• Borde, R., , and Oyama R. , 2008: A direct link between feature tracking and height assignment of operational atmospheric motion vectors. Proc. Ninth Int. Winds Workshop, Annapolis, MD, EUMETSAT, 8 pp. [Available online at http://www.eumetsat.int/Home/Main/AboutEUMETSAT/Publications/ConferenceandWorkshopProceedings/2008/groups/cps/documents/document/pdf_conf_p51_s3_13_borde_v.pdf.]

  • Borde, R., , and Dubuisson P. , 2010: Sensitivity of atmospheric motion vectors height assignment methods to semitransparent cloud properties using simulated Meteosat-8 radiances. J. Appl. Meteor. Climatol., 49, 12051218.

    • Search Google Scholar
    • Export Citation
  • Bormann, N., , Hernandez-Carrascal A. , , Borde R. , , Lutz H.-J. , , Otkin J. A. , , and Wanzong S. , 2014: Atmospheric motion vectors from model simulations. Part I: Methods and characterization as single-level estimates of wind. J. Appl. Meteor. Climatol.,in press.

    • Search Google Scholar
    • Export Citation
  • Büche, G., , Karbstein H. , , Kummer A. , , and Fischer H. , 2006: Water vapor structure displacements from cloud-free Meteosat scenes and their interpretation for the wind field. J. Appl. Meteor. Climatol., 45, 556575.

    • Search Google Scholar
    • Export Citation
  • Chang, F.-L., , Minnis P. , , Lin B. , , Khaiyer M. M. , , Palikonda R. , , and Spangenberg D. A. , 2010: A modified method for inferring upper troposphere cloud top height using the GOES 12 imager 10.7 and 13 μm data. J. Geophys. Res., 115, D06208, doi:10.1029/2009JD012304.

    • Search Google Scholar
    • Export Citation
  • Cho, H.-J., , Sohn E.-H. , , and Ou M.-L. , 2008: The impact of target box size on wind speed biases in satellite-derived atmospheric motion vectors. Proc. 2008 EUMETSAT Meteorological Satellite Conf., Darmstadt, Germany, EUMETSAT, 5 pp. [Available online at www.eumetsat.int/groups/cps/documents/document/pdf_conf_p_s12_50_cho_p.pdf.]

  • Doutriaux-Boucher, M., , Forsythe M. , , Saunders R. , , and Francis P. , 2006: MSG wind height assignment problems. Met Office Forecasting Research Tech. Rep. 469, 30 pp.

  • EUMETSAT, 2011: MGS meteorological products extraction facility algorithm specification document. EUMETSAT Doc. EUM/MSG/SPE/022, 259 pp.

  • Hernandez-Carrascal, A., , and Bormann N. , 2014: Atmospheric motion vectors from model simulations. Part II: Interpretation as spatial and vertical averages of wind and role of clouds. J. Appl. Meteor. Climatol.,in press.

  • Holmlund, K., 1998: The utilization of statistical properties of satellite-derived atmospheric motion vectors to derive quality indicators. Wea. Forecasting, 13, 10931104.

    • Search Google Scholar
    • Export Citation
  • Holmlund, K., 2000: The atmospheric motion vector retrieval scheme for Meteosat Second Generation. Proc. Fifth Int. Winds Workshop, Lorne, Australia, EUMETSAT, 201–208.

  • Lean, P., , Migliorini S. , , and Kelly G. , 2012: Studying the relationship between synthetic NWP-derived AMVs and model winds. Proc. 11th Int. Winds Workshop, Auckland, New Zealand, EUMETSAT, 8 pp. [Available online at http://www.eumetsat.int/website/wcm/idc/idcplg?IdcService=GET_FILE&dDocName=PDF_CONF_P60_S5_04_LEAN_V&RevisionSelectionMethod=LatestReleased&Rendition=Web.]

  • Le Marshall, J., , Pescod N. , , Khaw A. , , and Allan G. , 1993: The real-time generation and application of cloud-drift winds in the Australian region. Aust. Meteor. Mag., 42, 89103.

    • Search Google Scholar
    • Export Citation
  • Menzel, W. P., , Smith W. L. , , and Stewart T. , 1983: Improved cloud motion wind vector and altitude assignment using VAS. J. Climate Appl. Meteor., 22, 377384.

    • Search Google Scholar
    • Export Citation
  • Nieman, N. J., , Schmetz J. , , and Menzel W. P. , 1993: A comparison of several techniques to assign heights to cloud tracers. J. Appl. Meteor., 32, 15591568.

    • Search Google Scholar
    • Export Citation
  • Roebeling, R., , Baum B. , , Bennartz R. , , Hamann U. , , Heidinger A. , , Thoss A. , , Walther A. , 2013: Evaluating and improving cloud parameter retrievals. Bull. Amer. Meteor. Soc.,94, ES41–ES44.

  • Schmetz, J., , and Nuret M. , 1989: Cloud motion wind estimates in Europe. ECMWF/EUMETSAT Workshop on the Use of Satellite Data in Operational Numerical Weather Prediction: 1989–1993, Vol. 2, ECMWF/EUMETSAT, 274283.

  • Schmetz, J., , Holmlund K. , , Hoffman J. , , Strauss B. , , Mason B. , , Gärtner V. , , Koch A. , , and van de Berg L. , 1993: Operational cloud motion winds from Meteosat infrared images. J. Appl. Meteor., 32, 12061225.

    • Search Google Scholar
    • Export Citation
  • Schmetz, J., , Pili P. , , Tjemkes S. , , Just D. , , Kerkmann J. , , Rota S. , , and Ratier A. , 2002: An introduction to Meteosat Second Generation (MSG). Bull. Amer. Meteor. Soc., 83, 977992.

    • Search Google Scholar
    • Export Citation
  • Schreiner, A. J., , Menzel W. P. , , Straka W. , , and Heidinger A. , 2012: Comparison of CO2 and H2O atmospheric motion vector height assignment techniques using the GOES-13 imager. Proc. 11th Int. Winds Workshop, Auckland, New Zealand, EUMETSAT, 6 pp. [Available online at http://www.eumetsat.int/website/wcm/idc/idcplg?IdcService=GET_FILE&dDocName=PDF_CONF_P60_S5_05_MENZEL_V&RevisionSelectionMethod=LatestReleased&Rendition=Web.]

  • Sèze, G., , Marchand S. , , Pelon J. , , and Borde R. , 2008: Comparison of AMV cloud top pressure derived from MSG with space based lidar observations (CALIPSO). Proc. Ninth Int. Winds Workshop, Annapolis, MD, EUMETSAT, 8 pp. [Available online at http://www.eumetsat.int/website/wcm/idc/idcplg?IdcService=GET_FILE&dDocName=PDF_CONF_P51_S3_15_BORDE_V&RevisionSelectionMethod=LatestReleased&Rendition=Web.]

  • Sohn, E., , and Borde R. , 2008: The impact of window size on AMV. Proc. Ninth Int. Winds Workshop, Annapolis, MD, EUMETSAT, 6 pp. [Available online at http://www.eumetsat.int/website/wcm/idc/idcplg?IdcService=GET_FILE&dDocName=PDF_CONF_P51_S4_18_SOHN_V&RevisionSelectionMethod=LatestReleased&Rendition=Web.]

  • Stubenrauch, C. J., , Eddounia F. , , and Sauvage L. , 2005: Cloud heights from TOVS Path-B: Evaluation using LITE observations and distributions of highest cloud layers. J. Geophys. Res., 110, D19203, doi:10.1029/2004JD005447.

    • Search Google Scholar
    • Export Citation
  • Szejwach, G., 1982: Determination of cirrus cloud temperature from infrared radiances: Application to Meteosat. J. Appl. Meteor., 21, 284293.

    • Search Google Scholar
    • Export Citation
  • Velden, C. S., , and Bedka K. M. , 2009: Identifying the uncertainty in determining satellite-derived atmospheric motion vector height attribution. J. Appl. Meteor. Climatol., 48, 450463.

    • Search Google Scholar
    • Export Citation
  • von Bremen, L., 2008: Using simulated satellite images to improve the characterization of atmospheric motion vectors (AMVs) and their errors for numerical weather prediction. NWP SAF Rep. NWPSAFEC-VS-015, Version 1.4, 41 pp. [Available from http://research.metoffice.gov.uk/research/interproj/nwpsaf/vs.html.]

  • Watts, P. D., , Bennartz R. , , and Fell F. , 2011: Retrieval of two‐layer cloud properties from multispectral observations using optimal estimation. J. Geophys. Res., 116, D16203, doi:10.1029/2011JD015883.

    • Search Google Scholar
    • Export Citation
  • View in gallery

    (a) Infrared counts within the target area in image 1 and image 2, and (b) CLA-CTH, CCij coefficients, and 25% coldest pixels (from image 2). AMV has been extracted using IR10.8 channel of SEVIRI (1145 and 1200 UTC 3 Aug 2006 images; lat: 15.26°, lon: 4.37°). In the classification tool bar, indexes 210–212, 220–222, and 230–232 correspond to low-level, midlevel and high-level clouds, respectively.

  • View in gallery

    Infrared counts within the target area are plotted against their individual pixel contributions for the corresponding target area illustrated in Fig. 1. Diamond symbols correspond to the clear-sky pixels and the plus symbols represent the cloudy pixels within the target area. Vertical dash line represents the averaged individual contribution.

  • View in gallery

    CLA-CTH of the cloudy pixels present in the target area plotted as function of CCij. Dark triangles represent the weighted pressure calculated using the pixels with a CCij larger than 〈CCij〉 and that are on the coldest branch. A clear diamond corresponds to the weighted pressure for the whole cold branch (CCij > 0). A solid line represents the histogram of the CLA-CTH product.

  • View in gallery

    As in Fig. 1, but for another AMV example (1145 and 1200 UTC 3 Aug 2006 images; lat: 14.76°, lon: 34.2°).

  • View in gallery

    Plots as in Figs. 2 and 3, but according to the AMV example presented in Fig. 4. This example corresponds to a multilayer cloudy situation.

  • View in gallery

    AMV pressure distribution for all the AMVs extracted in (top to bottom) channels Vis0.8, WV6.2, WV7.3, and IR10.8 (1–30 Jun 2012). The new CCC results are shown in black. The results of the former AMV algorithm are shown in gray.

  • View in gallery

    Amount of AMVs extracted in Vis0.8 channel and average QI for the CCC method (dashed line) and former AMV algorithm (solid line) during the period 1–30 Jun 2012: (two top plots) high-level winds (100–400 hPa), (two middle plots) midlevel winds (400–700 hPa), and (two lowest plots) low-level winds (700–1000 hPa). Only AMVs with a QI larger than 85 are used.

  • View in gallery

    As in Fig. 6, but for the WV6.2 channel. Only high-level and midlevel results are presented.

  • View in gallery

    As in Fig. 6, but for the IR10.8 channel.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 0 0 0
PDF Downloads 0 0 0

A Direct Link between Feature Tracking and Height Assignment of Operational EUMETSAT Atmospheric Motion Vectors

View More View Less
  • 1 EUMETSAT, Darmstadt, Germany
© Get Permissions
Full access

Abstract

Height assignment (HA) is currently the most challenging task in the operational atmospheric motion vectors’ (AMV) extraction scheme. Several sources of error are associated with the height assignment step, including the sensitivity of the HA methods to several atmospheric parameters. However, one of the main difficulties is to identify, for the HA calculation, the most significant image pixels used in the feature-tracking process. The most widely used method selects the coldest pixels in a representative target box (e.g., coldest 25%) to infer the height of the detected feature, irrespective of what was tracked. This paper presents a method based on a closer link between the pixels used for tracking and their HA. The individual contribution to the overall tracking cross-correlation coefficient is used to identify the most significant pixels contributing to the tracking. This approach has been implemented operationally at European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) to derive AMVs since September 2012. This paper details the method, gives specific examples, and provides a first glance at its performances and benefits for the operational AMV production.

Corresponding author address: Régis Borde, EUMETSAT, Eumetsat Allee 1, D-64295 Darmstadt, Germany. E-mail: regis.borde@eumetsat.int

Abstract

Height assignment (HA) is currently the most challenging task in the operational atmospheric motion vectors’ (AMV) extraction scheme. Several sources of error are associated with the height assignment step, including the sensitivity of the HA methods to several atmospheric parameters. However, one of the main difficulties is to identify, for the HA calculation, the most significant image pixels used in the feature-tracking process. The most widely used method selects the coldest pixels in a representative target box (e.g., coldest 25%) to infer the height of the detected feature, irrespective of what was tracked. This paper presents a method based on a closer link between the pixels used for tracking and their HA. The individual contribution to the overall tracking cross-correlation coefficient is used to identify the most significant pixels contributing to the tracking. This approach has been implemented operationally at European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) to derive AMVs since September 2012. This paper details the method, gives specific examples, and provides a first glance at its performances and benefits for the operational AMV production.

Corresponding author address: Régis Borde, EUMETSAT, Eumetsat Allee 1, D-64295 Darmstadt, Germany. E-mail: regis.borde@eumetsat.int

1. Introduction

Atmospheric motion vectors (AMVs) derived from all geostationary satellites constitute a significant part of the observation data assimilated in numerical weather prediction (NWP) models. This is because they are the only upper-level wind observations with good global coverage for the tropics and midlatitudes, especially over the large oceanic areas. AMVs are routinely extracted all around the world by a number of meteorological satellite operators such as the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT), the National Oceanic and Atmospheric Administration (NOAA)’s National Environmental Satellite, Data, and Information Service (NESDIS), the Cooperative Institute for Meteorological Satellite Studies (CIMSS), the Japan Meteorological Agency (JMA), the Korea Meteorological Administration (KMA), the China Meteorological Administration (CMA), and the India Meteorological Department (IMD).

The derivation of displacement vectors is realized by tracking clouds or water vapor features in consecutive satellite images. At EUMETSAT the final hourly AMV product extracted from the Meteorological Satellite (Meteosat) Second Generation (MSG) (Schmetz et al. 2002) is an average of three vectors calculated from a sequence of four consecutive Spinning Enhanced Visible and Infrared Imager (SEVIRI) images that observe the full Earth disk every 15 min. AMVs are derived for two visible [Vis0.8 and high-resolution visible (HRVis)] and water vapor (WV6.2 and WV7.3) channels and one infrared (IR10.8) channel. The spatial resolution at nadir is 3 km for all SEVIRI channels except the HRVis channel, where it is 1 km, allowing retrieval of smaller-scale motion winds. The basic steps of wind vector production at EUMETSAT are (Holmlund 2000) 1) selecting a feature to track, 2) tracking the selected target in a time sequence of images to obtain a relative motion (a vector), 3) assigning a pressure (altitude) to the vector, and 4) assessing the quality of the retrieved vector (Holmlund 1998). A complete description of the algorithm can be found in the EUMETSAT algorithm specification document (EUMETSAT 2011).

AMVs extracted from satellite imagery only represent the movement of a specific layer of the atmosphere (Velden and Bedka 2009) and are mainly a spatial and temporal average of local motions. Schmetz and Nuret (1989) stated that AMVs only give an unbiased estimate of winds if clouds are conservative tracers, randomly distributed and floating within the airflow. This is clearly not the case. Clouds are not randomly arranged but are associated with specific conditions such as, for example, ascending air masses. In addition, while some clouds do move with the wind, orographic clouds, for example, do not. It is then really difficult to know which motion the AMVs extracted by the algorithm really represent. Sohn and Borde (2008) and Cho et al. (2008) recently studied the temporal and spatial scaling issues that are associated with AMV extraction. Other recent studies based on simulated images have also helped to improve our understanding of what the AMVs are representing (von Bremen 2008; Lean et al. 2012; Bormann et al. 2014; Hernandez-Carrascal and Bormann 2014), even if the question still remains open.

Various sources of error are introduced in the AMV extraction process, but the vector height assignment is still recognized to be the most difficult part. The altitude of the AMVs is difficult to set for semitransparent and/or multilayer clouds (Roebeling et al. 2013). It is also the parameter that most impacts the quality of the AMVs, and therefore their role in NWP assimilation. Several methods are used to set the AMV altitude, depending on the type of target tracked. For opaque clouds the measured window channel IR10.8 brightness temperature is matched against a collocated temperature profile obtained from NWP model forecast data. Specific methods are applied for low-level AMVs that include setting the wind altitude to the cloud base (LeMarshall et al. 1993) or, when it exists, to the level of a temperature inversion. At high and midlevels, several techniques exist to estimate the cloud-top pressure of semitransparent clouds (Szejwach 1982; Nieman et al. 1993; Schmetz et al. 1993; Menzel et al. 1983), for which part of the signal received by the satellite is composed of the surface contribution and/or other lower cloudy layers. However, these methods are not always accurate, especially in the case of multilayer situations (Borde and Dubuisson 2010; Schreiner et al. 2012; Doutriaux-Boucher et al. 2006; Sèze et al. 2008). In addition to the error coming from cloud-top estimation, Borde and Oyama (2008) showed that, for these methods, the pixels influencing the tracking solution are not necessarily the same as those selected to assign a height to the tracer. The implied disconnection between the two processing steps can lead to poor-quality winds and large errors.

This paper describes the change recently implemented in the EUMETSAT operational AMV algorithm that aims to minimize a potential source of error in the AMV extraction. Use is made of the cross-correlation contribution (CCC) method to improve the height assignment estimate. The CCC method proposed in Borde and Oyama (2008) establishes a very strong link between the feature that is tracked in two consecutive satellite images and the estimation of the altitude. In the tracking process, the individual pixel contributions to the cross-correlation coefficient are used to weight the importance of the pixels used in the calculation of the pressure, as defined by Büche et al. (2006). The pressure assigned to the AMV is calculated as a weighted average of the individual cloud-top pressures estimated for each cloudy pixel, using the cross-correlation contribution weights. The first part of this paper describes the CCC method and its advantages. The second section shows how the method has been used to improve the quality of the EUMETSAT Meteorological Product Extraction Facility (MPEF) operational AMV products. Finally, the last section summarizes a discussion of the issues addressed.

2. CCC method

The CCC method is applied for all AMV channels. However, the results presented in this section show AMVs that are derived from tracking clouds in the infrared IR10.8-μm channel of Meteosat-8, using 24 × 24 pixel (3-km resolution) target boxes. No filters and/or enhancement processes are used in this study.

The first step of the AMV extraction scheme is to find a target (as a contrast feature) at a selected grid point. A search area of 48 × 48 pixels centered on the grid point is used. For each possible location of the target box in the search area, a measure of contrast is calculated using the following two methods: the local standard deviation (LSD) of 3 × 3 pixels computed at the center (contrast) and the number of pixels (NLSD) within the target box with an LSD larger than a configurable value. The target location is selected to be the 24 × 24 target box, which has maximum contrast AND with NLSD larger than a configurable number. Overlap of more than 50% is not allowed between adjacent targets. The scheme attempts to find a cloudy target, defined as containing more than 50 pixels classed as cloudy (based on the MSG cloud mask product). For water vapor (WV) channels only, if a cloudy target is not found, then a search for a clear-sky target is attempted.

Having located the target in this first image, a search for its corresponding position, 15 min later, in a second image is made. The search is carried out in an 80 × 80 pixel box centered on the target location in the first image. A cross-correlation method is used for the matching. The upper part of Fig. 1 shows an example of the target area identified in two consecutive 10.8-μm infrared channel MSG images, at 1145 and 1200 UTC 3 August 2006, respectively [latitude: 15.26°, longitude: 4.37°; AMV speed: 11 m s−1; final quality index (QI): 0.88]. The final QI is estimated in the range 0–1 following Holmlund (1998). Counts in the IR10.8 channel are plotted in Fig. 1a, and the corresponding cloud scene, cloud analysis (SCE-CLA) and cloud analysis cloud-top height (CLA-CTH) information are plotted in the top plots of Fig. 1b. High-level and low-level clouds correspond to white and gray colors, respectively, in the classification images picture.

Fig. 1.
Fig. 1.

(a) Infrared counts within the target area in image 1 and image 2, and (b) CLA-CTH, CCij coefficients, and 25% coldest pixels (from image 2). AMV has been extracted using IR10.8 channel of SEVIRI (1145 and 1200 UTC 3 Aug 2006 images; lat: 15.26°, lon: 4.37°). In the classification tool bar, indexes 210–212, 220–222, and 230–232 correspond to low-level, midlevel and high-level clouds, respectively.

Citation: Journal of Atmospheric and Oceanic Technology 31, 1; 10.1175/JTECH-D-13-00126.1

The pattern-matching scheme compares the individual pixel counts of the target box with all possible locations of the target box in the search area to find the best match. The correlation matching in the above-mentioned example uses count values, but radiance information can also be used. The degree of matching between pixel radiances a and b of the two images A and B, respectively, is classically given by the following two-dimensional cross-correlation coefficient:
e1
where m, n is the (lines, elements) displacement of the target box in (search) image B from its initial position in the first image A. The correlation coefficient CC(m, n) is normalized to values between −1 (mirror structures) and +1 (identical structures). The terms ā and σa represent the average and standard deviation of the radiances a in image A, respectively (and correspondingly for b in image B). Values M and N correspond to the box size (M × N = 24 × 24) for this study. For all m, n displacement values, the best pixel-accurate target match corresponds to the maximum of CC(m, n). According to Büche et al. (2006), for target boxes in the two images, the correlation coefficient can be written as Eq. (2), where the symbol CCij expresses the contribution of each individual pair of pixels (i, j) and (i + m, j + n) to the total correlation coefficient of the pair a and b(m, n):
e2
Figure 2 illustrates how the individual pairs of pixels taken from the 24 × 24 pixel target boxes contribute to the maximization of CC(m, n). Within the target area, the “diamond” symbol corresponds to a clear-sky pixel and the “plus” symbol to a cloudy pixel. Usually, the coldest and the warmest pixels in the target box contribute most to the CC(m, n). When a clear distinction between cold and warm scenes exists, the relative individual pixel contributions, CCij, present a “C shaped” distribution against infrared counts. The distance between the two branches corresponds to the contrast of the structures within the target area. The example plotted in Figs. 1 and 2 corresponds to a single-layer cloud moving above a homogeneous surface. The correlation is very high, CC = 0.95. In more complicated cases, the appearance and/or decay of clouds between images 1 and 2 generally induces a negative CCij. Pixels that contribute most to the CC(m, n) can be defined as those that have a CCij larger than the average contribution, 〈CCij〉. These are indicated by the dashed vertical line in Fig. 2.
Fig. 2.
Fig. 2.

Infrared counts within the target area are plotted against their individual pixel contributions for the corresponding target area illustrated in Fig. 1. Diamond symbols correspond to the clear-sky pixels and the plus symbols represent the cloudy pixels within the target area. Vertical dash line represents the averaged individual contribution.

Citation: Journal of Atmospheric and Oceanic Technology 31, 1; 10.1175/JTECH-D-13-00126.1

The two branches of this C-shaped curve are below and above the average count value of the pixels present within the target area, respectively. The cold branch corresponds to cloud-top pixels, whereas the warm branch is more frequently composed of clear-sky scenes, cloud edges, fractional clouds, semitransparent clouds and, in cases of multilayer scenes, low-level clouds.

a. Estimation of the pressure using CCij information

EUMETSAT provides specific information on cloud type and cloud-top height (CLA-CTH) for all the cloudy pixels in Meteosat images. A complete description can be found in the EUMETSAT algorithm specification document (ASD). The current operational CLA-CTH product estimates cloud-top height, using the water vapor ratioing method for semitransparent clouds and IR brightness temperature for opaque clouds. The CO2 slicing method (Menzel et al. 1983) can also be used instead of the water vapor ratioing method to correct the semitransparency effect. The CCC method estimates the AMV pressure P as the average CLA-CTH pressure of the selected pixels, weighted by their individual contribution to the correlation coefficient CCij.:
e3

If there are no pixels in the cold branch with an individual contribution larger than the average (CCij larger than 〈CCij〉), then all pixels in the cold branch (with CCij larger than 0) are instead used to estimate the pressure.

The weighted pressure standard deviation Psd (hPa) is calculated using the same set of pixels. This standard deviation gives information on the variability of the pixels used to estimate the altitude:
e4

For AMVs derived using infrared channels, only the pixels with a successfully extracted CLA-CTH value, which lie on the “cold branch” of the Fig. 2 plot (radiance smaller than the average radiance within the target area), and for which CCij is larger than the average CCij, 〈CCij〉, are selected to calculate the AMV pressure. The selected pixels are illustrated in the bottom-left panel of Fig. 1b. For comparison, the bottom-right panel of Fig. 1b shows the 25% coldest pixels within the target box.

Figure 3 shows the CLA-CTH of the cloudy pixels present in the target area, plotted as a function of CCij. The pressure contributions to the CCC method vary between 196 ± 88 hPa, using pixels of the cold branch having a CCij larger than 0, and 188 ± 50 hPa, using a CCij larger than the average 〈CCij〉. Nearly 16% of the coldest pixels contribute to the latter pressure estimate. The pressure estimated using the classical 25% coldest pixels solution is slightly larger, equal to 219 hPa, setting the cloud top 30 hPa lower in the troposphere than for the CCC method. The one estimated using the coldest peak of the CLA-CTH histogram is 193 hPa. The AMV example illustrated in Figs. 1 and 2 corresponds to the motion of a single cloudy layer at high levels. It is noted that, even for such an easy case, the various pressures calculated from the different methods are slightly different.

Fig. 3.
Fig. 3.

CLA-CTH of the cloudy pixels present in the target area plotted as function of CCij. Dark triangles represent the weighted pressure calculated using the pixels with a CCij larger than 〈CCij〉 and that are on the coldest branch. A clear diamond corresponds to the weighted pressure for the whole cold branch (CCij > 0). A solid line represents the histogram of the CLA-CTH product.

Citation: Journal of Atmospheric and Oceanic Technology 31, 1; 10.1175/JTECH-D-13-00126.1

Generally, for the AMV extraction process, there are many different situations, depending on the target types and local atmospheric conditions. Most of them are much more complicated than the example shown in Fig. 1 and include multilayer situations, thin semitransparent cirrus, or fractional clouds.

Figures 4a,b and 5 show another case that corresponds to a multilayer cloudy situation (latitude: 14.76, longitude: 34.2; AMV speed: 8.6 m s−1; final QI: 0.82). In Fig. 4b, low-level clouds are plotted in gray and high-level clouds in white. The correlation is high, CC ~ 0.97, and the group of pixels that really drives the tracking process is mainly located at the bottom right of the target box. They are high-level cloudy pixels, representing nearly 16% of the total, and their individual contribution to the correlation is very large. The low-level cloudy pixels do not contribute much to the tracking. Therefore, the detected speed, 8.6 m s−1, refers only to the motion of the highest layer.

Fig. 4.
Fig. 4.

As in Fig. 1, but for another AMV example (1145 and 1200 UTC 3 Aug 2006 images; lat: 14.76°, lon: 34.2°).

Citation: Journal of Atmospheric and Oceanic Technology 31, 1; 10.1175/JTECH-D-13-00126.1

Fig. 5.
Fig. 5.

Plots as in Figs. 2 and 3, but according to the AMV example presented in Fig. 4. This example corresponds to a multilayer cloudy situation.

Citation: Journal of Atmospheric and Oceanic Technology 31, 1; 10.1175/JTECH-D-13-00126.1

The pressures estimated using the CCC method are 259 ± 59 hPa, using pixels of the cold branch having a CCij larger than 0, and 252 ± 44 hPa, using CCij larger than the average 〈CCij〉, respectively. The pressure estimated using the classical 25% coldest pixels is equal to 357 hPa, nearly 100 hPa lower than pressure obtained using the CCC method. The pressure estimated from the coldest peak is similar to the CCC method, about 250 hPa.

These examples illustrate two of the many situations that can be encountered during the AMV extraction process, and the inherent difficulty in calculating the altitude of the features that represent the atmospheric motion. A detected motion vector does not necessarily correspond to the displacement of a single cloud layer. It generally represents an average displacement of cloud segments that are moving at different speeds and altitudes within the target box. The use of CCij for the final height estimation enables a dynamic and specific selection of the feature that contributes most significantly to the detected motion. It also allows the derivation of an error bar associated with the AMV pressure, which is a strong request from NWP centers to help improve the assimilation process.

b. AMVs extracted from the visible channels

At EUMETSAT cloud motion winds are also derived at low levels from the visible channel at 0.8 μm and in the HRVis channel. For the visible channels, the same AMV extraction algorithm is used as for the IR channels. The height assignment (HA) is also the major problem in this process. As the scattering of photons on cloud particles dominates the radiative transfer processes in the visible part of the spectra, the radiance in a visible channel cannot be converted directly into a temperature using the Planck equation. Indeed, even for thick opaque clouds the cloud-top height cannot be estimated directly from the visible channel, but instead uses the corresponding infrared channel at 10.8 μm. This is a major source of error in the case of multilayer situations, because the coldest pixels in the infrared channel within the target box do not necessarily correspond to the tracked feature in the visible channel.

Using Lidar In-Space Technology Experiment (LITE) observations, Stubenrauch et al. (2005) found the frequency of occurrence of multilayer clouds equal to 55% over land and 38% over ocean. Indeed, using 24 × 24 pixel target boxes to track AMVs with geostationary satellites, the likelihood of multilayer situations within the box is very high. Satellite systems with retrieval methods that assume a single-layer cloud do not provide correct cloud-top height information in the case of multilayer situations (Borde and Dubuisson, 2010; Chang et al. 2010). An appropriate strategy is then required to correctly select the pixels that represent the feature tracked in the visible channel and then to use the IR channel radiance to estimate the altitude of the wind vector.

As for the infrared channel, the cross-correlation process uses the contrast of the pixels present within the target box, and a plot similar to Fig. 2 can be produced by considering the visible reflectance as a function of the CCij. The cloud tops that correspond to pixels with a smaller radiance in the IR channel correspond to the pixels with a larger reflectance in the visible channels. According to Eq. (3), the pixel selection area that must be used to estimate the altitude of the visible-channel-extracted AMVs is the brightest part of the C-shape curve. For AMVs derived using visible channels, only the pixels with a successfully extracted CLA-CTH value on the “warm branch” of the plot of the Fig. 2 (reflectance larger than the average reflectance within the target area), and which have a CCij larger than the average CCij, 〈CCij〉, are selected to calculate the pressure.

3. Operational implementation at EUMETSAT

a. Main impact of CCC change on AMV characteristics

The performance of the CCC method has been compared against the performance of the former operational AMV extraction algorithm used in the EUMETSAT MPEF. During the month of June 2012, a special running configuration was set in the EUMETSAT MPEF to allow a direct comparison between the two algorithms. A month of data from Meteosat-9 was collected. The distribution of the AMV pressure retrieved for all of the channels is shown in Fig. 6. Using similar pixels for the tracking and the height identification clearly leads to a vertical redistribution of AMV pressure from high levels to midlevels in the Vis0.8 and IR10.8 channels. The AMV pressure distribution for the two configurations is also slightly different at low levels of the atmosphere, the CCC method setting the AMVs slightly higher in the atmosphere than the former algorithm. The difference mainly appears in areas where an atmospheric temperature inversion is present. The CCC method uses the CLA-CTH product, which sets the low-level cloudy pixels to the bottom of the inversion temperature layer found in the forecast profiles. This postprocessing is done for every cloudy pixel with a cloud-top height below 650 hPa. In this process the cloud can also be corrected upward as well as downward in the atmosphere. With the former AMV algorithm, for a cloud below 650 hPa, such a correction was only made downward (in cases for which the final AMV pressure was found above the inversion layer). For cases in which the AMV pressure was found below the inversion layer, it was never corrected upward. In addition a cloud-base height (CBH) method (Le Marshall et al. 1993) was used in the former algorithm, further assigning artificially the AMV pressure downward. These different processes mainly account for the differences observed for low-level AMVs in the IR10.8 and Vis0.8 channels. In the water vapor channels, the AMV pressure distributions remain nearly the same. There are more AMVs at midlevel and fewer AMVs at the high level using the CCC method.

Fig. 6.
Fig. 6.

AMV pressure distribution for all the AMVs extracted in (top to bottom) channels Vis0.8, WV6.2, WV7.3, and IR10.8 (1–30 Jun 2012). The new CCC results are shown in black. The results of the former AMV algorithm are shown in gray.

Citation: Journal of Atmospheric and Oceanic Technology 31, 1; 10.1175/JTECH-D-13-00126.1

Figure 7 shows the number of AMVs and their average QI without a model first-guess check, extracted in the Vis0.8 channel 3 times a day during the test period, for the CCC method and the former AMV algorithm. Results are split between high levels (100–400 hPa), midlevels (400–700 hPa), and low levels (700–1000 hPa) of the atmosphere. Similarly, Figs. 8 and 9 show the number of AMVs and their associated average quality in the other two channels—WV6.2 (high and midlevels) and IR10.8, respectively. Similar trends can be observed for every channel. There is a higher average QI at high levels and midlevels using the CCC method, associated with a larger amount of AMVs at midlevels. The number of AMVs is slightly reduced at high levels in Vis0.8 and WV6.2 channels because of the vertical redistribution at midlevels. The number of AMVs is similar (Vis0.8 channel) or to some extent larger (IR10.8 channel) at low levels for the CCC method, while the average QI is slightly degraded. We have shown only the results that correspond to AMVs having a very high QI (larger than 0.85) because these are mostly the AMVs that will be assimilated by the NWP models. However, the trends described above remain the same when considering a smaller QI threshold; the positive impact of the CCC method at high and midlevels is especially noticeable.

Fig. 7.
Fig. 7.

Amount of AMVs extracted in Vis0.8 channel and average QI for the CCC method (dashed line) and former AMV algorithm (solid line) during the period 1–30 Jun 2012: (two top plots) high-level winds (100–400 hPa), (two middle plots) midlevel winds (400–700 hPa), and (two lowest plots) low-level winds (700–1000 hPa). Only AMVs with a QI larger than 85 are used.

Citation: Journal of Atmospheric and Oceanic Technology 31, 1; 10.1175/JTECH-D-13-00126.1

Fig. 8.
Fig. 8.

As in Fig. 6, but for the WV6.2 channel. Only high-level and midlevel results are presented.

Citation: Journal of Atmospheric and Oceanic Technology 31, 1; 10.1175/JTECH-D-13-00126.1

Fig. 9.
Fig. 9.

As in Fig. 6, but for the IR10.8 channel.

Citation: Journal of Atmospheric and Oceanic Technology 31, 1; 10.1175/JTECH-D-13-00126.1

b. Comparison against collocated radiosonde observations

Table 1 shows the statistics of AMVs extracted in all channels against radiosonde observations for the test period. The operational Coordination Group for Meteorological Satellites (CGMS) collocation criteria defined at the third International Wind Workshop in Ascona, Switzerland, have been used. These criteria, shown in Table 2, are commonly used by all AMV producers to monitor the quality of the AMV product.

Table 1.

Statistics of cloudy AMVs collocated with radiosonde observations. (Abbreviations are given as Reg for region, Algo for algorithm, GLO for global, and TRO for tropics.)

Table 1.
Table 2.

CGMS criteria used to collocate a satellite AMV observation with a radiosonde observation.

Table 2.

The number of available collocations depends directly on the amount of high-quality AMVs. The statistics have been done on a similar set of data for both configurations in Table 1. The differences in collocation numbers at high, mid-, and low levels between the new CCC method and the former operational method reflect the different vertical distribution of AMVs indicated in Fig. 6.

The statistics regarding the bias and RMS show a general positive or neutral overall impact for all channels using the CCC method, especially for high-level and midlevel winds. For low-level winds in the IR10.8 and Vis0.8 channels, the impact of the CCC method is slightly negative. This is directly linked to the problem identified in temperature inversion areas, which was discussed in section 3a.

4. Discussion

This paper presents the CCC method recently implemented operationally at EUMETSAT in September 2012 to derive AMV height. This method establishes a clear link between the tracking step and the height assignment step, selecting only the pixels that drive the tracking (correlation) process to estimate the AMV altitude. The altitude is derived directly from the CLA-CTH product and is not recomputed within the wind algorithm (in contrast to the formal AMV algorithm). This increases the consistency between geostationary products. Results of the operational tests show a general improvement of the AMV product quality in every channel. The amount of good (high quality) AMVs generally increases. The winds are redistributed within the atmosphere, leading to more midlevel vectors. The low-level AMVs extracted from the IR10.8 and Vis0.8 channels have been assigned slightly higher in the atmosphere. Using the CLA-CTH product to estimate the height implies a different approach for the inversion correction. Because the CLA-CTH product represents the top of the cloud, an inversion correction is applied to every cloudy pixel irrespective of its location within the inversion layer. The previous AMV algorithm did not apply any correction to cloudy pixels below the inversion layer. In addition to the inversion correction, the CBH method used an empirical equation that assigned downward the AMV altitude in order to set it to the level of the cloud base—that explains why the low-level AMVs for the CCC method are retrieved at a higher altitude.

However, recent studies suggest that the cloud-top height should not be the most appropriate parameter to set the AMV altitude of low-level winds. Hernandez-Carrascal and Bormann (2014) and Bormann et al. (2014) showed using simulated images that the AMV speed corresponds neither to the speed at the cloud top nor to the speed at the cloud base, but to a speed somewhere in the middle of the cloud. Because lower low-level AMVs seem better for assimilation in NWP models, a temporary solution has been introduced operationally at EUMETSAT to assign downward the low-level AMVs. As a postprocess, an inversion correction similar to the former algorithm is reapplied to every low-level wind found in an inversion area, meaning that winds found below the inversion layer stay below the inversion layer. However, the question of setting the AMV altitude at low levels is still open, and additional research and discussion between AMV producers and AMV users will be needed to improve the use of the AMV product in NWP models.

Use of the CCC method brings a high degree of flexibility to the AMV algorithm, in that it can easily accommodate cloud-top heights derived from other methods. For example, operational implementation of the optimal estimation cloud analysis (OCA) product (Watts et al. 2011) is being planned at EUMETSAT. Use of this new cloud product to set the altitude of the winds is expected to further improve the AMV product quality.

REFERENCES

  • Borde, R., , and Oyama R. , 2008: A direct link between feature tracking and height assignment of operational atmospheric motion vectors. Proc. Ninth Int. Winds Workshop, Annapolis, MD, EUMETSAT, 8 pp. [Available online at http://www.eumetsat.int/Home/Main/AboutEUMETSAT/Publications/ConferenceandWorkshopProceedings/2008/groups/cps/documents/document/pdf_conf_p51_s3_13_borde_v.pdf.]

  • Borde, R., , and Dubuisson P. , 2010: Sensitivity of atmospheric motion vectors height assignment methods to semitransparent cloud properties using simulated Meteosat-8 radiances. J. Appl. Meteor. Climatol., 49, 12051218.

    • Search Google Scholar
    • Export Citation
  • Bormann, N., , Hernandez-Carrascal A. , , Borde R. , , Lutz H.-J. , , Otkin J. A. , , and Wanzong S. , 2014: Atmospheric motion vectors from model simulations. Part I: Methods and characterization as single-level estimates of wind. J. Appl. Meteor. Climatol.,in press.

    • Search Google Scholar
    • Export Citation
  • Büche, G., , Karbstein H. , , Kummer A. , , and Fischer H. , 2006: Water vapor structure displacements from cloud-free Meteosat scenes and their interpretation for the wind field. J. Appl. Meteor. Climatol., 45, 556575.

    • Search Google Scholar
    • Export Citation
  • Chang, F.-L., , Minnis P. , , Lin B. , , Khaiyer M. M. , , Palikonda R. , , and Spangenberg D. A. , 2010: A modified method for inferring upper troposphere cloud top height using the GOES 12 imager 10.7 and 13 μm data. J. Geophys. Res., 115, D06208, doi:10.1029/2009JD012304.

    • Search Google Scholar
    • Export Citation
  • Cho, H.-J., , Sohn E.-H. , , and Ou M.-L. , 2008: The impact of target box size on wind speed biases in satellite-derived atmospheric motion vectors. Proc. 2008 EUMETSAT Meteorological Satellite Conf., Darmstadt, Germany, EUMETSAT, 5 pp. [Available online at www.eumetsat.int/groups/cps/documents/document/pdf_conf_p_s12_50_cho_p.pdf.]

  • Doutriaux-Boucher, M., , Forsythe M. , , Saunders R. , , and Francis P. , 2006: MSG wind height assignment problems. Met Office Forecasting Research Tech. Rep. 469, 30 pp.

  • EUMETSAT, 2011: MGS meteorological products extraction facility algorithm specification document. EUMETSAT Doc. EUM/MSG/SPE/022, 259 pp.

  • Hernandez-Carrascal, A., , and Bormann N. , 2014: Atmospheric motion vectors from model simulations. Part II: Interpretation as spatial and vertical averages of wind and role of clouds. J. Appl. Meteor. Climatol.,in press.

  • Holmlund, K., 1998: The utilization of statistical properties of satellite-derived atmospheric motion vectors to derive quality indicators. Wea. Forecasting, 13, 10931104.

    • Search Google Scholar
    • Export Citation
  • Holmlund, K., 2000: The atmospheric motion vector retrieval scheme for Meteosat Second Generation. Proc. Fifth Int. Winds Workshop, Lorne, Australia, EUMETSAT, 201–208.

  • Lean, P., , Migliorini S. , , and Kelly G. , 2012: Studying the relationship between synthetic NWP-derived AMVs and model winds. Proc. 11th Int. Winds Workshop, Auckland, New Zealand, EUMETSAT, 8 pp. [Available online at http://www.eumetsat.int/website/wcm/idc/idcplg?IdcService=GET_FILE&dDocName=PDF_CONF_P60_S5_04_LEAN_V&RevisionSelectionMethod=LatestReleased&Rendition=Web.]

  • Le Marshall, J., , Pescod N. , , Khaw A. , , and Allan G. , 1993: The real-time generation and application of cloud-drift winds in the Australian region. Aust. Meteor. Mag., 42, 89103.

    • Search Google Scholar
    • Export Citation
  • Menzel, W. P., , Smith W. L. , , and Stewart T. , 1983: Improved cloud motion wind vector and altitude assignment using VAS. J. Climate Appl. Meteor., 22, 377384.

    • Search Google Scholar
    • Export Citation
  • Nieman, N. J., , Schmetz J. , , and Menzel W. P. , 1993: A comparison of several techniques to assign heights to cloud tracers. J. Appl. Meteor., 32, 15591568.

    • Search Google Scholar
    • Export Citation
  • Roebeling, R., , Baum B. , , Bennartz R. , , Hamann U. , , Heidinger A. , , Thoss A. , , Walther A. , 2013: Evaluating and improving cloud parameter retrievals. Bull. Amer. Meteor. Soc.,94, ES41–ES44.

  • Schmetz, J., , and Nuret M. , 1989: Cloud motion wind estimates in Europe. ECMWF/EUMETSAT Workshop on the Use of Satellite Data in Operational Numerical Weather Prediction: 1989–1993, Vol. 2, ECMWF/EUMETSAT, 274283.

  • Schmetz, J., , Holmlund K. , , Hoffman J. , , Strauss B. , , Mason B. , , Gärtner V. , , Koch A. , , and van de Berg L. , 1993: Operational cloud motion winds from Meteosat infrared images. J. Appl. Meteor., 32, 12061225.

    • Search Google Scholar
    • Export Citation
  • Schmetz, J., , Pili P. , , Tjemkes S. , , Just D. , , Kerkmann J. , , Rota S. , , and Ratier A. , 2002: An introduction to Meteosat Second Generation (MSG). Bull. Amer. Meteor. Soc., 83, 977992.

    • Search Google Scholar
    • Export Citation
  • Schreiner, A. J., , Menzel W. P. , , Straka W. , , and Heidinger A. , 2012: Comparison of CO2 and H2O atmospheric motion vector height assignment techniques using the GOES-13 imager. Proc. 11th Int. Winds Workshop, Auckland, New Zealand, EUMETSAT, 6 pp. [Available online at http://www.eumetsat.int/website/wcm/idc/idcplg?IdcService=GET_FILE&dDocName=PDF_CONF_P60_S5_05_MENZEL_V&RevisionSelectionMethod=LatestReleased&Rendition=Web.]

  • Sèze, G., , Marchand S. , , Pelon J. , , and Borde R. , 2008: Comparison of AMV cloud top pressure derived from MSG with space based lidar observations (CALIPSO). Proc. Ninth Int. Winds Workshop, Annapolis, MD, EUMETSAT, 8 pp. [Available online at http://www.eumetsat.int/website/wcm/idc/idcplg?IdcService=GET_FILE&dDocName=PDF_CONF_P51_S3_15_BORDE_V&RevisionSelectionMethod=LatestReleased&Rendition=Web.]

  • Sohn, E., , and Borde R. , 2008: The impact of window size on AMV. Proc. Ninth Int. Winds Workshop, Annapolis, MD, EUMETSAT, 6 pp. [Available online at http://www.eumetsat.int/website/wcm/idc/idcplg?IdcService=GET_FILE&dDocName=PDF_CONF_P51_S4_18_SOHN_V&RevisionSelectionMethod=LatestReleased&Rendition=Web.]

  • Stubenrauch, C. J., , Eddounia F. , , and Sauvage L. , 2005: Cloud heights from TOVS Path-B: Evaluation using LITE observations and distributions of highest cloud layers. J. Geophys. Res., 110, D19203, doi:10.1029/2004JD005447.

    • Search Google Scholar
    • Export Citation
  • Szejwach, G., 1982: Determination of cirrus cloud temperature from infrared radiances: Application to Meteosat. J. Appl. Meteor., 21, 284293.

    • Search Google Scholar
    • Export Citation
  • Velden, C. S., , and Bedka K. M. , 2009: Identifying the uncertainty in determining satellite-derived atmospheric motion vector height attribution. J. Appl. Meteor. Climatol., 48, 450463.

    • Search Google Scholar
    • Export Citation
  • von Bremen, L., 2008: Using simulated satellite images to improve the characterization of atmospheric motion vectors (AMVs) and their errors for numerical weather prediction. NWP SAF Rep. NWPSAFEC-VS-015, Version 1.4, 41 pp. [Available from http://research.metoffice.gov.uk/research/interproj/nwpsaf/vs.html.]

  • Watts, P. D., , Bennartz R. , , and Fell F. , 2011: Retrieval of two‐layer cloud properties from multispectral observations using optimal estimation. J. Geophys. Res., 116, D16203, doi:10.1029/2011JD015883.

    • Search Google Scholar
    • Export Citation
Save