Abstract

Uncertainties in the height assignment of atmospheric motion vectors (AMVs) are the main contributor to the total AMV wind error, and these uncertainties introduce errors that can be horizontally correlated over several hundred kilometers. As a consequence, only a small fraction of the available AMVs are currently used in numerical weather prediction systems. For this reason, alternative approaches for the height assignment of AMVs are investigated in this study: 1) using collocated airborne lidar observations and 2) treating AMVs as layer winds instead of winds at a discrete level. Airborne lidar observations from a field campaign in the western North Pacific Ocean region are used to demonstrate the potential of improving AMV heights in an experimental framework. On average, AMV wind errors are reduced by 10%–15% when AMV winds are assigned to a 100–150-hPa-deep layer beneath the cloud top derived from nearby lidar observations. In addition, the lidar–AMV height correction is expected to reduce the correlation of AMV errors as lidars provide independent cloud height information. This suggests that satellite lidars may be a valuable source of information for the AMV height assignment in the future. Furthermore, AMVs are compared with dropsonde and radiosonde winds averaged over vertical layers of different depth to investigate the optimal height assignment for AMVs in data assimilation. Consistent with previous studies, it is shown that AMV winds better match sounding winds vertically averaged over ~100 hPa than sounding winds at a discrete level. The comparison with deeper layers further reduces the RMS difference but introduces systematic differences of wind speeds.

1. Introduction

Atmospheric motion vectors (AMVs) derived by tracking the drift of cloud or water vapor features in satellite imagery are a key element of the global observing system for the initialization of numerical weather prediction (NWP) models. They particularly constrain the wind field in remote areas of the Southern Hemisphere and above the world's oceans where hardly any other wind observations exist. Several studies have documented the positive contribution of AMVs to the forecast skill of global NWP models (Bormann and Thépaut 2004; Velden et al. 2005; Gelaro et al. 2010). All major NWP centers now assimilate AMVs from several geostationary and polar-orbiting satellites that together provide nearly global coverage.

Despite improvements of the retrieval algorithms over the last decades, however, the height assignment of AMVs introduces significant errors. Velden and Bedka (2009, hereinafter VB2009) estimated that the height assignment is the dominant factor in AMV uncertainty and contributes up to 70% of the total error. In addition, those errors can be horizontally correlated over several hundred kilometers (Bormann et al. 2003). As a consequence, AMVs are usually thinned rigorously for the assimilation in NWP models.

Spaceborne lidars such as the one on the Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) satellite can accurately determine the height of cloud tops. Therefore, the combination of AMVs with cloud-top information from satellite lidars is seen as a promising approach to reduce the error and error correlation of AMVs. Di Michele et al. (2012) compared cloud-top heights derived from CALIPSO lidar data and AMV heights, but did not investigate the possibility of correcting AMV heights.

The present study intends to develop a height correction for AMVs based on lidar observations. Particular emphasis is given to investigating if the AMV should actually be assigned to the lidar-observed cloud top itself or to some layer around or beneath the lidar cloud-top observations. Instead of satellite lidar observations, the height correction is tested with airborne lidar observation during The Observing System Research and Predictability Experiment (THORPEX) Pacific Asian Regional Campaign (T-PARC) 2008 (Weissmann et al. 2011, 2012). The use of airborne observations has the advantage that more than 300 dropsondes are available to validate AMV winds before and after the correction. Observations of the lidar backscatter ratio at 1064 nm and dropsondes were performed during 24 research flights of the Falcon 20 research aircraft of the Deutsches Zentrum für Luft- und Raumfahrt (DLR).

In addition, the present paper investigates the appropriate layer depth and its vertical position for the assimilation of AMVs. The study of VB2009 indicates that AMVs represent the wind in a tropospheric layer rather than at a finite level. For this reason they calculated the vector root-mean-square (VRMS) difference between AMVs and radiosonde winds averaged over layers of different depth from the AMV height downward. The present paper additionally investigates the effect of layer averages on the wind speed bias as systematic errors are particularly crucial in data assimilation. Furthermore, we test different vertical positions of the averaging layer relative to the original AMV height. For this purpose we compare AMVs with sounding winds averaged over layers of different depth, and we shift these layers from above to beneath the AMV. Systematic wind speed differences are calculated in addition to VRMS differences as deeper layers lead to systematically weaker winds. The comparison is based on several thousand vertical soundings (dropsondes and special radiosondes from ships and small islands) during T-PARC.

The paper is outlined as follows: Section 2 describes T-PARC and the dataset. Section 3 first compares AMV heights with lidar cloud-top heights and then evaluates an AMV height correction using lidar observations. Section 4 compares AMVs with layer-averaged radiosonde and dropsonde winds, and section 5 discusses and summarizes the results.

2. Dataset

a. T-PARC observations

The summer component of the multinational T-PARC field campaign was conducted in August to October 2008 in the western North Pacific Ocean region. T-PARC and the associated projects Tropical Cyclone Structure 2008 (TCS-08) and Dropwindsonde Observations for Typhoon Surveillance near the Taiwan Region (DOTSTAR; Chou et al. 2011) aimed to investigate the genesis of tropical cyclones (TCs), to improve typhoon track and intensity forecasts by targeted observations, and to investigate the extratropical transition of TCs and their downstream impact in midlatitudes (for more information see http://www.eol.ucar.edu/projects/t-parc/). The main observational platforms were four research aircraft launching dropsondes: The German DLR Falcon 20, the U.S. Navy P-3, the U.S. Air Force WC-130, and the Taiwanese DOTSTAR Astra Jet. Altogether, over 500 flight hours were spent and over 1300 dropsondes were launched in a period from 1 August to 3 October 2008. In addition to dropsondes, additional radiosondes were launched from Japanese research vessels and from small islands. The right panel of Fig. 1 shows the location of radiosondes and dropsondes used for the comparison of AMVs with layer-averaged sounding winds in section 4.

Fig. 1.

(left) Location of airborne Falcon observations used in section 3; lidar observations are represented by gray lines and dropsondes by plus signs. (right) Location of soundings used in section 4; plus signs mark the location of dropsondes, and gray circles mark special T-PARC radiosondes from ships or small islands. The size of the circles representing sounding stations with more than 500 AMV matches used for the comparison is scaled linearly by the number of matches; the largest circle represents 1221 matches.

Fig. 1.

(left) Location of airborne Falcon observations used in section 3; lidar observations are represented by gray lines and dropsondes by plus signs. (right) Location of soundings used in section 4; plus signs mark the location of dropsondes, and gray circles mark special T-PARC radiosondes from ships or small islands. The size of the circles representing sounding stations with more than 500 AMV matches used for the comparison is scaled linearly by the number of matches; the largest circle represents 1221 matches.

The DLR Falcon was additionally equipped with a scanning wind lidar and a differential absorption lidar (DIAL) system for water vapor observations (Wirth et al. 2009; Harnisch et al. 2011). As a by-product of water vapor profiles, the DIAL system also observes vertical profiles of the backscatter ratio (BSR) beneath the aircraft at a wavelength of 1064 nm. These profiles can be used to accurately determine the height of cloud tops. After testing different approaches for deriving cloud-top heights, the maximum of the BSR gradient plus a threshold for the BSR gradient were used for cloud detection. There are many different approaches for deriving cloud tops from lidar observations, but in general it was found that the differences between cloud heights derived using different approaches or slightly modified thresholds are clearly smaller than the differences between lidar cloud-top heights and AMV heights [see Folger (2012) for details of the applied cloud detection method and differences of several approaches]. The average horizontal resolution of the lidar observations is about 2 km. As AMVs are derived from a significantly larger area, the median of all cloud-top heights derived from lidar observations within 100-km distance and 60-min time difference is used for the comparison with AMV heights and the correction of these heights. The DLR Falcon performed 24 research flights with dropsonde and lidar BSR observations during T-PARC (left panel in Fig. 1). About 50 flight hours with lidar BSR observations were available for the AMV height correction after quality screening.

The Cooperative Institute for Meteorological Satellite Studies (CIMSS) produced an experimental dataset for T-PARC consisting of hourly AMVs from images of the operational Japanese Multifunctional Transport Satellite 1R (MTSAT-1R) in four channels: (i) infrared (IR) observations at 10.8 μm; (ii) visible (VIS) observations at 0.73 μm for daytime low clouds; (iii) shortwave infrared (SWIR) observations at 3.75 μm for nighttime low clouds; (iv) observations in a channel sensitive to water vapor. The present study uses AMVs in the first three channels (IR, VIS, and SWIR) that track cloud features. The CIMSS algorithm for deriving AMVs is close to that used operationally by the National Environmental Satellite, Data, and Information Service of the National Oceanic and Atmospheric Administration. The size of the target box for deriving VIS and IR AMVs was 15 × 15 pixels (about 60 km), and the target box for SWIR AMVs was 7 × 7 pixels (about 28 km). AMV heights were determined using the H2O intercept method, the IR histogram method, and the cloud-base method [for more details, see Nieman et al. (1993) and Olander (2001)]. The first two methods estimate the cloud-top height (as the lidar observations) whereas the last one estimates the cloud-base height. On average, the cloud-base method was applied for about one quarter of AMVs beneath 600 hPa.

The mean flight level of the DLR Falcon was 11 km MSL, but lidar cloud observations within 150 hPa from the aircraft downward were not used to assure that no AMVs from clouds above the aircraft with a lower erroneous height assignment appear in the dataset. Because of this selection criterion and a minimum of the number of cloud AMVs in the middle troposphere, only AMVs beneath 500 hPa are used for the height comparison and height correction. As a consequence, all AMVs in the resulting dataset were derived from water clouds and none were from ice clouds. About 50% of these AMVs are VIS, about 30% are SWIR, and 20% are IR (Fig. 2). The low fraction of IR AMVs is due to the fact that those are mainly located at higher altitudes. SWIR AMVs are only derived during nighttime, whereas most flights were performed during daylight.

Fig. 2.

(a) Height distribution of AMVs used for the lidar–AMV height comparison in section 3a. (b) Height distribution of AMVs used for the height correction in section 3b.

Fig. 2.

(a) Height distribution of AMVs used for the lidar–AMV height comparison in section 3a. (b) Height distribution of AMVs used for the height correction in section 3b.

b. Selection criteria for observations

Observations for the lidar–AMV height comparison and correction in section 3 were selected with the condition that there are at least 10 lidar cloud observations within 60-min time difference and 100-km horizontal distance from an AMV. A representative cloud-top height is then derived from these lidar observations by taking the median of all individual lidar cloud-top observations. Whenever the rms of differences between the individual observations and the median exceeds 70 hPa, the observations are discarded. CIMSS also provides a quality indicator (QI) for AMVs that ranges from 0 to 100 with 100 indicating the highest quality. This QI must be at least 50 for observations used in section 3. These thresholds were chosen to exclude lidar observations and AMVs from different clouds without reducing the sample size too much.

In addition, the height comparison in section 3a applied the criterion that the AMV height and the lidar cloud top must be within 150 hPa vertically to discard values where AMVs and the lidar cloud signal originate from clouds at very different heights because of the temporal or horizontal displacement of the observations. Applying these criteria resulted in 656 AMVs with nearby lidar cloud-top observations heights for the comparison.

For the AMV height correction using lidar observations in section 3b, the pressure difference criterion was replaced by the criterion that the applied height correction is not more than 100 hPa; that is, the center of the layer to which the AMV is shifted must be within 100 hPa from the original AMV height. This criterion is based on the assumption that AMV height errors are usually less than about 100–150 hPa, on sensitivity studies with different limits, and on the visual comparison of lidar BSR cross sections and AMV heights. The evaluation of the lidar–AMV height correction was performed with wind observations from the two nearest dropsondes released by the DLR Falcon. Only one dropsonde was used for the first and last observations of a flight. For this evaluation, the additional criterion was applied that there is at least one dropsonde within 100 km and 60 min from the AMV and the lidar observations used for the correction. There were 369 available data matches that fulfilled the collocation criterion for dropsondes, AMVs, and lidar observations for the height correction. About one quarter of these were discarded by the pressure difference criterion.

The comparison of AMVs with layer-averaged radiosonde and dropsonde winds in section 4 applied the same thresholds for temporal and horizontal displacement that were used in section 3 (100 km and 60 min), but the threshold for the CIMSS QI was increased to 70 as the dataset was significantly larger and therefore allowed a more rigorous limit. Altogether 13 000 matches of AMVs and sounding winds are used in section 4.

3. Lidar–AMV height comparison and correction

a. Height comparison

The histogram of height differences between AMVs and representative lidar cloud tops is shown in Fig. 3. Representative lidar cloud tops for every AMV are derived by taking the median of all lidar cloud-top observations within 100 km and 60 min from the respective AMV as described in section 2. The distribution strongly depends on the AMV type. VIS AMVs are distributed throughout the vertical range of ±150 hPa from the lidar cloud-top height with more AMVs above than beneath the lidar cloud top. The majority of IR AMVs and nearly all SWIR AMVs in contrast are located beneath the lidar cloud tops. On average, the pressure of VIS AMVs is 21 hPa lower than at the lidar cloud tops, which means that VIS AMVs are on average located higher than the lidar cloud tops. The pressure of IR AMVs is 15 hPa higher than at the lidar cloud top and the pressure of SWIR AMVs is 53 hPa higher.

Fig. 3.

Histogram of height differences (hPa) between AMVs and lidar cloud-top heights. Positive values indicate AMV heights that are lower than lidar cloud-top heights.

Fig. 3.

Histogram of height differences (hPa) between AMVs and lidar cloud-top heights. Positive values indicate AMV heights that are lower than lidar cloud-top heights.

The important question is where AMVs should be located relative to the lidar cloud top as AMVs may represent the wind in a layer beneath the cloud top. In that case, the AMV height should actually be lower than the lidar cloud-top observations. For this reason, we computed the mean VRMS differences between the AMVs shown in Fig. 3 and layer-averaged dropsonde winds to aid the interpretation of systematic height differences. Once, the dropsonde wind is averaged over a layer starting at the AMV height and going downward by 50, 100, or 150 hPa (also referred to as comparing or assigning a layer beneath hereinafter) and second, the wind is averaged over a layer of the same depth centered at the original AMV height. In case such a layer extends beneath the ground level, the layer depth is reduced accordingly to end at ground level. Mean VRMS differences are calculated as the mean of the square root of the sum of the squared differences of both wind components. Figure 4 shows the relative reduction of mean VRMS differences when a layer beneath the AMV height is compared instead of a layer of the same depth centered at the AMV height.

Fig. 4.

Relative reduction of mean AMV–dropsonde VRMS difference when AMVs are compared with a layer beneath the original AMV height instead of a layer centered at the original AMV height. The left three bars represent results for 50-hPa-deep layers, the center three bars represent results for 100-hPa layers, and the right three bars represent results for 150-hPa layers.

Fig. 4.

Relative reduction of mean AMV–dropsonde VRMS difference when AMVs are compared with a layer beneath the original AMV height instead of a layer centered at the original AMV height. The left three bars represent results for 50-hPa-deep layers, the center three bars represent results for 100-hPa layers, and the right three bars represent results for 150-hPa layers.

For VIS AMVs, the mean dropsonde–AMV VRMS difference is 9%–15% lower when dropsonde winds are averaged over layers beneath the original AMV height than for averaging over layers of the same depth centered at the original AMV height (Fig. 4). The reduction is significant at the 99% confidence level for layer depths of 50, 100, and 150 hPa using a Student's t test for dependent samples. This indicates that VIS AMVs are clearly too high as the center of these layers is 25–75 hPa beneath the original AMV height.

The differences of SWIR AMVs and dropsondes are also reduced when dropsonde winds are averaged over layers of 100–150 hPa beneath the original AMV although those AMVs are already located 53 hPa beneath lidar cloud tops on average. The AMV–dropsonde differences for IR AMVs, in contrast, increase by nearly 10% for 150-hPa layers beneath the original AMV height. However, the reduction or increase of mean VRMS differences is not significant for both IR and SWIR AMVs. Further discussion of the optimal layer assignment relative to the original AMV height based on a much larger sample size is presented in section 4, whereas the purpose of Fig. 4 is only to aid the interpretation of differences between AMV and lidar cloud-top observations shown in Fig. 3.

Overall, we conclude that AMV heights should be located lower than lidar cloud tops. VIS AMVs are on average located 21 hPa above lidar cloud tops, but they appear to represent winds in a layer that is centered 50–75 hPa lower than their current height and therefore 29–54 hPa lower than lidar cloud tops. IR and SWIR AMVs are already located 15 and 53 hPa beneath lidar cloud tops, respectively.

b. Height correction

This section describes the correction of AMV heights with airborne lidar cloud-top observations and the evaluation of wind differences to dropsondes before and after the height correction. The correction shifts the AMV wind vertically to a layer relative to the median height of a nearby lidar cloud-top observations (see section 2 for the description of the dataset and selection criteria). Figure 5 shows the differences of AMV and dropsonde winds using different layer depths and three different layer positions relative to the lidar cloud-top observations for the correction. Mean VRMS differences generally decrease with increasing layer depth. The lowest mean VRMS difference is reached when AMVs are shifted to a layer from the lidar cloud top to 100–150 hPa beneath or, in other words, when AMV winds are compared with dropsonde winds averaged over a layer from the lidar cloud top to 100–150 hPa beneath. Assigning a 150-hPa layer that is 25% above and 75% beneath the lidar cloud tops leads to about the same result. The bias is less than 0.2 m s−1 for all layer depths and position that were tested. Results for the individual channels (VIS, IR, and SWIR, not shown) are similar and in general it seems best to assign AMVs to 100–150-hPa deep layers beneath the lidar cloud top or to 150-hPa layers with 25% above and 75% beneath the lidar cloud top.

Fig. 5.

Mean VRMS and wind speed bias of differences between AMVs (VIS, IR, and SWIR combined) and dropsondes when AMVs are assigned to a layer relative to nearby lidar cloud-top observations. The x axis denotes the depth of the assigned layer. The three different line types denote layers centered at the lidar cloud top (black dashed line), layers from the lidar cloud top downward (solid black line), and layers with 25% above and 75% beneath the lidar cloud top (solid gray line).

Fig. 5.

Mean VRMS and wind speed bias of differences between AMVs (VIS, IR, and SWIR combined) and dropsondes when AMVs are assigned to a layer relative to nearby lidar cloud-top observations. The x axis denotes the depth of the assigned layer. The three different line types denote layers centered at the lidar cloud top (black dashed line), layers from the lidar cloud top downward (solid black line), and layers with 25% above and 75% beneath the lidar cloud top (solid gray line).

Figure 6 shows the relative improvement, that is, relative reduction of mean VRMS differences between AMV and dropsonde winds for assigning AMV winds to a layer of 100 or 150 hPa beneath the lidar cloud-top observation instead of a layer of the same depth centered at their original height. On average, this height correction reduces VRMS differences by 14%. The average error reduction is largest for IR AMVs (14%–19%), followed by VIS AMVs (~16%) and SWIR AMVs (9%–11%). The reduction is statistically significant at the 99% confidence level for all AMVs and the subsets of VIS and SWIR AMVs. The reduction of the IR subset is significant at the 95% confidence level.

Fig. 6.

Relative reduction of mean VRMS differences between AMV and dropsonde winds when AMVs are assigned to a layer beneath the lidar cloud top instead of a layer centered at the original AMV height. The depth of the layer is 100 hPa for the left bars and 150 hPa for the right ones.

Fig. 6.

Relative reduction of mean VRMS differences between AMV and dropsonde winds when AMVs are assigned to a layer beneath the lidar cloud top instead of a layer centered at the original AMV height. The depth of the layer is 100 hPa for the left bars and 150 hPa for the right ones.

As shown in section 3a, the errors of VIS AMVs are also reduced by ~14% when a layer beneath the original AMV height is assigned. However, the optimal layer relative to the original AMV height likely depends on the AMV dataset and processing, whereas the lidar information is independent of the AMV processing. For SWIR and IR AMVs, the lidar correction leads to clearly lower errors than assigning a layer beneath the original AMV height. Further discussion on the potential of reducing AMV errors through assigning a layer relative to the original AMV height is provided in the following section.

4. Comparing AMVs with layer-averaged radiosonde and dropsonde winds

Section 3a suggests that the error of VIS AMVs is substantially reduced when they are assigned to a layer beneath their original height and section 3b demonstrates that the error of AMVs in all channels is reduced when a layer beneath lidar cloud-top observations is assigned to them. These results motivated a further investigation of how deep the atmospheric layer is that AMVs represent and how this layer should be positioned vertically if no additional lidar information is available. To increase the sample size, this section uses dropsondes from all four T-PARC aircraft and special T-PARC radiosondes. The use of radiosondes also allows us to compare AMVs at altitudes of 100–500 hPa, whereas the section 3 only uses AMVs beneath 500 hPa.

VB2009 already compared VRMS differences between AMV winds and layer-averaged sounding winds for layers of different depth from the AMV height downward. Their results suggest that the treatment of AMVs as layer-averaged winds beneath their original height in data assimilation could lead to a significant improvement. This section intends to complement the study of VB2009 by testing different positions of the layer relative to the original AMV height and by investigating the effect on the wind speed bias in addition.

Assigning a layer beneath the original AMV height results in a systematic height reduction and may therefore introduce systematic wind errors in case the height was correct or too low before. A larger averaging volume additionally leads to lower wind speeds. As data assimilation systems are particularly sensitive to systematic errors, we investigate the wind speed bias in addition to VRMS differences and we shift the averaging layers from 50 or 100 hPa above the AMV height to 100 hPa beneath. Whenever the averaging layer would extend beneath the ground, we use the lowest possible layer instead. The intention for shifting the layer is to find the optimal position of the layer relative to the AMV height and also to detect if the reduction of the difference is due to compensating systematic height errors by extending the layer to the correct height of the AMV wind or to the fact that AMVs really represent a layer wind.

The VRMS differences and the wind speed bias between AMV winds and layer-averaged sounding winds is shown in Fig. 7 for different layer depths (different line types) as a function of the vertical offset of the center of the averaging layer relative to the AMV height. The line type with the lowest minimum indicates the optimal (or appropriate) layer depth concerning mean VRMS or wind speed bias. The position of this minimum on the x axis indicates the optimal (or appropriate) position of the layer relative to the AMV height. The fact that the line with the lowest VRMS minimum is also the one with the lowest VRMS value at x = 0 hPa in all panels indicates that the compensation of systematic height errors is not the main effect for the VRMS reduction.

Fig. 7.

Mean VRMS and wind speed bias of differences between AMV winds and layer-averaged winds from dropsondes and radiosondes. The panel titles denote the AMV type [(a),(b) IR; (c),(d) SWIR; (e),(f) VIS], the height range of compared values (hPa), and the number of compared values. Different line types represent different layer depths for the vertical averaging of dropsonde and radiosonde winds: gray dashed line for 10 hPa, gray solid line for 50 hPa, black dashed line for 100 hPa, black dash–dotted line for 150 hPa [(a) only], and black solid line for 200 hPa [(b)–(f)]. Note that the scales for bias and mean VRMS values are different.

Fig. 7.

Mean VRMS and wind speed bias of differences between AMV winds and layer-averaged winds from dropsondes and radiosondes. The panel titles denote the AMV type [(a),(b) IR; (c),(d) SWIR; (e),(f) VIS], the height range of compared values (hPa), and the number of compared values. Different line types represent different layer depths for the vertical averaging of dropsonde and radiosonde winds: gray dashed line for 10 hPa, gray solid line for 50 hPa, black dashed line for 100 hPa, black dash–dotted line for 150 hPa [(a) only], and black solid line for 200 hPa [(b)–(f)]. Note that the scales for bias and mean VRMS values are different.

The results for IR AMVs above 499 hPa indicate that the lowest mean VRMS difference is reached when these AMV are assigned to a 100-hPa layer centered ~20 hPa beneath their original height. Deeper or shallower layers both lead to larger differences. However, a 50-hPa layer centered 16 hPa beneath the original AMV height may be the best choice in case a low wind speed bias is particularly important.

AMVs beneath 500 hPa (Figs. 7b–f) generally show a less distinct minimum of VRMS differences, presumably because of lower vertical wind gradients at lower levels. The lowest mean VRMS differences are generally reached for the deepest layer that is shown, that is, for 200 hPa. Averaging over these layers, however, leads to an increase of the wind speed bias that may not be desirable for data assimilation purposes. Thus, the choice of the optimal (appropriate) layer is to some extend a trade-off between mean VRMS and bias.

The mean VRMS difference is systematically reduced by 0.2–0.4 m s−1 when AMVs are assigned to a 100-hPa layer centered at their original height instead of a 10-hPa layer centered at that height (upper part in Table 1). Assigning such a layer also does not seem to lead to a significant increase of the wind speed bias. Table 1 also lists the mean VRMS and bias difference for one subjectively chosen optimal layer for every panel of Fig. 7 (lower part in Table 1). The 150- and 200-hPa layers were not selected because of the increase of the wind speed bias mentioned above. It is notable that these optimal layers are sometimes centered above and sometimes beneath the original AMV height in contrast to the assumption of VB2009 that layers beneath the AMV level are appropriate. The results in Fig. 7 suggest that a layer centered at the AMV height may be the best choice for all AMVs above 799 hPa. Only for VIS and SWIR AMVs beneath 800 hPa, there is an indication that assigning a 100-hPa layer beneath the original AMV height is more appropriate as both the mean VRMS and bias are reduced. For VIS AMVs beneath 800 hPa, such a layer beneath the original height even leads to lower VRMS differences than a layer at the original AMV height while the wind speed bias is about the same. For SWIR AMVs beneath 800 hPa, mean VRMS and bias for the layer beneath are comparable to the values for the layer centered at the original AMV height.

Table 1.

Reduction of mean VRMS and wind speed bias (both: m s−1) of differences between AMV and sounding winds when sounding winds are averaged over a layer described in the left column instead of a 10-hPa layer centered at the original AMV height. Negative values indicate lower values for the layer described in the left column. The upper part of the table presents the results for 100-hPa layers centered at the original AMV height, and the lower part presents the results for one other selected layer for every panel in Fig. 7.

Reduction of mean VRMS and wind speed bias (both: m s−1) of differences between AMV and sounding winds when sounding winds are averaged over a layer described in the left column instead of a 10-hPa layer centered at the original AMV height. Negative values indicate lower values for the layer described in the left column. The upper part of the table presents the results for 100-hPa layers centered at the original AMV height, and the lower part presents the results for one other selected layer for every panel in Fig. 7.
Reduction of mean VRMS and wind speed bias (both: m s−1) of differences between AMV and sounding winds when sounding winds are averaged over a layer described in the left column instead of a 10-hPa layer centered at the original AMV height. Negative values indicate lower values for the layer described in the left column. The upper part of the table presents the results for 100-hPa layers centered at the original AMV height, and the lower part presents the results for one other selected layer for every panel in Fig. 7.

5. Discussion and summary

This study develops a method for correcting lower-tropospheric AMV heights with airborne lidar cloud-top observations and investigates if AMVs rather represent winds in a layer instead of at a distinct level. The field campaign T-PARC in the western North Pacific offered a unique opportunity for such an investigation as it provided an experimental dataset of hourly MTSAT AMVs produced by CIMSS, airborne lidar observations from 24 flights of the DLR Falcon, more than 300 dropsondes from the same flights for an independent evaluation of the AMV height correction using lidar observations, and several thousand additional soundings for the comparison of AMVs to layer-averaged winds. The lidar–AMV height comparison and correction are limited to AMVs beneath 500 hPa because of the aircraft flight altitude, whereas the layer comparison also includes AMVs up to 100 hPa.

The first part of the study demonstrates that lower-tropospheric AMV heights can be corrected using airborne lidar observations. The best match of AMV and dropsonde winds is found when AMVs are assigned to 100- or 150-hPa-deep layers beneath the median of cloud-top heights derived from nearby lidar observations. On average, the wind error of AMVs is reduced by 10%–15% when AMVs are assigned to layers beneath the median of lidar cloud tops instead of layers centered at the original AMV height. The reduction is even larger when AMV–dropsonde differences at a discrete level are used as reference. Vertical layers with 75% beneath and 25% above the median of lidar cloud tops lead to similar results. The improvement is consistent for all channels (VIS, SWIR, and IR) and the results are statistically significant. For the correction, we use lidar observations within up to 100-km distance and 60-min time difference from the AMVs. A tighter criterion may be advisable for future applications, but in the current study this drastically limits the sample size and is therefore not shown.

The second part of the study compares AMV winds with layer-averaged radiosonde and dropsonde winds from T-PARC. Several layers of different depth are tested and these layers are shifted from above to beneath the AMV to investigate the depth of the layer that AMV winds represent and the appropriate position of such a layer relative to the original AMV height. It is found that the VRMS differences are reduced by 5%–10% when AMVs are assigned to a 100-hPa layer centered at their original height in comparison to a 10-hPa layer. Layer depths of 150–200 hPa lead to a slight further reduction of VRMS errors, but also tend to increase the bias of AMV–dropsonde wind speed differences. The optimal position of such a layer likely depends on the individual dataset and the processing algorithm.

Our findings generally confirm that AMVs represent winds in a tropospheric layer rather than at a discrete level as demonstrated by VB2009 and emphasize that AMVs should be assimilated as layer wind in NWP models. Depending on the AMV channel and the geographical region, VB2009 suggest 50–150 hPa as appropriate layer depth for VIS, IR, and SWIR AMVs, which is on average similar to our findings despite the different dataset and methodology. In addition to VB2009, we demonstrated that treating AMVs as layer winds has no negative effect on systematic errors unless the averaging layers are significantly thicker than 100 hPa.

In summary, we conclude that the errors of T-PARC AMVs could be reduced by about 5%–10% when AMV winds were assigned to 100-hPa layers centered at their original height. Further error reduction by 10%–15% could be reached when lower-tropospheric AMVs were assigned to 100- or 150-hPa layers beneath the median height of nearby lidar cloud-top observations and using 100-hPa layers centered at their original AMV height as reference.

The error reduction is demonstrated for a particular experimental AMV dataset processed by CIMSS in one particular geographical region. The dataset for the lidar height correction consists of 369 matches of AMV and lidar observations. Thus, results may be different for different regions, seasons, and datasets. Nevertheless, it is demonstrated that lidar information can be used to determine AMV heights with reasonable accuracy and such lidar observations constitute an independent and uncorrelated source of information. Thus, we think these results highlight the potential for reducing the errors and especially the error correlation of AMVs through the use of satellite lidar observations from CALIPSO or other spaceborne lidars in the future.

Acknowledgments

AMV, lidar, and dropsonde observations were made as part of T-PARC, TCS-08, and DOTSTAR. T-PARC and TCS-08 were sponsored by an international consortium from the United States (National Science Foundation, Office of Naval Research, Naval Research Laboratory, and U.S. Air Force), Germany (DLR and Forschungszentrum Karlsruhe), Japan (Japan Meteorological Agency), South Korea (National Institute of Meteorological Research), and Canada (Environment Canada). The role of the National Center for Atmospheric Research's Earth Observing Laboratory (NCAR EOL) for campaign and data management is acknowledged. DOTSTAR is funded by the National Science Council of Taiwan, the U.S. Office of Naval Research, and the Central Weather Bureau Taiwan.

Particular thanks are given to Oliver Reitebuch, Martin Wirth, and Christoph Kiemle from the lidar group of the DLR Institute for Atmospheric Physics who supported the analysis of lidar observations and to Chris Velden and Steve Wanzong at CIMSS for processing the AMVs and providing information on the dataset. The authors also acknowledge one anonymous reviewer, who made the important remark that AMVs should be compared with cloud heights derived from multiple lidar observations to avoid a mismatch of scales.

This study was carried out in the Hans Ertel Centre for Weather Research. This research network of universities, research institutes, and the Deutscher Wetterdienst is funded by the BMVBS (Federal Ministry of Transport, Building and Urban Development).

REFERENCES

REFERENCES
Bormann
,
N.
, and
J.-N.
Thépaut
,
2004
:
Impact of MODIS polar winds in ECMWF's 4DVAR data assimilation system
.
Mon. Wea. Rev.
,
132
,
929
940
.
Bormann
,
N.
,
S.
Saarinen
,
G.
Kelly
, and
J.-N.
Thépaut
,
2003
:
The spatial structure of observation errors in atmospheric motion vectors from geostationary satellite data
.
Mon. Wea. Rev.
,
131
,
706
718
.
Chou
,
K.-H.
,
C.-C.
Wu
,
P.-H.
Lin
,
S. D.
Aberson
,
M.
Weissmann
,
F.
Harnisch
, and
T.
Nakazawa
,
2011
:
The impact of dropwindsonde observations on typhoon track forecasts in DOTSTAR and T-PARC
.
Mon. Wea. Rev.
,
139
,
1728
1743
.
Di Michele
,
S.
,
T.
McNally
,
P.
Bauer
, and
I.
Genkova
,
2012
:
Quality assessment of cloud-top height estimates from satellite IR radiances using the CALIPSO lidar
.
IEEE Trans. Geosci. Remote Sens.
,
51
,
2454
2464
,
doi:10.1109/TGRS.2012.2210721
.
Folger
,
K.
,
2012
: Höhenkorrektur von Satelliten-Windvektoren (AMVs) mit flugzeuggetragenen Lidarmessungen. M.Sc. thesis, Meteorologisches Institut, Ludwig-Maximilians-Universität München, Germany, 81 pp. [Available online at http://www.meteo.physik.uni-muenchen.de/dokuwiki/lib/exe/fetch.php?media=lscraig:herz:masterarbeit_folger.pdf.]
Gelaro
,
R.
,
R. H.
Langland
,
S.
Pellerin
, and
R.
Todling
,
2010
:
The THORPEX observation impact intercomparison experiment
.
Mon. Wea. Rev.
,
138
,
4009
4025
.
Harnisch
,
F.
,
M.
Weismann
,
C.
Cardinali
, and
M.
Wirth
,
2011
:
Assimilation of DIAL water vapour observations in the ECMWF global model
.
Quart. J. Roy. Meteor. Soc.
,
137
,
1532
1546
.
Nieman
,
S. J.
,
J.
Schmetz
, and
W. P.
Menzel
,
1993
:
A comparison of several techniques to assign heights to cloud tracers
.
J. Appl. Meteor.
,
32
,
1559
1568
.
Olander
,
Z.
,
2001
: UW-CIMSS satellite-derived wind algorithm user's guide (version 1.n). CIMSS, 87 pp. [Available online at http://cimss.ssec.wisc.edu/iwwg/Docs/windcoug.pdf.]
Velden
,
C. S.
, and
K. M.
Bedka
,
2009
:
Identifying the uncertainty in determining satellite-derived atmospheric motion vector height attribution
.
J. Appl. Meteor. Climatol.
,
48
,
450
463
.
Velden
,
C. S.
, and
Coauthors
,
2005
:
Recent innovations in deriving tropospheric winds from meteorological satellites
.
Bull. Amer. Meteor. Soc.
,
86
,
205
223
.
Weissmann
,
M.
, and
Coauthors
,
2011
:
The influence of assimilating dropsonde data on typhoon track and midlatitude forecasts
.
Mon. Wea. Rev.
,
139
,
908
920
.
Weissmann
,
M.
,
R. H.
Langland
,
C.
Cardinali
,
P. M.
Pauley
, and
S.
Rahm
,
2012
:
Influence of airborne Doppler wind lidar profiles near Typhoon Sinlaku on ECMWF and NOGAPS forecasts
.
Quart. J. Roy. Meteor. Soc.
,
138
,
118
130
.
Wirth
,
M.
,
A.
Fix
,
P.
Mahnke
,
H.
Schwarzer
,
F.
Schrandt
, and
G.
Ehret
,
2009
:
The airborne multi-wavelength water vapor differential absorption lidar WALES: System design and performance
.
Appl. Phys. B: Lasers Opt.
,
96
,
201
213
.