Atmospheric motion vectors (AMVs) provide valuable wind information for the initial conditions of numerical weather prediction models, but height-assignment issues and horizontal error correlations require a rigid thinning of the available AMVs in current data assimilation systems. The aim of this study is to investigate the feasibility of correcting the pressure heights of operational AMVs from the geostationary satellites Meteosat-9 and Meteosat-10 with cloud-top heights derived from lidar observations by the polar-orbiting Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) satellite. The study shows that the wind error of AMVs above 700 hPa is reduced by 12%–17% when AMV winds are assigned to 120-hPa-deep layers below the lidar cloud tops. This result demonstrates the potential of lidar cloud observations for the improvement of the AMV height assignment. In addition, the lidar correction reduces the “slow” bias of current upper-level AMVs and is expected to reduce the horizontal correlation of AMV errors.
Observations from various geostationary and polar-orbiting satellites are used to derive atmospheric motion vectors (AMVs) by tracking clouds or water vapor structures in consecutive satellite images. AMVs provide outstanding global wind field coverage, especially over oceans, where in situ wind observations are rare. Wind observations are particularly important for the initialization of numerical weather prediction (NWP) models (Baker et al. 2014), and therefore AMVs are an essential ingredient for NWP. The positive impact of AMV assimilation in NWP models has been shown in several studies (e.g., Bormann and Thépaut 2004; Velden et al. 2005). The vertical height assignment remains a challenging task, however, and introduces significant errors. These errors contribute up to 70% to the total AMV error (Velden and Bedka 2009) and can be horizontally correlated over several hundred kilometers (Bormann et al. 2003). Hence, AMVs are drastically thinned for the assimilation in NWP models and only a small fraction of the available observations is used.
Preceding studies (Velden and Bedka 2009; Weissmann et al. 2013) demonstrated that AMVs actually represent the wind in a vertically extended layer, although they are traditionally assimilated at discrete levels. In addition, Weissmann et al. (2013) showed that the height of AMVs can be corrected using airborne lidar cloud-top observations. The study presented here further investigates these two approaches that can potentially reduce the errors of AMVs. First, we treat AMVs as vertically extended layer observations instead of single-level observations. Second, satellite lidar cloud-top observations are used to correct AMV pressure heights. This paper is a follow-up study to Weissmann et al. (2013), in which a small, regional sample of airborne lidar observations was used as a test bed for AMV height correction with lidar cloud-top observations. As suggested in Weissmann et al. (2013), the current study conducts the transition to larger scales using a sample of satellite lidar observations with significantly larger size and longer temporal extent. Lidar cloud-top height observations from the polar-orbiting Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) satellite are used to correct the heights of Meteosat-9 and Meteosat-10 AMVs. A number of suitable vertical layers relative to the lidar cloud tops and relative to the original AMV heights are investigated. Furthermore, different depths of the vertical layers are tested to find an appropriate layer that should be assigned to AMVs in data assimilation systems. Operational collocated radiosondes are used to validate AMV winds before and after the height correction.
2. Data and method
The study comprises eight months (1 April–6 October 2012 and 16 April–13 June 2013) of operational AMVs that were derived hourly from the geostationary satellites Meteosat-9 (2012 period) and Meteosat-10 (2013 period) by the European Organization for the Exploitation of Meteorological Satellites (EUMETSAT). Both satellites are positioned at 0° longitude, and most of the height-corrected AMVs are located over Europe and Africa, where radiosondes are available for wind verification.
Meteosat AMVs from four different satellite channels are used: infrared observations (IR) at 10.8 μm, visible observations (VIS) at 0.8 μm, and observations from the water vapor channels (WV) at 6.2 and 7.3 μm. VIS-based AMVs can only be tracked during daylight and are derived for clouds in the lower troposphere, whereas IR-based AMVs occur throughout the troposphere and lower stratosphere. WV-based AMVs from the two water vapor channels are mainly positioned in the upper troposphere. The AMVs considered in this study are derived by tracking cloud structures, whereas WV AMVs tracking water vapor structures in cloud-free areas are excluded. The final AMV pressure height for Meteosat AMVs is determined by different height-assignment methods or techniques: IR window, IR/WV ratio, H2O intercept, and CO2 slicing [details can be found in Di Michele et al. (2013)]. On 5 September 2012, the EUMETSAT height-assignment algorithm changed to the cross-correlation contribution (CCC) method (Borde et al. 2014). This method provides a more consistent height assignment because the pixels that contribute most to the tracking process are used to set the AMV height. The information about the specific height-assignment method is no longer available in the final data product, however.
Corresponding lidar cloud observations were obtained by the polar-orbiting satellite CALIPSO that was launched in 2006 and flies at an inclination of 98.2° in a sun-synchronous orbit at 705-km altitude. CALIPSO is part of the A-Train, which is a constellation of several international science satellites that fly in formation and therefore facilitate a wide variety of different observations of the same scenery from space. The Cloud–Aerosol Lidar with Orthogonal Polarization (CALIOP) aboard CALIPSO measures vertical profiles of the atmospheric backscatter at two wavelengths (532 and 1064 nm), which enables determination of the cloud-top height with high horizontal and vertical resolution. Additional measurements of the depolarization at 532 nm allow determination of the cloud phase. In this study, the official CALIPSO level-2 cloud-layer product is used. It provides the 1-km horizontally averaged cloud-top height from CALIOP, the number of superimposed cloud layers, the cloud phase, and a quality index for clouds. The vertical resolution of CALIOP is 30 m at altitudes from −0.5 to 8.2 km and 60 m from 8.2 to 20.1 km [for more information on CALIPSO, see Winker et al. (2009, 2010) and Hunt et al. (2009)]. Missing CALIPSO observations on 27 days of the 8-month study period lead to 220 days of available data.
b. Collocation requirements
AMVs are corrected with nearby CALIPSO lidar observations that are within 50-km horizontal distance and 30-min time difference of the location and time of each AMV. The median value of all available (at least 20) individual CALIPSO cloud-top observations within this range is taken as the representative cloud top. In addition, the root-mean-square differences between single lidar cloud observations and their median value must not exceed 70 hPa. All multilayer cloud scenes are discarded. The EUMETSAT AMV quality index (QI) must be greater than 50, with 100 indicating the best possible value and 0 the worst value. The QI for CALIPSO observations ranges from −100 to 100 and distinguishes between aerosol (−100) and cloud layers (100). To ensure that the detected lidar signal definitely represents a cloud, this index has to exceed a value of 90.
In addition, the AMVs must be less than 100 hPa above and 200 hPa below the corresponding CALIPSO cloud-top height. This interval is chosen to account for the fact that the lidar observation and the AMV may “see” different clouds because of the temporal and/or horizontal displacement and is based on the assumption that AMVs represent the wind below the actual cloud top (Weissmann et al. 2013).
Figure 1 shows the position of Meteosat-9 AMVs and CALIPSO lidar observations on 1 April 2012 that match the described collocation requirements. For this day, we found 1247 collocated observations within the Meteosat-9 domain (approximately ±63° in each direction from 0° longitude and 0° latitude). There are typically around 1000–1300 Meteosat AMVs per day that could be corrected with CALIPSO observations. Altogether, 243 097 matches of Meteosat AMVs and CALIPSO lidar observations are found in the complete period of 220 days.
The AMV wind is evaluated using nearby operational radiosondes. These measurements will serve as a reference of the true state of the atmosphere in the following. Because the wind field is usually horizontally more uniform than cloud-top heights, the collocation criteria for nearby radiosondes are extended to 150 km and 90 min from the corresponding AMV. Thereby, both the original AMV pressure height and the lidar cloud-top height must be located at least 50 hPa below the highest level of the corresponding radiosonde. Given the comparably low number of operational radiosondes, the sample size reduces to 4478 matches of Meteosat AMVs, CALIPSO lidar observations, and operational radiosondes for the complete period.
The sample is divided into high-level AMVs with pressure heights < 300 hPa, midlevel AMVs with pressure heights between 300 and 700 hPa, and low-level AMVs with pressure heights ≥ 700 hPa. In total, 1259 high-level AMVs derived from the IR and WV channels (337 and 922 matches, respectively) are available. The respective CALIPSO observations are all classified as ice clouds. The midlevel dataset consists of 1576 AMVs (611 IR AMVs and 965 WV AMVs), and the corresponding CALIPSO cloud products comprise 67% ice clouds and 33% water clouds. The 1643 low-level AMVs from the IR and VIS channels (219 and 1424 matches, respectively) are expected to correspond to water clouds only. Figure 2 shows the vertical distribution of all AMVs that are used in this study.
c. Height-correction method
The applied AMV height correction with satellite lidar observations from CALIPSO follows Weissmann et al. (2013) and is illustrated in Fig. 3. AMV winds are compared with radiosonde winds that are vertically averaged over layers of varying depth (0–200 hPa): first for layers relative to the originally assigned AMV height and second for layers relative to the CALIPSO lidar cloud-top height. If a layer reaches the lowest or highest radiosonde level, the layer depth is reduced accordingly. Three different layer positions are considered: 1) layers centered at the corresponding AMV height or lidar cloud-top height, 2) layers with 25% above and 75% below the corresponding height, and 3) layers from the corresponding height downward.
Mean vector root-mean-square (VRMS) differences for all considered layers are calculated as
with dui = ui(AMV) − ui(layer_average), and accordingly for dυi. In this case, N corresponds to the number of collocated matches of AMVs, CALIPSO lidar observations, and radiosondes.
a. VRMS differences and wind speed bias
Figure 4 shows the mean VRMS differences of AMVs and radiosonde winds. VRMS values are calculated for assigning AMVs to vertical layers of increasing depth, which are computed by averaging radiosonde winds over the respective layer. The first set of layers uses the original AMV height as reference (gray lines); the second set uses lidar cloud-top observations as reference (black lines). The corresponding wind speed bias is shown in Fig. 5.
VRMS differences for high- and midlevel AMVs above 700 hPa (Figs. 4a–d) from WV and IR channels exhibit a distinct error reduction when AMVs are treated as vertically extended layers instead of as single-level observations (which are the values for 0 hPa on the x axis). Lowest VRMS differences are achieved either by layers below the lidar cloud tops or by layers with 25% above and 75% below the lidar cloud tops. The optimal depth of these layers varies from 120 to 200 hPa. Layers below the lidar cloud tops exhibit their lowest VRMS differences for a depth of 100–150 hPa, and layers with 25/75% above/below the lidar cloud tops yield their best results for a depth of 150–200 hPa. Overall, the shape of the curves for these two lidar layers is fairly similar for the different subsets presented in Figs. 4a–d, and small differences in the position of the minimum may also be a result of the limited sample size of individual subsets instead of systematic differences between them. For all four subsets, the minimum of VRMS differences for layers relative to the lidar cloud top is in the range of 0.5–1.5 m s−1 lower than the lowest values reached with layers relative to the original AMV height.
Figures 5a and 5b exhibit a significant slow bias of high-level AMVs assigned to their original discrete height (values for 0 hPa on the x axis). Such a slow bias has also been found in other recent studies (e.g., Bresky et al. 2012). The bias is generally reduced when AMVs are assigned to deeper layers, and results indicate that assigning them, for example, to layers of 100–150 hPa below the lidar cloud tops can also largely remove the slow bias of current upper-level AMVs. Overall, the results presented in Fig. 5 show that layers leading to low VRMS differences tend to be similar to layers leading to a low wind speed bias.
In contrast to upper-level AMVs, low-level AMVs (Figs. 4e,f) are typically assigned to an estimated cloud-base height rather than to a level near the cloud top. Averaging over layers that are centered at the original AMV height shows a slight advantage over the discrete value with increasing layer depth. Poorer results are revealed for layers below the original AMV height, which might be due to the fact that these AMV heights are relative to the cloud base and a layer below the cloud base consequently does not represent the wind conditions of the tracked cloud correctly.
Slightly better results are obtained when lidar cloud-top information is incorporated, but the benefit is less distinct than for mid- and high-level AMVs. The 200-hPa layers with 25/75% above/below lidar cloud tops and 200-hPa layers below lidar cloud tops (for IR and VIS, respectively) lead to the lowest VRMS differences, but results for layers of the same depth centered at the original AMV heights are only 0.1–0.2 m s−1 higher. Because low-level AMVs are located at pressure heights greater than 700 hPa, the 200-hPa layers below the lidar cloud tops are mostly layers from the lidar cloud top to the lowest radiosonde level. The lower benefit of lidar cloud-top heights for the correction of low-level AMVs may result from the relation of low-level AMVs to cloud-base winds and the inability of satellite lidars to observe these cloud bases accurately.
High- and midlevel AMVs overall exhibit similar behavior, and therefore all AMVs above 700 hPa are combined in Fig. 6. The combination of high- and midlevel AMVs will be referred to as upper-level AMVs in the following. Results indicate that lowest VRMS differences in combination with lowest wind speed bias values are achieved for either 120–130-hPa layers below the lidar cloud tops or for 200-hPa layers with 25/75% above/below the lidar cloud tops.
b. Relative VRMS reduction for lidar layers and lidar levels
Figure 7 shows the relative reduction of VRMS differences when results for layers below the lidar cloud tops are compared with results of layers of the same depth centered at the original AMV heights (Fig. 7a) and results using the discrete original AMV heights (Fig. 7b). The shape of the curves in Figs. 7a and 7b is similar. For upper-level AMVs (black lines), best results are yielded for layer depths of 100–120 hPa. Highest error reduction values are ~12% for lidar layers in comparison with layers centered at the original AMV heights (Fig. 7a) and are ~17% in comparison with the discrete original AMV heights (Fig. 7b). The improvement is apparent in both upper-level channels IR and WV (black dotted and dashed lines). Dividing between upper-level ice clouds and water clouds leads to a similar error reduction for both subsets and is therefore not shown. About 59.4% (64.6%) of the 2835 upper-level AMVs show reduced VRMS differences for the 120-hPa layers below the lidar cloud top in relation to the 120-hPa layers centered at the original AMV heights (to the discrete original AMV heights).
Correcting the height of low-level AMVs (gray lines) with lidar information only leads to a small error reduction, but the averaging over deep layers shows advantages over using discrete heights. The VRMS differences of 200-hPa layers below the lidar cloud tops are predominantly superior to the VRMS differences of 200-hPa layers centered at the original AMV heights and of the discrete original AMV heights (50.8% and 59.2%, respectively).
After demonstrating the benefit of assigning AMVs to vertical layers below lidar cloud tops, we now investigate how much of that error reduction could be achieved by assigning them to one representative discrete level relative to the lidar cloud top instead. The black solid line in Fig. 8 represents the treatment of AMVs as a layer average below the lidar cloud top (equivalent to the black solid line in Fig. 7b), whereas the dash–dotted line represents the assignment of AMVs to the discrete mean pressure height of that lidar layer (i.e., a discrete level located at one-half of the layer depth below the lidar cloud top). Results indicate that assigning AMVs to the mean pressure of the lidar layers achieves most of the reduction of assigning AMVs to vertically extended lidar layers. Interpreting AMVs as layer-averaged winds leads to a relative reduction that is ~3% higher, however. For both approaches, the maximum of the curves occurs at ~120 hPa, which corresponds to using discrete levels 60 hPa below the lidar cloud tops. The corresponding wind speed bias values at this maximum are close to zero for both approaches (not shown).
Figure 9 illustrates the distribution of differences between the original AMV pressure and the mean pressure level of 120-hPa-deep layers below the lidar cloud top for upper-level AMVs. About 75% of the AMVs are located above the mean pressure of the lidar layers and are thus shifted to lower altitudes (negative values) with the lidar height correction. Because AMVs are derived by tracking the motion of the cloud, the lidar cloud top (dashed line) marks the natural upper edge where AMVs should be located. Approximately 30% of the AMVs are located above the cloud, which may be related to an erroneous height assignment as well as to the temporal and horizontal displacement of AMV and CALIPSO lidar observations. On average, upper-level AMVs are located 31 hPa above the lidar layer center (and, correspondingly, 29 hPa below the lidar cloud top), with only small differences between the single channels WV and IR. In summary, this indicates that the operational processing of upper-level AMVs should consider that AMVs represent wind in a layer below the actual cloud tops, but the systematic height differences are likely dependent on the applied AMV processing systems and their settings.
c. Effects of using different subsamples
To investigate the effect of changes in the height-assignment algorithm of EUMETSAT, the analyzed 220 days are divided into three different time periods in Table 1. The first one comprises 142 days before 5 September 2012, the day on which the height-assignment algorithm was changed to the CCC method. The second period consists of 32 days starting on 5 September 2012, and the last period consists of 46 days from 16 April until 12 June 2013. According to the preceding results (see Fig. 7), the lidar-layer depth is set to 120 hPa for upper-level AMVs and 200 hPa for low-level AMVs. For upper-level AMVs, the error reduction for assigning layers below the lidar cloud tops instead of the discrete original AMV heights is apparent in all three periods, ranging from 11.4% to 18.9%. As stated before, low-level AMVs do not show a clear error reduction through the lidar height correction. One noticeable feature, however, is the high error reduction for low-level AMVs in the second period from 5 September to 6 October 2012. This feature is likely related to a temporary degradation of the quality of low-level AMVs in the time period after the height-assignment algorithm changed to the CCC method (Salonen and Bormann 2012).
To utilize a reasonably large sample size, the collocation criteria for AMVs and radiosondes in this study are set to 150 km and 90 min (see section 2b). The temporal and spatial displacements of AMVs and verification radiosondes introduce an additional error component that is expected to be independent of the AMV error itself and the height correction, however. Therefore, weak collocation criteria lead to an underestimation of the actual relative error reduction. Figure 10 shows how the relative error reduction for upper-level AMVs increases as the horizontal collocation criteria are tightened. As expected, the number of matches decreases for smaller distances. The error reduction for 120-hPa layers below the lidar cloud tops relative to layers centered at the originally assigned AMV heights shows a strong increase from ~12% at 150 km to ~21% at 40 km (black solid line). When compared with the discrete original AMV height, the relative error reduction increases from ~17% to ~25% (gray dashed line). Reducing the time difference does not lead to clearly larger improvements and is therefore not shown.
This study uses a threshold for the AMV QI of 50 (see section 2b). Restricting it to higher values (up to ≥80) reduces the sample size by up to ~60%. Table 2 lists the relative error reduction for assigning 120-hPa layers (upper-level AMVs) and 200-hPa layers (low-level AMVs) below the lidar cloud tops instead of the discrete original AMV heights for different quality thresholds. Restricting the sample to upper-level AMVs with QI ≥ 80 shows slightly less improvement than including lower-quality AMVs, but the differences are less than 2.5%. For low-level AMVs, the error reduction slightly increases when only AMVs with higher quality are regarded.
In this study, we use satellite lidar observations to correct the height of AMVs from Meteosat-9 and Meteosat-10 with lidar cloud-top observations from CALIPSO. Here, 220 days of data with altogether ~4500 collocated AMVs, CALIPSO observations, and radiosondes are analyzed. We investigate appropriate layer depths and layer positions relative to the lidar cloud tops and relative to the original AMV heights by comparing AMV winds with radiosonde winds averaged over layers of the respective depth and position.
For upper-level AMVs, we found that assigning 120-hPa layers below the lidar cloud tops led to an improvement of ~12% relative to assigning layers of the same depth centered at the original AMV heights and of ~17% relative to using the discrete original AMV heights. Similar results are found for 200-hPa layers with 25% of the layer above and 75% below the lidar cloud top. For AMVs above 700 hPa, the improvement is apparent in both channels and for both ice and water clouds.
The error reduction for AMVs below 700 hPa is less distinct when layers relative to the lidar cloud tops are used instead of layers relative to the originally assigned AMV heights. Although there is only a slight error reduction for these AMVs using lidar information, there is an indication that lidar observations can reduce AMV errors in periods with lower AMV quality due to changes in the AMV processing. The reasons why the lidar height correction is showing much better results for upper-level AMVs may result from the relation of low-level AMVs to cloud-base winds and the inability of satellite lidars to observe these cloud bases accurately.
A tighter threshold for the horizontal distance between AMVs and radiosondes used for verification leads to a clearly even larger effect of the lidar height correction. The results imply that the lidar height correction can actually reduce the AMV wind error by over 20% relative to assigning AMVs to layers that are relative to the original heights and over 25% relative to using the discrete original AMV heights, but the sample size gets smaller for a tight threshold.
Our results confirm the findings of preceding studies that AMVs are more representative of a vertically extended layer wind instead of the wind at a discrete level (Velden and Bedka 2009; Weissmann et al. 2013). Hernandez-Carrascal and Bormann (2014) showed in a simulated framework that AMVs represent the wind within the cloud instead of the wind at the cloud-top or cloud-base level. This is consistent with our finding that layers below the lidar cloud tops yield the best results. Alternatively, assigning AMVs to a level centered at the mean pressure of the lidar cloud layer achieves most of the benefit of assigning AMVs to layers below lidar cloud tops. This is also similar to the results of Hernandez-Carrascal and Bormann (2014), in which a discrete level at an adjusted pressure height can have effects that are similar to those of a layer-averaged wind.
In summary, the results of this study demonstrate that the errors of Meteosat AMVs above 700 hPa can be significantly reduced when information from lidar cloud-top observations is incorporated. As already stated by other studies (Weissmann et al. 2013; Hernandez-Carrascal and Bormann 2014), the best layer depth and layer position relative to the original AMV height likely depend on the AMV processing and therefore vary from one dataset to another. Lidars, in contrast, provide high-resolution cloud-top observations that are expected to be independent of the height-assignment method used in the AMV processing. This implies that the horizontal correlation of AMV errors can also be reduced.
This study uses a sample size of ~4500 collocated AMVs, CALIPSO observations, and radiosondes. The strongest restriction, however, involves the availability of radiosondes for verification that are not required for the lidar height correction itself. Per day, there are about 1000–1300 Meteosat AMVs with nearby CALIPSO observations that could be directly corrected with lidar information. About 3300 operational Meteosat-10 AMVs are assimilated every 6 h in the current global forecasting model of the Deutscher Wetterdienst (DWD), leading to ~13 200 assimilated Meteosat-10 AMVs per day. Accordingly, the number of assimilated Meteosat-10 AMVs could be increased by 8%–10% when the additional lidar-corrected AMVs are included. Furthermore, it is assumed that the lidar height correction can also reduce the errors of AMVs from other geostationary and polar-orbiting satellites. Altogether, this could provide a considerable number of additional higher-quality observations with errors that can be expected to be significantly less correlated with other AMVs.
Assimilating such lidar-corrected observations requires a forward operator for treating AMVs as vertical layers, an adjustment of the assigned error, and some technical modifications (e.g., for thinning and quality-control procedures). The layer operator has recently been implemented in the DWD system, and follow-on studies that assimilate lidar-corrected AMVs are planned.
Our study demonstrates the potential of using lidar cloud observations from CALIPSO or other future space-based lidars for the height correction of AMVs. It suggests that NWP may benefit from assimilating lidar-corrected AMVs and treating them as layer-averaged AMVs in the future. Even larger benefits for NWP may be achievable by using the lidar information to develop situation-dependent quality-control functions. Alternatively, lidar-derived heights for AMVs could be used to validate different AMV processing algorithms.
The operational AMV and radiosonde data were provided by the Deutscher Wetterdienst. We thank Alexander Cress and Harald Anlauf from DWD for helpful comments and for their support with the data acquisition. The CALIPSO data were obtained from the NASA Langley Research Center Atmospheric Science Data Center. We also thank Franziska Schnell (LMU Munich) for her help with data access and Régis Borde (EUMETSAT) for information on AMV height assignment. The study was carried out in the Hans Ertel Centre for Weather Research. This German research network of universities, research institutes, and DWD is funded by the BMVI (Federal Ministry of Transport and Digital Infrastructure).