean vector difference and the standard deviation about the mean vector difference,

 
formula

It is suggested that one report MVD and SD, along with mean radiosonde speed SPD and number of collocations with radiosonde data NC. Here the unit of MVD, RMSVD, SD, SPD, and BIAS is meters per second. These statistics can provide a fixed measure of product quality over time and can be employed in determining the observation weight in objective data assimilation.

Fig. 6.

Validation of derived WV tracer height [estimated pressure (hPa) on x axis] with corresponding EUMETSAT-derived height (y axis; hPa).

Fig. 6.

Validation of derived WV tracer height [estimated pressure (hPa) on x axis] with corresponding EUMETSAT-derived height (y axis; hPa).

To validate the present algorithm, we first applied this technique to 1-month of Meteosat-5 VHRR consisting of three triplets (centered at 0000, 0730, and 1200 UTC, respectively) for each day of October of 2006, and the corresponding EUMETSAT WV winds are acquired. The derived winds by the present algorithm from Meteosat-5 and corresponding EUMETSAT winds are compared with collocated radiosonde for each day by calculating different statistical parameters as discussed above for the region 50°N–50°S and 30°–130°E. During collocation, the 1.0° × 1.0° latitude/longitude grid point is considered, and speed and direction differences of more than 30 m s−1 and 90° relative to radiosonde, respectively, are filtered out.

The MVDs, mean speed differences MSD, mean direction differences MDD, total derived vectors TV, and NC for all available triplets are shown in Figs. 7a,c,e. Here MVD-M and MVD-S represent the MVD of EUMETSAT-derived WV winds and the WV winds derived by the present algorithm (in which spatial consistency check SCC is not considered in the QI when compared with collocated radiosonde winds), respectively, for the level 500–100 hPa. In a similar way, acronyms are used for MSD-M, MSD-S, MDD-M, MDD-S, TV-M, TV-S, NC-M, and NC-S. The same set of statistical parameters is also calculated using the present algorithm by taking into consideration SCC. It is observed from Fig. 7a that in 63 cases out of 91 the present algorithm has performed better as compared with the EUMETSAT-derived product. The difference of MVD-S and MVD-S (SCC) is shown in Fig. 7b. In Fig. 7b MVD-S (SCC) represents the mean vector difference calculated using the present algorithm by considering the spatial consistency check in the QI routine. When SCC is applied to the present algorithm, the MVD value reduced in 54 cases out of 91 (Fig. 7b) from the earlier cases when SCC is not applied in QI. The time series of MSD-M, MDD-M, MSD-S, and MDD-S are shown in Fig. 7c. The MSDs are represented by a bar diagram in Fig. 7c. In the cases of MSD and MDD (Fig. 7c), 52 cases out of 91, our algorithm is performing well as compared with the EUMETSAT product when both are compared with radiosonde data. The differences of MSD-S and MSD-S (SCC) and MDD-S and MDD-S (SCC) are shown in Fig. 7d. In Fig. 7d MSD-S (SCC) and MDD-S (SCC) represent the mean speed and direction differences calculated using the present algorithm by considering the SCC in the QI routine. When SCC is applied to the present algorithm, the MSD value was reduced in 51 cases out of 91 (Fig. 7d) from the earlier cases when SCC is not applied in QI. When only the derived wind speeds are compared at collocation points, both the EUMETSAT product and the present algorithms show similar error characteristics. The errors in wind speed are random in both the products with no distinguished bias. However, the present algorithm shows clear improvement over the EUMETSAT algorithm in capturing the wind direction. Out of 804 collocations where radiosonde, EUMETSAT, and the retrievals from the present algorithm were simultaneously available, the retrieved direction from the present algorithm is more accurate relative to the EUMETSAT product at 468 points. With the use of SCC, the present algorithm has performed better in 479 cases out of 826 collocations. This shows that the SCC factor in QI is not playing a significant role in the improvement of statistics in the present algorithm. Thus this improvement can be attributed to the more robust tracking method in which the number of false targets gets reduced significantly. Figure 7e represents the total number of derived winds (by lines) and number of collocations on each day (by bars), and 54 cases out of 91 total numbers of derived winds as well as the number of collocations are larger in our algorithm as compared with the corresponding EUMETSAT-derived product. Table 2 shows the MVD, RMSVD, SD, SPD, BIAS, NC, MSD, and MDD when the validation is done for the whole month of October of 2006 using the three triplets for each day centered at 0000, 0730, and 1200 UTC, respectively. It indicates that when a comparison is done on a monthly basis, the present algorithm shows some improvement over the EUMETSAT product. The use of SCC in the present algorithm has shown little improvement over the case in which SCC is not applied at 0000 and 1200 UTC, when more radiosonde data are available; however, this improvement is not noticed at 0730 UTC. Thus it can be concluded that there is no significant improvement in the accuracy even with the use of the SCC in the present algorithm.

Fig. 7.

(a) MVD without SCC in MVD-S, (b) difference of MVD-S and MVD-S with SCC, (c) mean speed and direction difference in MVD-S without SCC, (d) difference of MSD-S and MDD-S with and without SCC, and (e) total number of derived vectors and number of collocations on each day of October 2006, valid at 0000, 0730, or 1200 UTC.

Fig. 7.

(a) MVD without SCC in MVD-S, (b) difference of MVD-S and MVD-S with SCC, (c) mean speed and direction difference in MVD-S without SCC, (d) difference of MSD-S and MDD-S with and without SCC, and (e) total number of derived vectors and number of collocations on each day of October 2006, valid at 0000, 0730, or 1200 UTC.

Table 2.

Statistical parameters with and without SCC.

Statistical parameters with and without SCC.
Statistical parameters with and without SCC.

6. Conclusions

A new tracer selection based on local image variance and tracking procedure, itself based on Nash–Sutcliffe model efficiency, is demonstrated here successfully. Though the empirical height assignment technique based on the genetic algorithm looks promising, with mean absolute error in height of approximately 27 hPa and combined correlation of cloudy and noncloudy pixels of approximately 85%, the current GA-based approach is an ad hoc method and tries to statistically mimic the operational height assignment method used in Meteosat-5, which has its own limitations. It uses only image information—namely, the 1) average brightness temperature of the 25 coldest pixels, 2) average brightness temperature of the 25 warmest pixels, and 3) cosine of latitude at the center of the template window, and external information like numerical model outputs (both for tracking and height assignment) is not used in this algorithm. This algorithm is also capable of giving WV winds both in cloudy and cloud-free regions together. The verification with radiosonde data confirmed that this technique shows some improvement from the existing EUMETSAT technique, at least over the Indian Ocean region, when accuracy is judged in terms of mean vector difference, root-mean-square vector difference, standard deviation, mean speed difference, mean direction difference, and bias. At the collocations where radiosonde, EUMETSAT, and the retrievals from the present algorithm were simultaneously available, the retrieved wind direction from the present algorithm is more accurate than the EUMETSAT product. It is also found that there are no significant improvements in the accuracy with the use of the spatial consistency check in the present algorithm. Thus this improvement can be attributed to the more robust tracking method in which the number of false targets gets reduced significantly.

Acknowledgments

The authors thank the three anonymous reviewers for their critical and insightful comments/valuable suggestions, which were helpful in substantially improving the content and quality of presentation of this manuscript. The authors are also thankful to EUMETSAT for providing 1-month Meteosat-5 VHRR images and the corresponding derived winds. The director, the deputy director, RESA, and group director MOG/RESA of Space Applications Centre (SAC), ISRO Ahmedabad are acknowledged for their encouragement and help. The authors are also thankful to Shri A. S. Kirankumar, Deputy Director SEDA, Shri C. M. Nagrani of SEDA, and Dr. B. S. Gohil of MOG SAC Ahmedabad for their critical comments during the initial phase of the development of this algorithm.

REFERENCES

REFERENCES
Alvarez
,
A.
,
C.
Lopez
,
M.
Riera
,
E.
Hernandez-Garcia
, and
J.
Tintore
,
2000
:
Forecasting the SST space-time variability of the Alboran Sea with genetic algorithms.
Geophys. Res. Lett.
,
27
,
2709
2712
.
Alvarez
,
A.
,
A.
Orfila
, and
J.
Tintore
,
2001
:
DARWIN: An evolutionary program for nonlinear modeling of chaotic time series.
Comput. Phys. Commun.
,
136
,
334
349
.
Dew
,
G.
,
2004
:
Investigation of target tracking techniques in the MSG MPEF environment.
Proc. Seventh Int. Winds Workshop, Helsinki, Finland, EUMETSAT, 153–161
.
Dew
,
G.
, and
K.
Holmlund
,
1998
:
Improved computing efficiency in the Fourier domain.
Proc. Fourth Int. Winds Workshop, Saanenmoser, Switzerland, EUMETSAT, 289–298
.
Dew
,
G.
, and
K.
Holmlund
,
2000
:
Investigations of cross-correlation and Euclidian distance target matching techniques in the MPEF environment.
Proc. Fifth Int. Winds Workshop, Lorne, Australia, EUMETSAT, 235–243
.
EUMETSAT
,
2005
:
Wind vector automatic quality control scheme. EUMETSAT Rep. EUM/OPS/TEN/05/1747, 10 pp
.
Hayden
,
C. M.
, and
R. J.
Purser
,
1995
:
Recursive filter objective analysis of meteorological fields: Applications to NESDIS operational processing.
J. Appl. Meteor.
,
34
,
3
15
.
Holmlund
,
K.
,
1993
:
Operational water vapor wind vectors from Meteosat imagery.
Proc. Second Workshop on Wind Extraction from Operational Satellite Data, Tokyo, Japan, EUMETSAT, 77–84
.
Holmlund
,
K.
,
1995
:
Half hourly wind data from satellite derived water vapour measurements.
Adv. Space Res.
,
16
,
10
.
59
68
.
Holmlund
,
K.
,
1998
:
The utilization of statistical properties of satellite-derived atmospheric motion vectors to derive quality indicators.
Wea. Forecasting
,
13
,
1093
1104
.
Jianmin
,
X.
,
Z.
Qisong
,
Z.
Xiaohu
,
W.
Sujuan
, and
L.
Feng
,
2006
:
Status of operational AMVs from FY-2C.
Proc. Eighth Int. Winds Workshop, Beijing, China, EUMETSAT, P.47
.
Laurent
,
H.
,
1993
:
Wind extraction from Meteosat water vapor channel image data.
J. Appl. Meteor.
,
32
,
1124
1133
.
Le Marshall
,
J.
,
N.
Pescod
,
R.
Seecamp
,
K.
Puri
,
C.
Spinoso
, and
R.
Bowen
,
1999
:
Local estimation of GMS-5 water vapor motion vectors and their application to Australian region numerical weather prediction.
Aust. Meteor. Mag.
,
48
,
73
77
.
Le Marshall
,
J.
,
G.
Mills
,
N.
Pescod
,
R.
Seecamp
,
K.
Puri
,
P.
Stewart
,
L. M.
Leslie
, and
A.
Rea
,
2002
:
The estimation of high-density atmospheric motion vectors and their application to operational numerical weather prediction.
Aust. Meteor. Mag.
,
51
,
173
180
.
Merrill
,
R.
,
1989
:
Advances in automated production of wind estimates from geostationary satellite imagery. Preprints, Fourth Conf. on Satellite Meteorology, San Diego, CA, Amer. Meteor. Soc., 246–249
.
Moriasi
,
D. N.
,
J. G.
Arnold
,
M. W.
Van Liew
,
R. L.
Bingner
,
R. D.
Harmel
, and
T.
Veith
,
2007
:
Model evaluation guidelines for systematic quantification of accuracy in watershed simulations.
Trans. ASABE
,
50
,
885
900
.
Nash
,
J. E.
, and
J. V.
Sutcliffe
,
1970
:
River flow forecasting through conceptual models part I: A discussion of principles.
J. Hydrol.
,
10
,
282
290
.
Rattenborg
,
M.
, and
K.
Holmlund
,
1996
:
Operational wind products from new Meteosat Ground Segment.
Proc. Third Int. Winds Workshop, Ascona, Switzerland, EUMETSAT, 53–59
.
Singh
,
R.
,
C. M.
Kishtawal
, and
P. C.
Joshi
,
2005
:
Estimation of monthly mean air-sea temperature from satellite observations using genetic algorithm.
Geophys. Res. Lett.
,
32
.
L02807, doi:10.1029/2004GL021531
.
Singh
,
R.
,
C. M.
Kishtawal
,
P. K.
Pal
, and
P. C.
Joshi
,
2006
:
Surface heat fluxes over global ocean exclusively from satellite observations.
Mon. Wea. Rev.
,
134
,
965
980
.
Szpiro
,
G. G.
,
1997
:
Forecasting chaotic time series with genetic algorithm.
Phys. Rev.
,
55
,
2557
2568
.
Tokuno
,
M.
,
1996
:
Operational system for extracting cloud motion and water vapor motion winds from GMS-5 image data.
Proc. Third Int. Winds Workshop, Ascona, Switzerland, EUMETSAT, 21–30
.
Velden
,
C. S.
,
C.
Hayden
,
S.
Nieman
,
W. P.
Menzel
,
S.
Wanzong
, and
J.
Goerss
,
1997
:
Upper-tropospheric winds derived from geostationary satellite water vapor observations.
Bull. Amer. Meteor. Soc.
,
78
,
173
195
.
Weldon
,
R. B.
, and
S. J.
Holmes
,
1991
:
Water vapor imagery: Interpretation and applications to weather analysis and forecasting. NOAA Tech Rep. NESDIS 67, 213 pp
.

Footnotes

Corresponding author address: Dr. S. K. Deb, Atmospheric Sciences Division, Meteorology and Oceanography Group, Space Applications Centre, ISRO, Ahmedabad 380015, India. Email: sanjib_deb@sac.isro.gov.in