This study uses lidar observations from the polar-orbiting Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) satellite to correct operational atmospheric motion vector (AMV) pressure heights. This intends to reduce the height assignment error of AMVs for their use in data assimilation. Additionally, AMVs are treated as winds in a vertical layer as proposed by several recent studies. Corrected and uncorrected AMV winds are evaluated using short-term forecasts of the global forecasting system of the German Weather Service. First, a direct lidar-based height reassignment of AMVs with collocated CALIPSO observations is evaluated. Assigning AMV winds from Meteosat-10 to ~120-hPa-deep layers below the lidar cloud top reduces the vector root-mean-square (VRMS) differences of AMVs from Meteosat-10 by 8%–15%. However, such a direct reassignment can only be applied to collocated AMV–CALIPSO observations that compose a comparably small subset of all AMVs. Second, CALIPSO observations are used to derive statistical height bias correction functions for a general height correction of all operational AMVs from Meteosat-10. Such a height bias correction achieves on average about 50% of the reduction of VRMS differences of the direct height reassignment. Results for other satellites are more ambiguous but still encouraging. Given that such a height bias correction can be applied to all AMVs from a geostationary satellite, the method exhibits a promising approach for the assimilation of AMVs in numerical weather prediction models in the future.
Atmospheric motion vectors (AMVs) are retrievals of the atmospheric wind derived by tracking cloud and water vapor structures in successive satellite images. The displacement of these structures generally characterizes tropospheric motions, and thereby the horizontal wind speed and wind direction can be determined. By using imagery from geostationary and polar-orbiting satellites and also exploiting the possibility of combining images from different satellites (Lazzara et al. 2014; Hautecoeur et al. 2014), AMVs are globally available. AMVs provide wind information with a unique spatial and temporal coverage, especially over the oceans and in polar regions with traditionally rare in situ observations. Given that the current global observing system is heavily skewed toward mass–temperature observations, reliable wind observations in remote areas are an essential data source for global numerical weather prediction (NWP) models (Baker et al. 2014; Weissmann et al. 2012). AMVs are therefore assimilated routinely in all global NWP systems, and many studies have shown that this type of satellite data has a positive impact on the forecast skill of NWP models (Bormann and Thépaut 2004; Velden et al. 2005; Joo et al. 2013).
Although AMVs have proven to be an important source of wind information, some issues remain. One unsolved problem comprises spatially correlated errors up to horizontal distances of several hundred kilometers (Bormann et al. 2003). The main contributor to the AMV wind error is the height assignment, which can be particularly relevant when the wind varies strongly with height. Velden and Bedka (2009) estimated that 70% of the total AMV wind error arises from height assignment issues. A number of error sources contribute to this: Temperature and humidity model profiles that are used to retrieve the AMV height may contain errors that are often correlated horizontally, and multilayer clouds or semitransparent clouds pose a further challenge for the height assignment process. In addition, the assumption that clouds are passive tracers of the atmospheric wind is not always valid (Schmetz et al. 1993). Studies on AMV error characteristics and the improvement of the AMV height assignment are therefore an active field of research (e.g., Borde et al. 2014; Salonen et al. 2015; Bresky et al. 2012). In practice, these issues lead to a massive thinning of the originally dense AMV dataset for data assimilation. As an example, the German Weather Service [Deutsche Wetterdienst (DWD)] thins AMVs to a minimum horizontal distance of 240 km in their global NWP system.
AMVs are traditionally interpreted as single-level observations. However, recent research revealed that this assumption should be reconsidered. Several studies showed that AMVs rather represent winds in vertical layers instead of winds at discrete levels (Hernandez-Carrascal and Bormann 2014; Weissmann et al. 2013; Velden and Bedka 2009; Lean et al. 2015). Folger and Weissmann (2014) proposed to assign AMVs from Meteosat-9 and Meteosat-10 to vertical layers beneath lidar-derived cloud-top heights. The evaluation of this height reassignment using nearby operational radiosondes resulted in a significant reduction of AMV wind errors. The aim of the present paper is to further elaborate this concept and to overcome the limitations of spatially and temporally rare radiosonde observations by using model equivalents (O-B statistics; see, e.g., Cotton 2012) for the wind evaluation. As in Folger and Weissmann (2014), lidar observations from the polar-orbiting Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) satellite are used to correct AMV pressure heights. In the first part of the present study, a direct height reassignment of AMVs with collocated CALIPSO observations is evaluated using different layer depths and layer positions relative to the lidar cloud-top height and relative to the operationally assigned AMV height. In the second part of this study, height bias correction functions for a general height adjustment of operational AMVs are derived to proceed from an individual height reassignment to a larger scope of application. This approach allows us to use lidar information for the AMV height correction without the need of real-time lidar data.
2. Data and method
1) AMVs from geostationary satellites
The study is mainly focused on AMVs derived from images of the European geostationary satellite Meteosat-10, which is located at 0° longitude and covers Europe, Africa, and large parts of the Atlantic Ocean. The major instrument on board is the Spinning Enhanced Visible and Infrared Imager (SEVIRI). SEVIRI provides full-disk imagery every 15 min with a resolution of 3 km at nadir for all instrument channels, and 1 km for the high-resolution visible channel (Schmetz et al. 2002).
In addition to a detailed evaluation for Meteosat-10 AMVs, a brief assessment of results for AMVs from other geostationary satellites that are used routinely in global NWP models is provided. These are Meteosat-7 at 57°E, the Multifunctional Transport Satellite 2 (MTSAT-2) at 145°E, and the two Geostationary Operational Environmental Satellites (GOES) at 135°W (GOES-West) and 75°W (GOES-East). The geographical position of these satellites is shown in Fig. 1. Meteosat-10 belongs to the Meteosat Second Generation with in total 12 channels in the visible and infrared range. Meteosat-7 (Meteosat First Generation) is less sophisticated, with only 3 channels, and consequently considerably fewer AMVs are available. Meteosat AMVs are derived operationally by the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT). GOES satellites have six channels each, and GOES AMVs are provided by the National Environmental Satellite, Data, and Information Service (NESDIS) of the National Oceanic and Atmospheric Administration. MTSAT-2 was the operational Japanese geostationary satellite during the period evaluated in this study and had five channels; MTSAT-2 AMVs were provided by the Japan Meteorological Agency. As Meteosat-7 provides only a small AMV sample and results tend to be similar to Meteosat-10, only results for MTSAT-2 AMVs and for the two GOES satellites combined are shown in this paper.
AMVs are derived by using images from different channels. Generally, visible channels (VIS) are used in the lower troposphere below pressure heights of 700 hPa during daylight periods. AMVs from infrared channels (IR) are derived throughout the troposphere, whereas AMVs from water vapor channels (WV) are mainly found in upper levels above 600 hPa. The WV AMVs are derived by tracking cloud structures as well as water vapor gradients in clear sky conditions. As AMV pressure heights are compared with lidar cloud-top observations in this study, WV AMVs derived from water vapor structures are not considered. There are several AMV height assignment methods that are used operationally. The equivalent blackbody temperature (EBBT) method is the most common method for low-level opaque clouds, using brightness temperatures of IR satellite images to retrieve height levels for AMVs. For high-level AMVs, the CO2-slicing and the H2O-intercept method both utilize differences of two channels for the height assignment. More details on height assignment methods can be found in Di Michele et al. (2013) and Nieman et al. (1997). In September 2012, the EUMETSAT algorithm for the derivation of Meteosat AMVs changed to the cross-correlation-contribution (CCC) method (Borde et al. 2014). The pixels that contribute most to the tracking process are used to assign a representative pressure height for each AMV, therefore providing a more consistent height assignment. With the changeover to CCC, the resulting operational Meteosat AMV dataset during the study period unfortunately contains no information on the height assignment method applied for deriving individual AMVs.
2) Lidar observations from CALIPSO
The polar-orbiting satellite CALIPSO was launched in 2006 as part of the A-Train and flies in a sun-synchronous orbit in 705-km altitude, encircling Earth in about 100 min. The Cloud–Aerosol Lidar with Orthogonal Polarization (CALIOP) on board is currently the only spaceborne lidar and thus provides unique information on clouds and aerosols from space. This study uses the official level-2 cloud-layer product, which provides information on the lidar cloud-top height with high horizontal (1 km) resolution along track. The vertical resolution can be seen as a measure of the uncertainty of the observed cloud top and varies for different altitudes: At lower altitudes (from −0.5 to 8.2 km), the vertical resolution is 30 m and at higher altitudes (8.2–20.1 km) it is 60 m. The combined use of two wavelengths in the visible (532 nm) and IR (1064 nm) range allows us to retrieve additional information on the specific cloud scene, such as cloud phase or multilayer cloud situations. Additionally, the cloud aerosol distinguisher (CAD) is defined as a quality index that indicates the reliability of the retrieved lidar information, ranging from +100 (cloud observation) to −100 (aerosol observation). For more information on CALIPSO, see Winker et al. (2009, 2010) and Hunt et al. (2009).
3) Collocation of AMVs and lidar observations
To find suitable CALIPSO lidar cloud-top observations that are close to AMVs, different collocation criteria are applied that largely follow Folger and Weissmann (2014). Generally, the applied criteria are based on a trade-off between the amount of available AMV–CALIPSO pairs and the matching accuracy. Besides the horizontal and temporal displacement between CALIPSO lidar observations and the corresponding AMVs, it should also be kept in mind that CALIPSO overpasses certain areas at certain times of the day and the resulting comparison of CALIPSO cloud tops and AMVs therefore does not capture variations in height assignment biases due to the diurnal cycle.
For Meteosat AMVs, the maximum horizontal distance between AMVs and collocated CALIPSO lidar observations is set to 50 km, with a maximum time difference of 30 min. The median of all lidar cloud-top observations within this range is taken as a representative cloud top, which is then compared with the operational AMV pressure height. A threshold of at least 20 lidar observations within this range is applied. In addition, the root-mean-square difference between these lidar observations and their median must be smaller than 70 hPa. Multilayer cloud scenes as well as cloud observations with a CAD < 90 are discarded. The forecast-dependent AMV quality index (ranging from 0 to 100, with 100 indicating the best possible quality) must be greater than 50. However, situations may still arise where CALIPSO observes different clouds than the ones used for deriving AMVs because of the horizontal and temporal distance of AMV and CALIPSO lidar observation. In addition, CALIOP is capable of detecting faint cirrus cloud layers that cannot be observed by imaging instruments like SEVIRI. To mitigate these issues, only AMVs that are at most 100 hPa above and 200 hPa below the respective median cloud-top heights are considered. This asymmetric interval is chosen on the assumption that an AMV represents the atmospheric motion of a vertically extended cloud structure and is therefore located below the actual cloud top. All AMVs beyond this range are discarded.
AMVs from GOES or MTSAT-2 imagery are derived by different institutions using different processing algorithms and quality control procedures. Furthermore, the AMV height assignment is aided by different model fields and may therefore show different characteristics than Meteosat-10 for lidar-based height correction methods. Based on sensitivity studies using different settings for the height reassignment and height bias correction, the collocation criteria for GOES and MTSAT-2 AMVs are tightened relative to the ones used for Meteosat-10. First, the AMV quality index threshold is raised from 50 to 80. In addition, AMVs are only used if the CALIPSO flight path approaches the AMV position to less than 10 km, which corresponds to an average distance of ~25 km between the AMV and the available lidar cloud-top observations within the 50-km radius that is used for the calculation of the median lidar cloud-top height.
4) Evaluation periods
Two different periods are evaluated in this study. For the evaluation of the direct height reassignment in section 3a, the study period comprises 11 days (31 May–10 June 2013; first evaluation period) of operational AMVs. For the height bias correction in section 3b, a 6-day time interval is used that ranges from 7 to 12 May 2013 (second evaluation period). Figure 1 shows the position of AMVs with collocated CALIPSO lidar observations during the second evaluation period. The corresponding distribution during the first evaluation period is very similar (not shown). The slightly different time frame of the two evaluation periods is chosen because the height bias correction in section 3b requires continuous (or nearly continuous) CALIPSO lidar observations in the preceding 30-day time interval of the respective evaluation period. This criterion is not fulfilled for the first evaluation period because of significant gaps in the CALIPSO dataset.
The study uses high-level AMVs above a pressure height of 400 hPa and low-level AMVs below a pressure height of 700 hPa. As there are traditionally few midlevel AMVs, AMVs between 400 and 700 hPa are not evaluated because of an insufficient sample size. Figure 2 shows the vertical distribution of AMVs for both evaluation periods for Meteosat-10. Altogether, 13 200 AMVs in the first evaluation period and 7410 AMVs in the second evaluation period are analyzed, with about 70% of them located in low-level and 30% in high-level regions. For the other geostationary satellites, there are considerably fewer available AMVs. Table 1 lists the numbers of all used AMVs with collocated lidar observations for both periods. Only about 10% of the number of Meteosat-10 AMVs is found for MTSAT-2 and GOES AMVs. This is due to the smaller number of operationally available AMVs from these satellites [GOES AMVs were only available in 3-h intervals during the study period as compared with hourly Meteosat Second Generation (MSG) or MTSAT-2 AMVs], as well as due to the stricter collocation criteria [higher quality indicator (QI) threshold and smaller horizontal distance] applied for the height reassignment and height bias correction for AMVs from these satellites.
1) Layer-averaged GME model winds
Several studies have shown that AMVs should be interpreted as layer averages instead of winds at discrete levels. Satellite sensors detect radiation from finite vertical layers, and additionally, the motion of clouds represents a vertically averaged wind over a cloud layer rather than the wind at the cloud top (Hernandez-Carrascal and Bormann 2014). Folger and Weissmann (2014) used collocated operational radiosondes to show that layers relative to the lidar cloud top yield a reduction of vector root-mean-square (VRMS) errors of 12% (17%) relative to layers (levels) centered at the operational AMV height.
In situ wind observations by radiosondes have small errors and therefore provide an ideal data source for wind evaluation. However, the inhomogeneous spatial and temporal distribution of radiosondes complicates their use for evaluation purposes. The additional need of verification radiosondes massively limits the sample size of collocated AMVs and lidar observations to about one percent of the original number. To overcome this limitation, model equivalents from the “GME” global model (Majewski et al. 2002) are used in this study for the wind evaluation. GME was the operational global forecasting system of DWD until July 2015, with a horizontal grid spacing of 20 km and 60 pressure levels in the vertical. The GME data assimilation system is based on the three-dimensional variational technique (3DVAR) with a 3-h cycling (0000, 0300, …, 1800, 2100 UTC). Model fields can deviate from the true atmospheric state because of errors in the forecast model. However, model errors and AMV errors can be assumed to be uncorrelated. Based on this assumption, differences between observation and model winds provide a good data source for the evaluation of observations that is available for every observation and therefore commonly used in the context of data assimilation (see, e.g., Cotton 2012). Direct comparisons of results from the model evaluation presented in this paper with results from a radiosonde evaluation of a similar approach of a lidar-based height correction presented in Folger and Weissmann (2014) in section 4 will substantiate this assumption.
An observation operator for layer-averaged AMVs was recently implemented at DWD. This operator provides AMV model equivalents derived from 3-h short-term [first guess (FG)] forecasts that are used for the wind evaluation in the present study. The calculation of the 3-h forecast is performed with the operational GME settings, which also includes the assimilation of single-level AMVs assigned at their original height in the preceding time steps. AMVs within the 3-h assimilation window (±1.5 h) are compared with the FG field at the corresponding time step, leading to maximum temporal difference between AMV and model equivalents of 90 min. The geographical position of each AMV is horizontally interpolated between the model grid points, and then the vertical layer averaging over the respective layer is applied. Layer averages are computed according to Simpson’s rule. The model wind is calculated as a weighted average of the interpolated wind values at the layer center, layer top, and layer base. The weighting coefficients of both the layer-top and the layer-base wind are ⅙1/6, whereas the wind at the layer center is weighted with 4/6.
2) Direct height reassignment
Based on the results of Folger and Weissmann (2014), layers of varying depth ranging from 0 to 200 hPa at three positions are evaluated: (i) below the lidar cloud-top height, (ii) with 25% above and 75% below the lidar cloud-top height, and (iii) centered at the operational AMV height. This is schematically illustrated in Fig. 3. A layer depth of 0 hPa denotes a discrete level that corresponds to the procedure applied operationally for the original AMV height. For the wind evaluation, two error metrics are applied for all considered layers: the VRMS difference and the wind speed bias. These are calculated as follows:
with and analogously. The N is the number of available AMVs with corresponding CALIPSO lidar cloud-top observations, oper denotes operational wind values, and model denotes the GME model equivalents of the respective layers. The VRMS difference defined above is sometimes also referred to as mean vector difference (MVD; Menzel 1996).
3) Height bias correction
The previously described AMV height reassignment is based on actual cloud-top heights of collocated lidar observations of the respective AMV. However, this method is only applicable to a small number of operational AMVs, as time and position of the AMVs have to coincide with nearby CALIPSO lidar observations. As an alternative approach, height bias correction functions are calculated for a general mean adjustment of all AMV heights from a respective satellite. For this purpose, the direct height reassignment is applied to all available AMVs within a certain time frame, and then an average over the resulting height adjustment values is computed. This height bias correction is then applied to a subsequent independent verification period. Durations of 30 days and 10 days are used as averaging time frames, and the resulting corrections are applied during the second evaluation period (7–12 May 2013). The 30-day mean comprises the days from 1 April to 6 May 2013 (with missing CALIPSO data on six days within this interval). The 10-day mean is calculated from the actual preceding days of the respective date. This means that for example the height bias correction derived from the period 2–11 May 2013 is applied and evaluated on 12 May 2013. As a third approach, the 30-day period is subdivided for different latitude bands to determine separate correction functions for the Northern Hemisphere (latitude larger than 25°N), the Southern Hemisphere (latitude larger than 25°S), and a tropical region in between. Table 2 lists the numbers of AMVs with collocated CALIPSO lidar observations that are used for the different height correction periods for Meteosat-10. For the other geostationary satellites (section 3c), only a 30-day average is calculated for the height bias correction because of the smaller number of available AMVs. The numbers of GOES and MTSAT-2 AMVs available for the 30-day training period are listed in Table 3. Height bias correction functions are calculated separately for the different channels (VIS, IR, and WV) for 50-hPa altitude bins between 950 and 200 hPa plus one additional bin for AMVs below and above this range, respectively. Every bin must contain at least 30 individual adjustment values to determine a valid mean adjustment for the respective altitude bin and AMV channel.
VRMS differences and wind speed bias values are calculated for all AMVs for (i) discrete operational levels, (ii) levels at 60 hPa below the actual lidar cloud-top observation, and (iii) adjusted levels based on the height bias correction functions. In addition, 120-hPa-deep layer averages centered at these levels are considered. The level 60 hPa below the lidar cloud top is chosen as it represents the mean pressure of the 120-hPa layer.
The first two sections (3a and 3b) present results for lidar-based height correction methods for AMVs from the European geostationary satellite Meteosat-10. Section 3a comprises results of the direct height reassignment, and section 3b contains results of the height bias correction. Section 3c provides a brief evaluation of both methods for other geostationary satellites.
a. CALIPSO-based height reassignment
AMV winds are evaluated by assigning AMVs to different layers and levels relative to the original AMV height and relative to the lidar cloud-top height during the first evaluation period (31 May–10 June 2013). Figure 4 illustrates the distribution of height differences between operational AMV heights and cloud-top heights derived from collocated CALIPSO lidar observations for all used Meteosat-10 AMVs. More than 80% of all operationally assigned AMVs are located below the actual lidar cloud top, corresponding to positive height differences on the x axis. The highest number of AMVs occurs within the first 50 hPa below the lidar cloud top. A further subdivision into latitude bands reveals similar distributions for extratropical and tropical regions (not shown).
Figure 5 shows the VRMS difference (Fig. 5, top) and wind speed bias (Fig. 5, bottom) between operational AMV winds and layer-averaged model winds. Gray dashed lines represent layers that are centered at the operational AMV height, which serve as a reference for the reassigned layers relative to the lidar cloud-top height (black lines). High-level AMVs above a pressure height of 400 hPa comprise WV and IR AMVs, whereas low-level AMVs below a pressure height of 700 hPa consist mainly of VIS and IR AMVs. Dividing the AMVs dataset for different channels used for their derivation shows similar results for both high- and low-level AMVs and is therefore not shown.
For high-level AMVs (Fig. 5a), lowest VRMS differences are achieved for 120-hPa layers below the lidar cloud top, resulting in a relative VRMS reduction of about 10% when compared with reference layers of the same depth centered at the original AMV height (dashed line) and of about 15% when compared with the discrete operational AMV heights (dashed line at 0 hPa). The wind speed bias tends to be close to zero for 100-hPa layers below the lidar cloud top. Low-level AMVs (Fig. 5b) show lowest VRMS differences for 120-hPa layers below the lidar cloud top and for 200-hPa layers with 25% above and 75% below the lidar cloud top. For 120-hPa layers below the lidar cloud top, the reassignment reduces the VRMS difference by 8% and 15% relative to reference layers centered at the operational AMV height and relative to the discrete operational height, respectively. The wind speed bias is generally small for low-level AMVs, but layers below the lidar cloud top exhibit slightly smaller values than the 25%–75% layers.
To investigate the effect for different latitude bands, the AMV sample used in Fig. 5 is subdivided into extratropical and tropical regions in Fig. 6. Generally, 120-hPa layers below the lidar cloud top achieve lowest VRMS differences for high-level and low-level AMVs in both regions (Figs. 6a–d, top). For high-level AMVs in extratropical regions (Fig. 6a, bottom), the wind speed bias is close to zero for 120-hPa layers below the lidar cloud top and thus coincides with lowest VRMS differences. The wind speed bias in the tropics shows larger values for the 120-hPa layer (Fig. 6b, bottom) but still has about the same magnitude as the wind speed bias for layers of the same depth at the operational AMV height. As GME has some known shortcomings in high-level tropical regions because of a relatively poor convection scheme, the tropical wind speed bias may also result from model error in that region. Wind speed bias values for low-level AMVs (Figs. 6c,d, bottom) are generally small for extratropical and tropical regions.
b. Height bias correction
Folger and Weissmann (2014) and the previous section showed that a direct lidar-based height reassignment significantly reduces the VRMS difference and wind speed bias of AMVs. However, this direct reassignment can only be applied to a small fraction of the AMV dataset, where collocated CALIPSO observations are available for the correction. This section evaluates the potential of a general lidar-based height bias correction derived from a preceding 10–30-day period, with an additional hemispheric subdivision of the 30-day period. This approach has the advantage that the full operational AMV dataset for a respective satellite can be corrected.
The previous section revealed 120-hPa-deep layers below the lidar cloud top as an overall optimal configuration for the height reassignment. For this reason, layers of that position and depth are used as the basis for the height bias correction. Figure 7 shows the 30-day lidar-based height bias correction function for Meteosat-10 as a function of altitude and channel. The mean pressure of that optimal layer, meaning the discrete level at 60 hPa below the lidar cloud top, is used as height adjustment value on the x axis. Negative values indicate that the AMV is shifted downward in the atmosphere. Midlevel AMVs between 400 and 700 hPa are not used for the height bias correction given the comparably small AMV density in this range. On average, the adjustment of high-level AMVs is on the order of −20 hPa. Low-level AMVs are also shifted downward at most altitude levels. The largest adjustment of 60–80 hPa occurs for AMVs at 700–800-hPa altitude. Generally, the curves of height bias correction functions based on a 10-day mean as well as for the latitude subdivision (both not shown) tend to have a similar shape as the 30-day mean, with less pronounced height adjustment values for the tropics than for the extratropics.
Figure 8 shows the VRMS difference and the wind speed bias between AMV and model winds before and after applying the height bias correction to Meteosat-10 AMVs during the second evaluation period (7–12 May 2013). Black bars represent results for operational AMV heights, striped bars for the CALIPSO height reassignment with directly collocated lidar observations, and gray bars for the three different height bias correction functions. The left part of each panel shows the results when assigning AMVs to discrete levels, meaning the operational levels (black), levels at 60 hPa below the cloud top of a directly collocated lidar observation (striped), or three “adjusted” levels based on the height bias correction (gray). Accordingly, the right part represents layer-averaged values for 120-hPa-deep layers centered at the respective heights. Results for the direct height reassignment are generally similar to the results in section 3a, with slight variations due to the different evaluation periods considered in the two sections.
For high-level AMVs (Fig. 8a), lowest VRMS differences (Fig. 8a, top) are achieved for 120-hPa layers based on the CALIPSO-based height reassignment. However, the height bias correction reveals a distinct reduction of VRMS differences in comparison with levels/layers relative to the operational AMV height and achieves about 30%–50% of the reduction of the direct height reassignment, with no clear preference for a particular correction function. In addition, the wind speed bias (Fig. 8a, bottom) is clearly reduced for the height reassignment as well as for the height bias correction based on a 30-day mean and a 10-day mean when compared with the wind speed bias at the operational AMV height.
Low-level AMVs (Fig. 8b) exhibit a similar pattern as high-level AMVs. Again, the CALIPSO-based height reassignment shows best results when layer averages are used. In addition, layers relative to the adjusted heights based on the height bias correction show a clear reduction of VRMS differences relative to the operational values. In particular, the 10-day height bias correction exhibits VRMS differences that are almost equally low as those of the direct height reassignment. The wind speed bias for low-level AMVs is strongly reduced for the direct height reassignment as well as for the height bias correction, especially when a layer averaging is applied. Overall, VRMS differences are generally lower for layer averages than for discrete levels, which further emphasizes that AMVs represent the wind in a vertically extended layer.
The subdivision in extratropical and tropical regions for the height bias correction is investigated in Fig. 9 for both high- and low-level AMVs. For reasons of clarity, each subfigure shows only results for the operational, discrete AMV level (black bar on the left of each panel) and for 120-hPa-layer averages centered at the operational level and at lidar-corrected levels (right part of each panel), omitting results for the CALIPSO-based height reassignment to discrete levels. The general tendency seen in Fig. 8 is reflected in extratropical as well as tropical regions: The direct height reassignment reveals lowest VRMS differences and a slow wind speed bias in all subfigures (Figs. 9a–d, striped bars), but the height bias correction on average achieves about 50% of this reduction. As previously pointed out, high-level AMVs in the tropics tend to have a stronger wind speed bias than in the extratropics, which may be related to an insufficient representation of the atmospheric state by GME. Nevertheless, the application of both height correction methods in the tropics overall leads to a smaller bias than the operational value.
To investigate the effect of different layer depths and level positions relative to the lidar cloud-top height, Fig. 10 shows the relative reduction of VRMS differences of both lidar height correction methods (direct height reassignment and height bias correction) for discrete levels and finite layers when their results are compared directly with the results for the discrete operational AMV heights. The relative reduction of the VRMS difference is shown as a function of layer depth for all latitudes and for low- and high-level AMVs combined. Solid lines correspond to layers/levels of the direct height reassignment, whereas dashed lines represent layers/levels relative to the adjusted height based on a 30-day height bias correction. Overall, the best results are achieved using the direct height reassignment for 120-hPa layers below the lidar cloud top (solid black line) with a VRMS reduction of about 11% relative to the operational AMV heights. Again, this value deviates slightly from the reduction of VRMS differences found in section 3a (~15%), as two different evaluation periods are considered in the two sections. Using the height bias correction, the largest reduction of VRMS differences (~9%) is also achieved for 120-hPa layers (dashed black line), reaching about 80% of the VRMS reduction that is achieved with the direct height reassignment. The results for discrete levels below the lidar cloud top (solid gray line) are less distinct. The largest reduction of about 7% is achieved for levels at 50–60 hPa (drawn in the figure at 100–120 hPa) below the lidar cloud top. Results for the adjusted pressure heights that are based on the 30-day height bias correction are least pronounced (dashed gray line) and only show a slightly positive effect (3.5% reduction). According to this, applying a 30-day height bias correction leads to a 3-times-larger reduction when AMVs are interpreted as layer averages instead of assigning them to discrete levels at the mean pressure of these layers.
c. Other geostationary satellites
After demonstrating the benefit of the CALIPSO-based height reassignment as well as the height bias correction for Meteosat-10 AMVs, these height correction methods are now tested for other geostationary satellites. Results for GOES AMVs and MTSAT-2 AMVs are shown for the height reassignment applied during the first evaluation period (Fig. 12) and for the 30-day height bias correction applied during the second evaluation period (Fig. 13). Results for Meteosat-7 tend to be similar to Meteosat-10 and are therefore omitted. As described in section 2a(3), the collocation criteria for GOES AMVs and MTSAT-2 AMVs are slightly adjusted relative to the ones applied to Meteosat AMVs, using a stricter threshold for both the QI and the horizontal distance between AMV and CALIPSO lidar observation. While these stricter collocation criteria improved the results for GOES AMVs and MTSAT-2 AMVs, the results for the Meteosat-10 dataset were fairly independent of the applied criteria. This is illustrated in Fig. 11, which shows the reduction of VRMS differences when results for layers below the lidar cloud top are compared with results of layers of the same depth centered at the original AMV height for Meteosat-10 during the first evaluation period. Setting the QI threshold to 50 (black) and to 80 (light-gray shading) leads to a relatively constant reduction of VRMS differences for both cases by about 12%–14% when constraining the horizontal distance. A subdivision for tropical and extratropical regions also exhibits similar results (not shown). As the sample size gets comparably low for GOES and MTSAT-2 AMVs because of fewer available observations and the stricter collocation criteria, a Student’s t test for dependent samples is used in the following to evaluate the significance of the results.
The VRMS difference and wind speed bias for the direct height reassignment for GOES AMVs and MTSAT-2 AMVs are illustrated in Fig. 12. For GOES high-level AMVs (Fig. 12a), 100–120-hPa layers below the lidar cloud top show a small benefit over the operational AMV heights, with about 3% relative reduction of VRMS differences when compared with layers of the same depth centered at the operational AMV height (reduction not significant) and 9% when compared with discrete operational AMV heights (reduction significant at the 99% confidence level). The corresponding wind speed bias values are close to zero at layer depths of approximately 80 hPa. Low-level GOES AMVs (Fig. 12b) show a large reduction of VRMS differences for assigning AMVs to layers relative to the lidar cloud-top height for all layer depths, with minimum VRMS differences for 150-hPa layers. Here, the reduction reaches 22% (30%) when these layers are compared with layers (levels) at the operational AMV height (reduction in both cases significant at 99%). The corresponding wind speed bias is also clearly reduced relative to the operational values. High-level AMVs from MTSAT-2 (Fig. 12c) exhibit a similar pattern as high-level GOES AMVs, achieving lowest VRMS differences for 100-hPa-deep layers below the lidar cloud top. As for high-level GOES AMVs, the reduction for these layers is not significant when compared with layers of the same depth centered at the original AMV height, but is significant at the 99% confidence level when compared with the discrete operational AMV height. For low-level MTSAT-2 AMVs (Fig. 12d), no distinct improvement is found for assigning AMVs to layers relative to the lidar cloud-top height.
Figure 13 shows the VRMS difference and wind speed bias for the height bias correction for GOES and MTSAT-2 AMVs. Results for levels/layers based on a 30-day bias correction (gray bars) are compared with levels/layers relative to the operational AMV height (black bars) and relative to the lidar cloud-top height (striped bars). For high-level GOES AMVs (Fig. 13a), the VRMS differences of the direct lidar height reassignment as well as the height bias correction do not show advantages over the operational height. In addition, the wind speed bias deteriorates for both height correction methods. However, the sample size for high-level GOES AMVs is rather small. In addition, the NESDIS AMV algorithm applies a +8% speed bias adjustment for most of the high-level GOES AMVs. This automated speed bias correction complicates the use of an AMV height bias correction with lidar observations, as the adjusted wind speed does not correspond to the originally derived AMV height. In contrast to high-level GOES AMVs, low-level GOES AMVs (Fig. 13b) exhibit significantly lower VRMS differences with similarly positive results for the direct height reassignment and the 30-day height bias correction at the 99% confidence level. Again, layer averaging exhibits additional benefits when compared with using discrete levels. High-level AMVs from MTSAT-2 (Fig. 13c) show a similar pattern as Meteosat-10 AMVs in terms of VRMS differences. Lowest values are achieved for the direct height reassignment, but the height bias correction also leads to a reduction of VRMS differences relative to the operational values. This reduction is significant at the 99% confidence level when layer averages at the height-bias-corrected level are compared with the operational AMV levels or layers. In addition, the wind speed bias is slightly reduced for both lidar height correction methods. For low-level AMVs of MTSAT-2 (Fig. 13d), the height bias correction shows slightly lower VRMS differences than the operational AMV height, which are significant for layer averages on the 99% confidence level when compared with the operational AMV height. Wind speed bias values for both height correction methods are of similar magnitude as the operational bias.
4. Summary and discussion
In this paper, lidar observations from the polar-orbiting satellite CALIPSO are used to correct AMV pressure heights from different geostationary satellites, with a major focus on Meteosat-10.
In the first part, results for a direct height reassignment of AMVs with collocated CALIPSO lidar observations are presented. In contrast to Folger and Weissmann (2014), GME model equivalents are used for the wind evaluation instead of radiosondes. This allows us to analyze a considerably larger amount of AMVs. For Meteosat-10, both high-level and low-level AMVs exhibit lowest VRMS differences for assigning AMVs to 120-hPa-deep layers below the lidar cloud top. This leads to a reduction of VRMS differences of 8%–10% when compared with layers of the same depth centered at the operational AMV height, and about 15% when compared with the discrete operational AMV levels.
In the second part, statistical height bias correction functions are developed based on statistics of differences between AMVs and collocated lidar cloud-top heights. Different lengths of the training period and settings for deriving the height bias correction are tested and the resulting corrections are then applied to a subsequent evaluation period. This approach allows us to proceed from an individual, direct height reassignment for a small subset of AMVs to a height adjustment of all operational AMVs from the respective geostationary satellite. Generally, adjusting AMV pressure heights of Meteosat-10 according to a height bias correction derived from a 30- or 10-day height training period leads to lower VRMS differences and a lower wind speed bias relative to using the operational AMV heights. On average the reduction is about 50% of the reduction by the direct height reassignment but has the clear advantage that all AMVs can be corrected. A subdivision into tropical and extratropical regions leads to similar findings both for the evaluation of the height reassignment (section 3a) and the height bias correction (section 3b), supporting the robustness of the results presented in this study.
For other geostationary satellites, the positive effect of the direct height reassignment as well as the height bias correction is less distinct than for Meteosat-10 AMVs. However, both height correction methods overall indicate benefits in terms of VRMS differences and wind speed bias when compared with operational AMVs. For low-level GOES AMVs, results of the height reassignment as well as of the height bias correction indicate that lidar observations can reduce VRMS differences by up to 30%. This large reduction also reflects a well-known feature of low-level GOES AMVs in inversion regions, when AMVs are assigned too high in the atmosphere by NESDIS (Cotton 2012). In contrast, high-level GOES AMVs exhibit only small benefits for reducing VRMS differences, and even show degradation in terms of the wind speed bias. This may be related to a wind speed bias correction of +8% applied operationally by the data provider (NESDIS) to most GOES AMVs above a pressure height of 300 hPa. Consequently, there may be a need to either adapt the CALIPSO-based height bias correction for high-level GOES AMVs or to apply the height bias correction to GOES AMVs without the wind speed bias correction. However, it should also be kept in mind that the sample size for high-level GOES AMVs in this study is comparably small and that further studies are required to draw robust conclusions regarding the benefits of a lidar-based height correction of high-level GOES AMVs.
MTSAT-2 AMVs show both for high-level and low-level AMVs either neutral or slightly positive effects for both lidar height correction methods when compared with results at the operational AMV height.
The positive impact of assigning AMVs to layers instead of discrete levels shown in this study coincides with findings of preceding studies (Velden and Bedka 2009; Weissmann et al. 2013). In particular, results for the model evaluation conducted in the present study confirm the findings shown in Folger and Weissmann (2014), where a similar approach of the direct height reassignment was verified with radiosonde sounding data during an 8-month period. For radiosonde-evaluated upper-level AMVs that were mainly located over the European continent where radiosondes are frequently available, 120-hPa layers below the lidar cloud top yielded lowest VRMS differences and a wind speed bias close to zero. These previous results are very similar to the findings of the model evaluation presented in section 3a for extratropical regions. For low-level AMVs, radiosonde and model evaluation reveal different results. This is likely related to the evaluation period: The present study uses data after the changeover to the CCC method, which also switched off the application of the assignment of low-level AMVs to the estimated cloud base that is often used by AMV processing centers. The radiosonde verification in Folger and Weissmann (2014), in contrast, evaluated mostly AMVs before the CCC method was introduced. Overall, the consistency of the findings for the model evaluation presented in this study and the radiosonde verification presented in Folger and Weissmann (2014) implies that model error does not blur the results and emphasizes the validity of using short-term forecasts for the evaluation.
Using a simulated model framework, Hernandez-Carrascal and Bormann (2014) illustrated that AMVs represent winds averaged over a cloud layer instead of the cloud-top- or cloud-base-level wind. Lean et al. (2015) also quantified height assignment AMV error characteristics using a set of simulated AMVs and found the closest fit of AMVs to layer-averaged model winds that are most commonly located below the estimated cloud-top. This corresponds well to our results with lowest wind VRMS differences and wind speed bias values when assigning AMVs to layers below the lidar cloud-top height. Lean et al. (2015) and Hernandez-Carrascal and Bormann (2014) stated that a part of the benefit that is gained by assigning AMVs to layer averages over a cloud layer can be reached by using a discrete level positioned within the respective layer. This is also confirmed in the present study. However, the reduction of VRMS differences for assigning AMVs to a bias-corrected level tends to be relatively small in comparison with what can be gained by using a layer approach. Nevertheless, adjusted levels may be a compromise for current data assimilation systems, as this is easier to implement in NWP systems.
5. Conclusions and outlook
The present study clearly demonstrates the potential of incorporating lidar observations into AMV height assignment methods. Overall, the direct height reassignment for individual AMVs has proven to be a valuable approach for the reduction of the VRMS difference and the wind speed bias of operational AMVs. As this method requires collocated lidar observations for each AMV, its applicability is restricted to available spaceborne lidar observations in real time and therefore more complex to apply in operational data assimilation systems. The proposed height bias correction in the second part of this paper provides an alternative that does not require real-time data and would be easy to implement in an NWP system. Monthly or weekly updates of the height bias correction functions seem advisable to catch features due to seasonal variability or changes in the height assignment processing. However, the optimal update interval is still to be determined. Follow-on studies to evaluate the benefit of the lidar-based height bias correction for the forecast skill of NWP models are ongoing.
Spaceborne lidar observations provide reliable information on cloud-top heights that is independent from the AMV processing procedures and from model fields used for the processing. As model errors are usually correlated in space and time, the incorporation of independent lidar information in data assimilation systems is expected to also reduce horizontal error correlations as a positive ancillary effect. Generally, the application of height correction methods based on lidar information is not restricted to CALIPSO. Other spaceborne lidars are planned to be launched in the near future, for example, the Earth Clouds, Aerosols and Radiation Explorer (EarthCARE; Illingworth et al. 2015). Thus, the assimilation of AMVs as layer averages in combination with lidar information for the AMV height correction is seen as a promising approach to increase the benefit of AMVs for NWP. Furthermore, the height bias correction with independent lidar information may be useful to derive consistent datasets for climate research and to evaluate AMV height assignment methods.
The operational AMV data were provided by DWD. We are grateful to Harald Anlauf and Alexander Cress from DWD for the implementation of the observation operator for the treatment of AMVs as layer averages in GME and their ongoing support during the study. The CALIPSO data were obtained from the NASA Langley Research Center Atmospheric Science Data Center. The study was carried out in the Hans-Ertel Centre for Weather Research (Weissmann et al. 2014; Simmer et al. 2016). This German research network of universities, research institutes, and DWD is funded by the BMVI (Federal Ministry of Transport and Digital Infrastructure).