• Open Access

Improvements to MISR stereo motion vectors


  • Ákos Horváth

    Corresponding author
    1. Leibniz Institute for Tropospheric Research, Leipzig, Germany
    • Corresponding author: Á. Horváth, Leibniz Institute for Tropospheric Research, Permoserstrasse 15, D-04318, Leipzig, Germany. (horvath@tropos.de)

    Search for more papers by this author


[1] This study evaluated the effects of recent major modifications to the Multiangle Imaging SpectroRadiometer (MISR) wind retrieval algorithm by repeating an earlier yearlong comparison with Meteosat-9 cloud motion vectors. Algorithm upgrades included an area-based stereo matcher, enhanced quality control, georegistration corrections to focal plane distortions, and increased retrieval resolution. The new winds had better quality at all levels, yielding a particularly large increase in coverage at middle and high levels. The upgrades had an overall neutral impact on the E-W wind comparison statistics, which were already quite good in the preceding MISR data set. The comparison statistics for the more error-prone N-S wind, however, improved significantly on all scales—global, zonal, and regional—and throughout the entire atmosphere. Both the negative N-S wind bias and root mean square difference decreased, and correlation increased substantially, with middle and high levels, and tropics and subtropics experiencing the largest improvements. Subpixel georegistration corrections reduced cross-swath variations in N-S wind and height by half. As the net effect of these improvements, the error characteristics of the previously more uncertain N-S wind component became comparable to those of the E-W wind component at low and middle levels. Despite the substantial reductions in N-S wind errors, which are highly correlated with stereo height errors, the MISR – Meteosat-9 height comparison did not generally improve, strongly suggesting Meteosat-9 height assignment errors as the primary driver of discrepancy.

1 Introduction

[2] The Multiangle Imaging SpectroRadiometer (MISR) measures reflected solar radiation in nine distinct directions: at nadir and oblique angles of 26°, 46°, 60°, and 70°, distributed along track, both forward and aft relative to flight direction of the Terra satellite. Cloud motion and associated height are derived simultaneously by tracking cloud patterns over a 3.5 min interval in the nadir, 46°, and 70° views, separately for the forward and aft camera triplets [Horváth and Davies, 2001a]. Compared to cloud motion vectors (CMVs) from geostationary or polar-orbiter imagers, MISR stereo motion vectors (SMVs) offer potentially more accurate heights thanks to the purely geometric retrieval technique, which is insensitive to radiometric calibration drift and requires no ancillary data; however, precise coregistration of the multiangle views is crucial [Moroney et al., 2002]. The CMV height assignment, on the other hand, interprets the cloud radiometric signature in the infrared window, CO2, or water vapor channels using radiative transfer calculations and forecast temperature profiles [Nieman et al., 1993]. As a result, CMV heights are prone to substantial errors in broken, semitransparent, or multilayer clouds, and in the case of a low-level temperature inversion [Garay et al., 2008].

[3] The first-generation SMVs were evaluated against limited sets of GOES-10 (Geostationary Operational Environmental Satellite) CMVs [Horváth and Davies, 2001b] and radar wind profiler data [Marchand et al., 2007]. As predicted during the design of the stereo algorithm, MISR winds were found to be more accurate in the east-west direction than in the north-south direction, because aliasing between cloud motion and height parallaxes in the latter complicates the retrieval. However, retrieval errors were often higher than the prelaunch RMS estimate of ~4 m s−1 for wind, and ~400 m for height. Subsequently, Davies et al. [2007] introduced upgrades to the original algorithm. Image coregistration was improved to subpixel accuracy for all cameras by matching sea-ice patterns and land-surface features. This enabled a reliable SMV estimate from the aft camera triplet as well, which had been previously plagued by excessive registration uncertainties in the 70° aft view. Subpixel parallax assessment and tighter quality control based on forward-aft SMV consistency were also implemented. As shown by comparisons to model winds [Davies et al., 2007], and an extended set of wind profiler observations [Hinkelman et al., 2009], these second-generation SMVs had significantly improved quality, broadly comparable to prelaunch estimates, but this came at the expense of greatly reduced coverage.

[4] The latest assessment by Lonitz and Horváth [2011] evaluated one year of second-generation SMVs against state-of-the-art Meteosat-9 CMVs. This study provided the most comprehensive error characterization of MISR winds, including their geographic and vertical error distribution, identified regions and cloud types where stereo heights were a definite improvement over CMV heights, but also discovered a systematic SMV bias across the MISR swath. The cross-swath bias was among several issues specifically addressed by the second major overhaul of the MISR wind retrieval algorithm incorporated into the Level 2 cloud product software. The upgraded cloud product is now available for the entire mission starting from 1 March 2000.

[5] The current paper describes the main algorithm modifications and their effects on SMVs. I used the data set of collocated SMVs and CMVs from Lonitz and Horváth [2011], but replaced the second-generation SMVs with the latest version, allowing a direct comparison of my results with those of Lonitz and Horváth [2011]. The upgrades to the stereo algorithm significantly improved not only SMV-CMV comparison statistics, but also the coverage of SMVs.

[6] The remainder of the paper is organized as follows: Section 2 describes the MISR upgrades and comparison methodology. Section 3 documents resulting improvements to MISR winds as compared to Lonitz and Horváth [2011]. Finally, I offer a summary and concluding remarks in section 4.

2 Stereo Algorithm Upgrades and Comparison Method

2.1 Stereo Algorithm Upgrades

[7] The principle of the MISR wind retrieval is described in Horváth and Davies [2001a]. Technical details of its practical implementation are given in Diner et al. [1999], while those of the latest version are provided by Mueller et al. [2013]. Here, I summarize only the most significant algorithm upgrades implemented since the study of Lonitz and Horváth [2011].

[8] Coarse feature-based stereo matcher has been replaced with sophisticated area-based technique. For cloud tracking, all previous versions of the MISR wind algorithm used the Nested Maxima (NM) feature matcher, which tracked local reflectance maxima [Muller et al., 2002]. This stereo matcher was fast but had sparse coverage with typically 1–2% of data being matched, its accuracy was only to within 2 pixels, and the distribution of returned disparities exhibited a long tail. These limitations necessitated use of a relatively large (70.4 × 70.4 km2) retrieval domain to obtain the ~100 samples required to map the disparity field reasonably well. In addition, Nested Maxima often failed in relatively featureless cirrus clouds, especially in the most oblique 70° views. The new algorithm uses a hierarchical sum-of-absolute-differences area-based matcher with much better accuracy and lower failure rate. A three-level pyramidal scheme is employed, sampling image pairs at successively finer resolutions of 1100, 550, and 275 m. This multiresolution approach improves run time by decreasing the dimensions of the search window at the down-sampled upper levels and using a reduced search window at the highest resolution bottom level; thus, eliminating the need to stereo match entire images at full 275 m resolution.

[9] Thanks to the vastly improved coverage of the hierarchical sum-of-absolute-differences matcher, the operational wind retrieval resolution has been increased from 70.4 to 17.6 km. As section 3.8 demonstrates, meaningful SMVs can be obtained even at 4.4 km resolution in high-texture cloud scenes.

[10] Analysis of measured disparities has also been improved. The previous algorithm used two separate 2-D (cross-track, along-track) disparity histograms, one each for the 70°–46° and 46°–nadir camera pairs. The new approach forms a single, joint, 4-D disparity histogram for both directions and camera pairs simultaneously, allowing a more accurate estimation of the dominant disparity signal from the most-populated histogram bin.

[11] Corrections for focal plane distortions have been implemented. Lonitz and Horváth [2011] noticed a cross-swath bias in SMVs, resulting in less accurate retrievals on the eastern side of the MISR swath than on the western. This cross-swath artifact was traced back to focal plane distortions unaccounted for in the MISR Camera Geometric Model used during Level 1 processing. The new algorithm includes subpixel georegistration corrections in Level 2 processing to mitigate cross-swath bias; the effect of these corrections on SMVs is discussed in section 3.7.

[12] The quality control scheme has been completely revamped. The quality of an individual SMV was characterized in the old scheme by the WindQuality (WQ) flag, which took integer values from 0 to 4 representing ever more stringent limits on the maximum allowed forward-aft differences in wind speed, height, and direction. In contrast, the Quality Indicator (QI) of the new MISR algorithm varies smoothly between 0 and 100, and is calculated more in line with the EUMETSAT (European Organization for the Exploitation of Meteorological Satellites) quality control scheme [Holmlund, 1998; Holmlund et al., 2001]. The new scheme keeps the forward-aft vector comparison, which amounts to a temporal consistency test, but also adds a spatial consistency test of neighboring SMVs. A further modification relaxes the requirement for successful wind retrievals in both forward and aft camera triplets. While the previous data set only reported the average of forward and aft vectors if both existed, the new data set also includes cases with a single (forward or aft) valid retrieval, provided this vector passes the spatial consistency test with respect to neighboring vectors in the opposite (aft or forward) direction.

[13] For the old SMVs, the product guidelines advised using retrievals with WQ = 3 (“good”) or WQ = 4 (“best”) only. For the new SMVs, the MISR team uses a threshold of QI ≥ 50 [Mueller et al., 2012], while geostationary wind monitoring guidelines recommend a higher threshold of QI ≥ 80, given a EUMETSAT-type quality control scheme [Forsythe and Saunders, 2008]. In section 3.2, I analyze the dependence of wind comparison on quality indices, which can help in selecting the adequate SMV quality threshold for a specific application until the mapping between the MISR and EUMETSAT QI schemes is better understood.

[14] A new flag has been introduced to filter out cloud-free “ground retrievals.” Unlike traditional CMV methods, the MISR wind algorithm does not apply a priori cloud target selection, and obtains retrievals in cloud-free land scenes as well. Such ground retrievals were not flagged in prior versions of the data set and, thus, had to be screened out by ad hoc methods in postprocessing [Hinkelman et al., 2009; Lonitz and Horváth, 2011]. Adopting the logic of previous investigators, the new stereo algorithm now flags near-surface, low-speed retrievals unlikely to be associated with advecting clouds. This flag, called MotionDerivedCloudMask, classifies retrievals as high/low-confidence cloud or high/low-confidence clear.

[15] Cross-track winds are now also reported at 1.1 km resolution, in addition to the 17.6 km standard product. It is well established that the accuracy of MISR cross-track winds is significantly better than that of along-track winds, the latter of which suffer from aliasing between cloud height and cloud motion parallaxes [Marchand et al., 2007; Hinkelman et al., 2009; Lonitz and Horváth, 2011]. Taking advantage of this increased accuracy, the new SMV data set now also provides cross-track wind components at 1.1 km resolution. Such fine scale component vectors can be used to investigate cloud dynamical processes, as demonstrated by Wu et al.’s [2010] study on the inner-core dynamics of tropical cyclones.

[16] The improved capabilities of the new wind algorithm are demonstrated in Figure 1 for an extra-tropical cyclone in the South Atlantic. Here, the 70.4 km SMVs were from the final version F08_0017 of the old TC_STEREO product, also used in Lonitz and Horváth [2011], while the 17.6 km SMVs were from the new and renamed TC_CLOUD product version F01_0001. (For consistency with Lonitz and Horváth [2011], MISR wind is still labeled in this paper as stereo motion vector or SMV, although the new MISR product uses the customary label cloud motion vector or CMV, similar to the Meteosat-9 product.) The narrow MISR swath only allowed a maximum of six 70.4 km SMVs across the track, and a large number of the retrievals were of low quality (see the shorter vectors in Figure 1a). In contrast, the improved SMVs clearly had better coverage, and revealed much finer details of the wind field.

Figure 1.

MISR winds for path 191, orbit 45232, blocks 115–118: (a) 70.4 km vF08_0017 SMVs and (b) 17.6 km vF01_0001 SMVs. (a) The longer vectors indicate WQ ≥ 3 (“good” or “best”) winds, while the shorter vectors indicate lower quality winds. (b) Winds with QI ≥ 50. The white square marks a 70.4 × 70.4 km2 vF08_0017 domain.

2.2 Comparison Method

[17] This study was based on the Lonitz and Horváth [2011] data set, comprising ~200,000 MISR – Meteosat-9 wind pairs from the year 2008. The original data set included ground-filtered MISR SMVs with WQ ≥ 3 from the TC_STEREO product version F08_0017, and Meteosat-9 CMVs with QI ≥ 80 [EUMETSAT, 2011], which were matched using spatial and temporal collocation criteria exceeding those recommended by the Coordination Group for Meteorological Satellites [Velden and Holmlund, 1998]. In the current analysis, I simply replaced the old 70.4 km SMVs with the new 17.6 km SMVs from the TC_CLOUD product version F01_0001, but retained the Meteosat-9 CMVs. The old MISR ground filter was also replaced, and here I only considered retrievals flagged by the improved ground filter as high-confidence cloudy.

[18] As shown in Figure 1, there were a maximum of 16 new SMVs for any given old SMV. When analyzing the evolution of quality indices, I replaced each old SMV by the highest-quality new SMV within the corresponding original wind domain. For the analysis of SMV-CMV comparison statistics, however, all 17.6 km SMVs within a 70.4 km domain and above a certain QI threshold (0, 50, or 80) were averaged. This averaging was carried out to reduce representativeness biases in the comparison, because Meteosat-9 winds correspond to a nominal area of 72 × 72 km2 at the subsatellite point, which is more comparable to the original MISR domain size.

3 Results

3.1 Multiangle Imaging SpectroRadiometer Ground Retrievals

[19] Analysis of MISR ground retrievals provides a useful self-consistency validation to obtain minimum error bars on SMVs, because ground has no motion. The comparison of overall statistics between previous and current data sets is summarized in Table 1. The vF08_0017 results are from Lonitz and Horváth [2011], and correspond to retrievals with WQ ≥ 3 in the Meteosat-9 full disk area. For vF01_0001 data, calculations were carried out separately for the Meteosat-9 full disk and the entire globe using three different QI thresholds: 0 (all retrievals), 50, and 80. Here, I present only vF01_0001 results for their native 17.6 km resolution, because averaging to 70.4 km yielded similar statistics.

Table 1. MISR Ground Retrieval Statisticsa
 E-W Bias (m s−1)E-W RMSDb (m s−1)N-S Bias (m s−1)N-S RMSDb (m s−1)Height Bias (m)Height RMSDb (m)Sample Number
  1. a

    The vF08_0017 results are from Lonitz and Horváth [2011], corresponding to the Meteosat-9 full disk. For vF01_0001 data results are given separately for the Meteosat-9 full disk and the entire globe.

  2. b

    Root mean square difference.

WQ ≥ 3
vF01_00010.02 / 0.150.54 / 1.240.34 / 0.341.91 / 2.07−33 / −29175 / 2115,438,051 / 13,102,480
QI ≥ 0
vF01_00010.00 / 0.110.49 / 1.050.30 / 0.301.73 / 1.84−31 / −26161 / 1925,023,555 / 11,626,276
QI ≥ 50
vF01_0001−0.01 / 0.080.46 / 0.850.34 / 0.341.65 / 1.68−38 / −32148 / 1692,999,110 / 6,528,446
QI ≥ 80

[20] For “wind” components within the Meteosat-9 full disk, the new algorithm showed only slight general improvement, because previous ground retrievals were already fairly accurate. The E-W component was practically unbiased, with a root mean square difference (RMSD) of ~0.5 m s−1. The N-S component also had a small bias, changing from −0.13 m s−1 to ~0.3 m s−1, but a larger RMSD of ~1.7 m s−1. (Larger N-S errors are a general feature of MISR, caused by the aliasing of motion and height parallaxes in the along-track direction.) The dependence on QI was rather weak, indicating the quality of vF01_0001 retrievals is generally much better and more uniform than that of vF08_0017 retrievals (see also section 3.2.)

[21] Height bias also remained small, changing from +11 m to −30 m, which corresponded to the change in sign and magnitude of the N-S “wind” bias. (Height and N-S component errors are correlated with a sensitivity of approximately −70 to −80 m per +1 m s−1.) The largest improvement occurred in height RMSD, which reduced by half, from 331 to ~160 m. The reduction in height RMSD was not due to the smaller domain size of vF01_0001 retrievals, because averaging to 70.4 km resolution resulted in similar improvements; rather, it reflected the denser and more uniform sampling of a domain by the new area-based stereo matcher compared to the Nested Max matcher, the latter of which sparsely samples only the brightest ~100 points within a 70.4 × 70.4 km2 area.

[22] Extending analysis to the entire globe slightly worsened the statistics, especially for the E-W component. This was due to the inclusion of cloud motion retrievals in polar regions, where distinguishing between low-level clouds and snow/ice covered surfaces was problematic, potentially leading to false ground flags. High-latitude cloudy SMVs affect more the E-W component than the N-S component, partly because the dominant winds are zonal, and partly because the E-W direction corresponds to the more error-prone along-track direction in polar regions. (Whereas at low to middle latitudes the E-W component is more in the cross-track direction while the N-S component is more in the along-track direction.)

[23] Global distributions of ground retrieval “wind” biases are mapped in Figure 2. The E-W biases were mostly within ±1 m s−1 except in Greenland and Antarctica, where positive (westerly) biases could exceed 2 m s−1, possibly due to katabatic wind. The N-S biases were generally larger, typically being within ±2 m s−1, and tended to be more often positive (southerly) than negative (northerly). In polar regions, however, the N-S bias could be smaller than the E-W bias, reflecting the changed heading of the MISR ground track. In fact, for stereo retrievals the sharpest contrast in accuracy is between the cross-track and along-track directions, plotted in Figures 2c and 2d. The cross-track biases were smallest, rarely exceeding ±0.25 m s−1, even at high latitudes. Conversely, the along-track biases were largest, resembling N-S biases at low to middle latitudes and E-W biases in polar regions.

Figure 2.

Mean 1° × 1° “wind” component bias for MISR vF01_0001 ground retrievals with QI ≥ 80: (a) east-west, (b) north-south, (c) cross-track, and (d) along-track. In Figures 2c and 2d the ground track of MISR path 209 is indicated as a guide. Also note the different scale in Figure 2c.

[24] In Figure 3, corresponding retrieved mean surface elevations are compared with the digital elevation model (DEM) from the MISR Ancillary Geographic Product. Agreement between retrieved heights and the DEM was excellent, significantly improved since the previous stereo algorithm. In vF08_0017 data, 75% of height differences were ≤150 m, and 89% were ≤300 m [Lonitz and Horváth, 2011]. For vF01_0001 these values increased to 95 and 99%, respectively, in line with reduced height RMSD given in Table 1. Note the largest height differences were partly due to errors in the MISR DEM. In certain areas DEM errors are as high as several tens of meters, which introduce “wind” retrieval errors through increased coregistration uncertainties (V. Jovanovic, personal communication, 2013).

Figure 3.

Mean 1° × 1° ground elevation from (a) MISR vF01_0001 ground retrievals with QI ≥ 80 and (b) the digital elevation model (DEM) in the Ancillary Geographic Product, and (c) MISR-DEM ground elevation difference.

[25] Finally, I compared vertical profiles of ground retrieval “wind” biases between the data sets, as plotted in Figure 4. Extending analysis from the Meteosat-9 full disk to the entire globe increased the encountered surface elevation range from 0–3.5 km to 0–6.5 km. (Polar regions above 70° latitude were excluded to reduce the impact of false ground flags.) When results were binned according to DEM ground elevation, the most noticeable change was a slightly increased N-S component bias in the new data. Apart from that, both “wind” components in both data sets showed practically no vertical variation. Binning “winds” by their retrieved heights, however, introduced a systematic decrease with height in vF08_0017 data, especially for the N-S component. This vertical rearrangement was due to a negative correlation between N-S (along-track) “wind” errors and height errors, whereby negative N-S component biases result in positive height biases, and vice versa. The aliasing between “wind” and height errors was much less apparent in vF01_0001 retrievals, only affecting lowest and highest elevation bins. As shown in Lonitz and Horváth [2011], stereo matcher and, thus, wind error distributions can be well modeled by the sum of a Gaussian distribution and a uniform distribution representing blunders (gross wind and height errors). The reduced error aliasing reported above indicated less frequent blunders in the upgraded algorithm. Indeed, the fraction of ground retrieval N-S “wind” errors exceeding 5 m s−1, which cause height errors larger than the 500 m bin size, dropped from 2.5% in vF08_0017, to 0.5% in vF01_0001.

Figure 4.

Vertical variation of MISR ground retrieval E-W “wind” bias (asterisks) and N-S “wind” bias (diamonds) as a function of (a) true DEM ground elevation and (b) MISR-retrieved ground elevation, for vF08_0017 (grey, orange) and vF01_0001 (black, red) data. Horizontal bars indicate standard deviations, separately for “winds” larger/smaller than the bin mean. The vF08_0017 results are from Figure 3 in Lonitz and Horváth [2011] and correspond to retrievals with WQ ≥ 3 within the Meteosat-9 full disk, while the vF01_0001 results refer to retrievals with QI ≥ 80 over the entire globe within 70°S–70°N.

3.2 Dependence on Quality Indicators

[26] As described in section 2, the vF08_0017 MISR quality flag takes integer values between 0 and 4, with WQ ≥ 3 representing high-quality retrievals. The Meteosat-9 quality index ranges from 0 to 100, with guidelines prescribing a threshold of QI ≥ 80 for validation purposes [Forsythe and Saunders, 2008]. The new vF01_0001 MISR product employs a EUMETSAT-type quality control scheme with a recommended high-quality threshold of QI ≥ 50 [Mueller et al., 2013], although here I used a high-quality threshold of QI ≥ 80 for consistency with Meteosat-9. In the following, I analyze the vertical variation of retrieval quality in each data set, and dependence of the SMV-CMV comparison on quality indices. The results presented below might guide future investigators in striking the appropriate balance between SMV quality and coverage for their particular applications.

[27] The distribution of quality indices as a function of height is represented by box-whisker plots in Figures 5a, 5b, and 5c, with corresponding vertical profiles of high-quality fractions given in Figure 5d. The quality of vF08_0017 SMVs systematically shifted to lower values as height increased, with a large drop between low and middle levels, and a weaker decrease above 3 km. Correspondingly, the average fraction of high-quality retrievals was 58% at low levels (< 3 km), 33% at middle levels (3–7 km), and 18% at high levels (> 7 km). The quality of vF01_0001 SMVs also decreased with height; however, it was significantly better than that of vF08_0017 SMVs at all altitudes. In vF01_0001, midlevel winds had only slightly worse quality than the best low-level winds, and the largest quality decrease occurred at high levels. The average fraction of high-quality retrievals at low, middle, and high levels was 80%, 70%, and 47%, respectively, representing a factor of 1.4, 2.1, and 2.6 increase in coverage compared to vF08_0017. The fact that high-level winds were the most uncertain in both vF08_0017 and vF01_0001 reflected the general difficulties posed to the stereo technique by clouds at these altitudes. Tracking relatively featureless and low-contrast cirrus between the most oblique and nadir views can be challenging even for the new, area-based stereo matcher. Image matching difficulties might also arise for deep convective clouds, because the oblique cameras mostly view cloud sides, while the nadir camera views cloud tops. In addition, the MISR assumption of no vertical wind breaks down in rapidly developing convective clouds, resulting in large retrieval errors.

Figure 5.

Height dependence of retrieval quality. (a, b, and c) Vertical profiles of quality index distributions for MISR vF08_0017, MISR vF01_0001, and Meteosat-9 winds. The left and right edges of a box are the 25th and 75th percentiles, the line within a box is the median, and the × sign is the mean. Whiskers represent the 5th and 95th percentiles. The number of winds in a given height bin is also indicated. (d) Vertical variation of the fraction of high-quality retrievals defined as WQ ≥ 3 for MISR vF08_0017 SMVs (blue), and QI ≥ 80 for both MISR vF01_0001 SMVs (red) and Meteosat-9 CMVs (black).

[28] In contrast, for Meteosat-9 the quality of high-level CMVs was comparable to, or even slightly better than, that of low-level CMVs. The most noticeable deterioration occurred at middle levels, especially between 3 and 5 km, where not only the quality, but also the number, of retrievals decreased significantly. As discussed in Forsythe and Doutriaux-Boucher [2005] and Lonitz and Horváth [2011], CMV heights are more uncertain in this layer, because it is here where the CO2-slicing technique dominating at higher levels transitions to the brightness temperature-based method favored at lower levels. There is also a nonlinear shift in height assignment from cloud top to cloud base at these altitudes. This increased uncertainty/variability in height assignment can lower CMV QI through spatial and temporal height consistency tests included in the EUMETSAT quality control scheme [Holmlund, 1998; Holmlund et al., 2001]. As a result, the average fraction of high-quality CMVs was 62% at low levels, 47% at middle levels, and 66% at high levels.

[29] In summary, SMV quality significantly improved from vF08_0017 to vF01_0001, with middle and high levels showing the largest relative increases in coverage, by a factor of 2 to 3. At high levels, nevertheless, Meteosat-9 CMVs still had a better average QI than MISR SMVs. At low and middle levels, on the other hand, the QI of new stereo winds was better than that of geostationary winds. (Note, however, that the precise mapping between MISR and Meteosat-9 QI values representing comparable quality will require further experiments.)

[30] Next, I investigated the dependence of SMV-CMV comparison statistics on the various quality indices. Figures 6a and 6b show the number of matched pairs and wind RMSD in the more problematic N-S direction as a function of MISR vF08_0017 WQ and Meteosat-9 QI. Apart from minor discrepancies due to a slightly different ground filter used in the current study, these were essentially the same results as those obtained by Lonitz and Horváth [2011]. The comparison statistics were particularly sensitive to MISR WQ, suggesting more frequent blunders in SMVs than in CMVs. The N-S wind RMSD decreased by a factor of 4 between worst and best SMV bins; however, this came at the price of a comparable reduction in coverage. The dependence on the MISR quality indicator was also rather uneven, with a very sharp drop in N-S wind RMSD between WQ = 2 (“uncertain”) and WQ = 3 (“good”) winds.

Figure 6.

SMV-CMV comparison versus quality indices. (a and b) Sample number and N-S wind RMSD for 70.4 km vF08_0017 SMVs. (c) N-S wind RMSD for vF01_0001 SMVs averaged to 70.4 km and binned by vF08_0017 WQ, as in Figures 6a and 6b. (d and e) Sample number and N-S wind RMSD for the highest-quality 17.6 km vF01_0001 SMVs in 70.4 km domains. (f) N-S wind RMSD for vF01_0001 SMVs averaged to 70.4 km and binned by the mean QI.

[31] The N-S wind RMSD corresponding to 70.4 km mean vF01_0001 SMVs is plotted in Figure 6c. Here all 17.6 km MISR winds (0 ≤ QI ≤ 100) were included in the mean, and for consistency SMV-CMV pairs were binned according to the original vF08_0017 WQ values (amounting to a triple collocation between vF08_0017 SMVs, vF01_0001 SMVs, and Meteosat-9 CMVs). The variation with MISR quality index was now fairly smooth and the N-S wind RMSD decreased by half for previously low-quality MISR SMVs. The improvements for high-quality vF08_0017 MISR winds were more modest, indicating an already tight quality control in the original “good” and “best” wind categories. These comparison statistics, however, could be further improved by imposing a threshold on the vF01_0001 winds included in the averaging. For example, the N-S wind RMSD range between the worst and best bins in Figure 6c reduced from 7.31–3.66 m s−1 to 6.10–2.67 m s−1 and 5.45–2.62 m s−1 when a threshold of QI ≥ 50 or QI ≥ 80 were applied, respectively.

[32] To demonstrate the better coverage of the new MISR winds, in Figure 6d I binned SMV-CMV pairs according to the QI of the highest-quality 17.6 km vF01_0001 SMV within an original 70.4 km vF08_0017 domain. Comparison with Figure 6a indicated a 2.5-fold increase in number of wind pairs in the best bin, from 1.42 × 105 for SMVvF08_0017 WQ = 4 and CMV QI ≥ 80 to 3.57 × 105 for SMVvF01_0001 QI ≥ 80 and CMV QI ≥ 80. This estimate of coverage improvement agreed well with the analysis shown in Figure 5.

[33] The N-S wind RMSD corresponding to the binning in Figure 6d is plotted in Figure 6e. Although, as before, the comparison significantly improved for previously low-quality SMVs, the RMSD values in higher-quality bins here were 30–50% larger than the ones for the 70.4 km averages shown in Figure 6c. This was partly due to an increase in representativeness error when comparing 17.6 km resolution SMVs to nominally 72 km resolution Meteosat-9 CMVs. Indeed, RMSD reduced when vF01_0001 SMVs were averaged to 70.4 km and binned according to their mean QI as in Figure 6f, but differences were still slightly larger than in Figure 6c. (Note that Figures 6c and 6f contain the exact same SMV-CMV pairs, but the former used vF08_0017 WQ, while the latter vF01_0001 mean QI for binning.) For example, the RMSD in the best bin was 3.66 m s−1 in Figure 6c (for SMVvF08_0017 WQ = 4 and CMV QI ≥ 80) and 4.05 m s−1 in Figure 6f (for SMVvF01_0001 mean-QI ≥ 80 and CMV QI ≥ 80). This suggested the old MISR “best” wind category had slightly tighter quality control than the new one, albeit at the cost of much-reduced coverage.

3.3 Full-Disk Mean Comparison

[34] Full-disk annual-mean SMV-CMV comparison statistics are given in Tables 2, 3, and 4 for low- (< 3 km), middle- (3–7 km), and high-level (> 7 km) clouds. The vF08_0017 results are essentially the same as in Lonitz and Horváth [2011], with only minor differences due to the different ground filter and height bins (geometric vs. pressure height) used in the current study. The vF01_0001 results correspond to 70.4 km average SMVs calculated separately for three different QI thresholds.

Table 2. Full-Disk Annual-Mean Comparison of Meteosat-9 CMVs (QI ≥ 80) and MISR vF08_0017 SMVs (WQ ≥ 3) or vF01_0001 SMVs for Low-Level (< 3 km) Clouds, as Determined by MISR SMV Heighta
 E-W MDbE-W RMSDcE-W Corr.dN-S MDbN-S RMSDcN-S Corr.dHeight MDbHeight RMSDcHeight Corr.dSpeed MDbMVDeSDMVDfNRMSVDg
  1. a

    For vF01_0001 data, all 17.6 km SMVs exceeding a certain QI threshold (0, 50, or 80) were averaged within a 70.4 km domain.

  2. b

    Mean difference (m s−1 or m).

  3. c

    Root mean square difference (m s−1 or m).

  4. d


  5. e

    Mean vector difference (m s−1).

  6. f

    Standard deviation about MVD (m s−1).

  7. g

    Root mean square vector difference normalized by MISR wind speed.

WQ ≥ 3 (199831)
QI ≥ 0 (197453)
QI ≥ 50 (197362)
QI ≥ 80 (172311)
Table 3. Full-Disk Annual-Mean Comparison of Meteosat-9 CMVs (QI ≥ 80) and MISR vF08_0017 SMVs (WQ ≥ 3) or vF01_0001 SMVs for Mid-Level (3–7 km) Clouds, as Determined by MISR SMV Heighta
 E-W MDbE-W RMSDcE-W Corr.dN-S MDbN-S RMSDcN-S Corr.dHeight MDbHeight RMSDcHeight Corr.dSpeed MDbMVDeSDMVDfNRMSVDg
  1. a

    For vF01_0001 data, all 17.6 km SMVs exceeding a certain QI threshold (0, 50, or 80) were averaged within a 70.4 km domain.

  2. b

    Mean difference (m s−1 or m).

  3. c

    Root mean square difference (m s−1 or m).

  4. d


  5. e

    Mean vector difference (m s−1).

  6. f

    Standard deviation about MVD (m s−1).

  7. g

    Root mean square vector difference normalized by MISR wind speed.

WQ ≥ 3 (13300)
QI ≥ 0 (13361)
QI ≥ 50 (13031)
QI ≥ 80 (10059)
Table 4. Full-Disk Annual-Mean Comparison of Meteosat-9 CMVs (QI ≥ 80) and MISR vF08_0017 SMVs (WQ ≥ 3) or vF01_0001 SMVs for High-Level (> 7 km) Clouds, as Determined by MISR SMV Heighta
 E-W MDbE-W RMSDcE-W Corr.dN-S MDbN-S RMSDcN-S Corr.dHeight MDbHeight RMSDcHeight Corr.dSpeed MDbMVDeSDMVDfNRMSVDg
  1. a

    For vF01_0001 data, all 17.6 km SMVs exceeding a certain QI threshold (0, 50, or 80) were averaged within a 70.4 km domain.

  2. b

    Mean difference (m s−1 or m).

  3. c

    Root mean square difference (m s−1 or m).

  4. d


  5. e

    Mean vector difference (m s−1).

  6. f

    Standard deviation about MVD (m s−1).

  7. g

    Root mean square vector difference normalized by MISR wind speed.

WQ ≥ 3 (12024)
QI ≥ 0 (9266)
QI ≥ 50 (8954)
QI ≥ 80 (5741)

[35] Compared to Meteosat-9 CMVs, vF08_0017 SMVs showed negative biases in both components at all levels, that is, MISR winds had weaker westerlies and southerlies (or stronger easterlies and northerlies). The magnitude of the E-W wind bias was typically <1 m s−1 with an RMSD between 2 and 4 m s−1. The N-S wind bias and RMSD were larger, and steadily increased with height, from −0.8 m s−1 to −4.3 m s−1, and from 3.5 m s−1 to 9.2 m s−1, respectively.

[36] In vF01_0001 data, the already quite good E-W wind statistics showed only small changes. The magnitude of the bias decreased to <0.5 m s−1; the RMSD and correlation, however, remained practically unchanged. The only notable exception was a 0.5–0.8 m s−1 (15–20%) increase in RMSD at middle level. Recall that this is the layer where Meteosat-9 height assignment techniques are the least consistent, as indicated by the reduced CMV quality in Figure 5. Therefore, the slightly increased E-W wind scatter at middle level might have been indicative of CMV height uncertainties more than anything else.

[37] In contrast, the N-S wind component showed clear and substantial improvement in new stereo data at all heights. The negative bias was eliminated at low level, and reduced by 2.5–3.0 m s−1 at middle and high levels. The RMSD also decreased by 0.8–0.9 m s−1 at low level, and 3–4 m s−1 at high and middle levels, accompanied by significant increases in correlation. As a result, N-S wind errors became comparable to E-W wind errors at low and middle levels, and only at high level did a significant contrast in uncertainty remain between wind components.

[38] The reduction in negative N-S wind bias led to a general decrease in vF01_0001 SMV heights. At low and high levels the new MISR heights still exceeded Meteosat-9 heights, but the mean SMV-CMV height difference decreased from 430–630 m to 200–300 m. At middle level, however, the mean height difference did not decrease in magnitude, but simply changed sign from +335 m to −[260–350] m; that is, the new SMV heights were, on average, lower than CMV heights. The RMSD and correlation values indicated that agreement between MISR and Meteosat-9 heights remained generally poor, and at middle and high levels, even deteriorated compared to vF08_0017 data. Lack of height comparison improvements corresponding to MISR N-S wind improvements strongly suggested the dominance of Meteosat-9 height assignment errors in the SMV-CMV height differences.

[39] It is also worth noting that the MISR QI threshold had little influence on wind comparison statistics at low and middle levels, and most affected results at high level. This indicated an across-the-board SMV improvement for low-level and midlevel cloud types, and revealed that MISR quality control best captures retrieval uncertainties in cirrus and deep convection.

[40] To analyze sampling differences between the old and new MISR data sets, in Table 5 I divided the original MISR vF08_0017 SMV – Meteosat-9 CMV matched pairs into two complementary groups: one with and one without corresponding high-quality (QI ≥ 80) vF01_0001 SMVs (triple collocation versus double collocation). The second group contained original wind pairs that were removed from a given height bin in the new data set, mostly by improved quality control, but in some cases by virtue of the new SMV height falling into a different height category. The comparison statistics were at all levels better for wind pairs with corresponding high-quality vF01_0001 SMVs than for pairs without, confirming the overall positive effect of the new quality control scheme. The contrast in wind difference statistics between the two groups was relatively small for low-level clouds, but it was substantial for midlevel and high-level clouds. For example, the midlevel E-W wind mean difference, N-S wind mean difference, and RMSD doubled between the groups.

Table 5. Full-Disk Annual-Mean Comparison of Meteosat-9 CMVs (QI ≥ 80) and MISR vF08_0017 SMVs (WQ ≥ 3) With or Without Corresponding High-Quality vF01_0001 SMVs in the Same MISR Height Bina
 E-W MDbE-W RMSDcE-W Corr.dN-S MDbN-S RMSDcN-S Corr.dHeight MDbHeight RMSDcHeight Corr.dSpeed MDbMVDeSDMVDfNRMSVDg
  1. a

    High-quality vF01_0001 SMVs were defined as QI ≥ 80.

  2. b

    Mean difference (m s−1 or m).

  3. c

    Root mean square difference (m s−1 or m).

  4. d


  5. e

    Mean vector difference (m s−1).

  6. f

    Standard deviation about MVD (m s−1).

  7. g

    Root mean square vector difference normalized by MISR wind speed.

Low Level
with vF01_0001 (169312)−0.352.250.96−0.803.360.8644410190.350.102.872.840.44
without vF01_0001 (30519)−0.483.010.94−0.904.160.8139812610.28−0.033.663.610.55
Mid Level
with vF01_0001 (8010)−0.743.480.96−1.974.960.839319300.590.374.444.110.47
without vF01_0001 (5290)−1.354.270.94−4.659.390.6270222610.611.817.107.480.75
High Level
with vF01_0001 (5510)−0.543.030.98−3.517.260.9054813070.781.146.224.820.35
without vF01_0001 (6514)−0.823.870.98−4.9410.600.8370214530.791.787.588.360.46

[41] Height difference statistics, on the other hand, were fairly comparable between groups for low- and high-level winds, showing significant contrast only for midlevel winds. At middle level, the vF01_0001 data set preferentially removed wind pairs in which the vF08_0017 MISR heights largely overestimated Meteosat-9 heights (by 702 m on average); as discussed above, these data pairs also showed greatly increased wind differences. The rest of the midlevel MISR vF08_0017 SMVs showed good height agreement with Meteosat-9 CMVs, with a mean difference of only 93 m. These sampling differences between vF08_0017 and vF01_0001 MISR winds help explain changes in the SMV-CMV mean height differences given in Tables 2, 3, and 4. For low- and high-level winds, the SMV-CMV mean height difference decreased by 200–300 m between old and new MISR data sets, due mainly to a decrease in new SMV heights resulting from an increased southerly wind component (or reduced negative N-S wind bias). For midlevel winds, the change in the sign of the SMV-CMV mean difference implied a much larger, 600–650 m reduction in new SMV heights. As Table 5 indicates, however, the large change in the midlevel SMV-CMV mean height difference was a combination of two factors: a decrease of 200–300 m in new SMV heights—similar to low- and high-level winds—and preferential exclusion by the new QI scheme of wind pairs with large vF08_0017 height overestimations.

[42] To further demonstrate MISR improvements from vF08_0017 to vF01_0001, Figure 7 plots SMV-CMV vector differences in a wind rose. As discussed by Lonitz and Horváth [2011], vector differences increase in the 14°/194° direction, which corresponds to Terra/MISR along-track direction. Dominant features of the observed pattern could be explained by retrieval sensitivities to stereo matching errors, especially in the D camera. The vF08_0017 difference distribution was asymmetric with a larger frequency peak along 14° than along 194°, causing the negative SMV bias (Figure 7a). The weak but clear dependence of height differences on N-S wind differences was also indicative of the error correlations in MISR retrievals (Figure 7b). The difference distribution for vF01_0001 data, on the other hand, was fairly symmetric (Figure 7c), with only a faint trace of the MISR error pattern (Figure 7d), thanks to the much-reduced SMV bias and RMSD.

Figure 7.

Distribution of MISR SMV – Meteosat-9 CMV vector difference colored according to (a and c) frequency and (b and d) mean-difference-corrected height difference. Figures 7a and 7b are from Figure 10 in Lonitz and Horváth [2011] and correspond to vF08_0017 SMVs with WQ ≥ 3, while Figures 7c and 7d refer to 70.4 km averages of vF01_0001 SMVs with QI ≥ 80. The Meteosat-9 CMVs have QI ≥ 80 in both cases. Triangles indicate error vector directions due to stereo matching errors in the D along-track (red), D cross-track (blue), B along-track (green), and B cross-track (cyan) positions.

3.4 Meridional Variations

[43] Meridional variations in zonal-mean SMV-CMV comparison statistics are shown in Figure 8, separately for low-, middle-, and high-level clouds as determined by MISR height. The vF08_0017 results are practically the same as in Lonitz and Horváth [2011], while the vF01_0001 results refer to 70.4 km averages of SMVs with QI ≥ 80. The conclusions drawn below regarding the new stereo data set are largely independent of QI threshold, because zonal means, similar to full-disk means, were sensitive to QI at high levels only, and even there only moderately so.

Figure 8.

Meridional variation of MISR – Meteosat-9 annual (a, d, and g) mean wind difference, (b, e, and h) wind RMSD, and (c, f, and i) wind correlation, separately for low-, middle-, and high-level clouds, as determined by MISR height. Dotted and solid lines correspond to MISR vF08_0017 SMVs with WQ ≥ 3 and 70.4 km averages of vF01_0001 SMVs with QI ≥ 80, respectively. The E-W and N-S wind components are plotted in black and red.

[44] Zonal-mean statistics for the more accurate E-W wind component showed generally small changes between the old and new MISR data sets. The largest improvements occurred in the N-S wind statistics at middle and high levels, although the improvement at low level was also significant. The negative N-S MISR wind bias greatly decreased in vF01_0001 data at all latitudes, with particularly large, 2 to 4 m s−1, reductions for middle- and high-level clouds. The hemispheric asymmetry in vF08_0017 N-S wind bias at middle level, whereby MISR biases were 1 to 2 m s−1 larger in the Southern Hemisphere than in the Northern, was mostly eliminated in the new data set. These bias improvements were accompanied by significant reductions in N-S wind RMSD. As a result, the bias and RMSD in N-S winds became comparable to those in E-W winds at all latitudes for low- and middle-level clouds. For high-level clouds, the N-S wind RMSD was still 1.0–1.5 m s−1 larger than the E-W wind RMSD; however, the large peaks at the equator and 50°S/N in vF08_0017 data disappeared in vF01_0001 data. The N-S wind RMSD of the updated high-level retrievals now showed a smooth meridional variation, with a slight increase from equator to higher latitudes. Simultaneously, the N-S wind correlation increased significantly in the tropics/subtropics (30°S–30°N), especially at middle and high levels.

[45] Although changes to the E-W wind statistics were typically small, resulting in a slight bias improvement, at certain latitudes the RMSD increased, and the correlation decreased noticeably: 30°–35°S for midlevel and 20°–25°S/N for high-level clouds. The deterioration of the E-W wind comparison in these regions was probably related to CMV height assignment errors or sampling differences between vF08_0017 and vF01_0001. Recall, only high-confidence cloud retrievals were considered from the new data set, but low-confidence cloud SMVs were excluded. Nevertheless, this issue warrants further investigation.

[46] The zonal-mean comparison for wind height differences between the new SMV and CMV did not improve, compared to the results in Lonitz and Horváth [2011]. In line with full-disk mean results, MISR SMV heights and, thus, SMV-CMV height differences, decreased by a few hundred meters at all latitudes, but otherwise meridional variations remained largely unchanged. In terms of RMSD and correlation, height comparison even worsened at middle and high levels. Some of these changes might have been due to the different behavior of old and new stereo matchers in multilayer cloud fields. The vF08_0017 Nested Max matcher is proven to favor well-textured low-level clouds. Much less is known about the performance of the vF01_0001 area-based matcher, although it is expected to be less error-prone in cirrus than Nested Max. Therefore, it is possible that in certain cases, the new matcher tracked upper level clouds, while Nested Max and Meteosat-9 both tracked the lower levels, causing increased height differences in the updated data set. Lack of an overall improvement in wind height comparison mirroring the vast improvements in MISR N-S wind retrievals, however, reinforced the notion that SMV-CMV height differences were primarily driven by Meteosat-9 height assignment errors.

3.5 Regional Variations

[47] Regional variations of changes in SMV-CMV comparison statistics between MISR vF01_0001 and vF08_0017 are given for the E-W and N-S wind components in Figures 9 and 10, respectively. The vF01_0001 – vF08_0017 change in magnitude of SMV-CMV mean wind difference, RMSD, and correlation was calculated separately for low-, middle-, and high-level clouds. As shown, high-quality low-level SMVs were found mostly over ocean, and to a lesser degree in Europe and East Africa, while high-quality land retrievals were predominantly at middle and high levels.

Figure 9.

Change in E-W wind absolute mean difference, RMSD, and correlation between MISR vF01_0001 SMV – Meteosat 9 CMV and MISR vF08_0017 SMV – Meteosat 9 CMV comparison, for (a, b, and c) low-level, (d, e, and f) middle-level, and (g, h, and i) high-level clouds, as determined by MISR height. Results are significant at the 95% confidence level and correspond to vF08_0017 SMVs with WQ ≥ 3 and 70.4 km averages of vF01_0001 SMVs with QI ≥ 80. Negative/positive values indicate a decrease/increase in the validation statistics of vF01_0001 retrievals. Grid size is 2° × 2° for low-level clouds, but 4° × 4° for middle- and high-level clouds, given the much-reduced sampling in the latter two altitude bins.

Figure 10.

Same as Figure 9 but for the N-S wind component.

[48] Relative to Meteosat-9 CMVs, the MISR E-W wind validation statistics showed both increased and decreased differences between new and old stereo algorithms, depending on location (Figure 9). The distribution of the new-old validation differences, however, was symmetric, centering on zero for all parameters at all levels, and yielding only small overall changes when averaged over the full-disk (see Tables 2, 3, and 4.) Changes in SMV-CMV absolute mean wind difference between vF01_0001 and vF08_0017 were typically within ±0.4, ±0.9, and ±1.0 m s−1 for low-, middle-, and high-level clouds. Corresponding changes in RMSD varied within ±1.0, ±1.7, and ±1.6 m s−1, while correlation changes were usually within ±10%. The smallest regional changes occurred in boundary layer clouds over the central and eastern South Atlantic (±0.25 m s−1 for mean difference and RMSD, and ±2.5% for correlation), where old MISR winds showed generally the best performance [Lonitz and Horváth, 2011].

[49] In contrast, the vF01_0001 N-S wind comparison statistics significantly improved most everywhere and in all height bins, as compared to vF08_0017 statistics (Figure 10). The SMV-CMV absolute mean wind difference decreased by 1 to 2 m s−1 or more, with middle and high levels experiencing the larger reductions. Simultaneously, RMSD decreased by similar amounts in most locations, with the correlation increasing by 10 to 20% or more, especially in the tropics and subtropics (30°S–30°N). The only deterioration of validation statistics was found in low-level winds, where the N-S mean wind difference noticeably increased in two regions: (i) a narrow band over the Atlantic Ocean running between continents, and, (ii) the oceanic ITCZ around the Gulf of Guinea (5°S–5°N, 20°W–10°E). Compared to Meteosat-9, the low-level vF08_0017 MISR N-S winds were negatively biased everywhere, except these two regions where they showed a positive bias [Lonitz and Horváth, 2011]. The vF01_0001 stereo algorithm made meridional winds generally more positive/southerly (or less negative/northerly) and, therefore, further increased positive MISR bias in these two regions, while reducing negative MISR bias everywhere else.

[50] As also discussed by Lonitz and Horváth [2011], region (i) corresponded to the oceanic MISR paths 207–213, which had only a few land tie points for georegistration. Consequently, these paths were more often flagged for uncertain geolocation, resulting in reduced sampling that likely caused the increase in low-level N-S SMV bias compared to other paths. For middle- and high-level clouds other errors, such as in stereo matching, were the dominant ones, whose large improvements overcompensated registration errors, overall yielding reduced middle- and high-level SMV-CMV differences even in this oceanic region.

[51] The SMVs in region (ii), on the other hand, were from paths running over Europe and the Sahara, which contained a large number of land tie points for georegistration. MISR biases due to orbit-level geolocation uncertainties were, thus, unlikely to explain the increased N-S mean wind difference here. Note low-level Meteosat-9 winds also show considerably larger observation-model differences in this region than in surrounding marine areas. (See maps on the Satellite Application Facility on Numerical Weather Prediction AMV monitoring web site at http://www.nwpsaf.org.) Visible and near-infrared CMVs below 600 hPa over sea near Guinea are even blacklisted in the UK Met Office NWP model; therefore, increased N-S wind difference in the Gulf of Guinea might actually be the result of comparing improved vF01_0001 SMVs to biased CMVs. Future studies could perhaps investigate the effect of biomass burning smoke and desert dust on tracer selection and height assignment, which could be partially responsible for CMV biases in this area.

[52] In accordance with full-disk and zonal mean results, regional height differences generally did not improve in vF01_0001 data, and for middle- and high-level clouds the comparison even worsened. The only notable exception was a reduction of a couple of hundred meters in the SMV-CMV mean height difference over low-level marine clouds, caused by a similar decrease in the new SMV heights. CMV heights, nevertheless, remained biased low by 500 to 1000 m in marine boundary layer clouds compared to MISR, lidar, and model cloud heights, even after considering cloud base adjustments in the Meteosat-9 retrievals. As demonstrated by Lonitz and Horváth [2011], this low bias in brightness temperature-based CMV heights was likely due to contributions from the warm sea surface in broken cloud scenes.

3.6 Vertical Variations

[53] Vertical variations in SMV-CMV comparison statistics are given in Figure 11 at a scale finer than the coarse low-, middle-, and high-level categories discussed previously. In vF08_0017 data, the mean difference was negative for both wind components at all levels except the lowermost. This was partly caused by the general error aliasing in stereo retrievals, whereby negative/positive N-S wind errors entailed overestimation/underestimation of heights. (A similar vertical rearrangement due to binning by MISR height was also apparent in ground retrievals, as discussed in section 3.1.) The magnitude of the E-W wind mean difference was relatively small at all levels, typically ≤1.5 m s−1. The magnitude of the N-S wind mean difference, however, rapidly increased with height, from +2 m s−1 at the surface to −8 m s−1 at 14 km altitude. There was also a large local maximum in N-S wind mean difference at 3–4 km.

Figure 11.

Vertical variation of MISR – Meteosat-9 E-W wind (asterisks) and N-S wind (diamonds) comparison statistics for vF08_0017 (grey, orange) and vF01_0001 (black, red) MISR data: (a) mean wind difference, (b) RMSD, and (c) correlation. The vF08_0017 results are from Figure 11 in Lonitz and Horváth [2011] and correspond to SMVs with WQ ≥ 3, while vF01_0001 results refer to 70.4 km averages of SMVs with QI ≥ 80. Data were binned according to MISR height.

[54] In vF01_0001 data, the E-W wind mean difference remained negative at low and middle levels, but turned positive at high level, with magnitude reducing to ≤0.5 m s−1. The N-S wind mean difference remained negative everywhere except the lowermost layer; however, its magnitude was drastically reduced to ≤1.5 m s−1. The stereo error aliasing and local N-S wind difference maximum at 3–4 km also became less prominent in the new retrievals.

[55] The E-W wind RMSD showed relatively small vertical variation for both old and new SMVs. However, its value slightly increased from 2–4 m s−1 in vF08_0017, to 2–5 m s−1 in vF01_0001, the largest increases of 0.5–1.0 m s−1 occurring at ~4 km and above 12 km altitude. The vF08_0017 N-S wind RMSD, on the other hand, strongly increased with height throughout the troposphere, from 3 m s−1 to 12 m s−1. The vertical variation of N-S wind RMSD in vF01_0001 was significantly reduced to 3–6 m s−1, with middle and high levels experiencing the largest decreases. As a result of these improvements in N-S wind mean difference and RMSD, MISR retrieval errors in the two wind components became comparable at low and middle levels, and the meridional wind remained slightly less accurate than the zonal wind only at high level.

[56] The E-W wind correlation tended to slightly increase with height in both data sets, reflecting the stronger and predominantly zonal flow at higher altitudes. Similar to E-W wind RMSD, the E-W wind correlation also showed a minor deterioration for new vF01_0001 retrievals. The vF08_0017 N-S wind correlations were smaller and had a more complex vertical pattern, with a large drop to 0.63 in the 3–4 km height range. In vF01_0001 data, the N-S wind correlation showed similar relative variation, but significantly increased at all levels, typically by 5 to 15%.

[57] In summary, the MISR – Meteosat-9 E-W wind mean difference improved, but the RMSD and correlation slightly worsened, especially at middle and high levels, for vF01_0001 SMVs. Rather than being indicative of increased MISR errors, the slight worsening of E-W wind comparison at middle and high levels, where wind shear and variability are strong, might have actually been due to improvements in SMV heights relative to systematically biased CMV heights. The N-S wind statistics, on the other hand, showed definite and significant improvements throughout the entire atmosphere. A local deterioration in N-S wind comparison was still apparent at 3–4 km, although it was less dramatic for new MISR retrievals. The larger differences in this layer were most likely related to reduced CMV quality (Figure 5), and the shift in CMV height assignment from cloud top to cloud base, which could have resulted in height-mismatched SMV-CMV pairs.

3.7 Cross-Swath Variations

[58] Lonitz and Horváth [2011] noticed cross-swath biases in vF08_0017 MISR data, whereby SMV errors were significantly larger on the eastern side of the swath than on the western. These biases affected N-S wind and height the most, and were traced back to uncorrected lens aberrations during L1B2 georectification and registration [Mueller et al., 2013]. Although MISR geolocation accuracies were well within design specifications, the Camera Geometric Model did not fully account for focal plane distortions [Jovanovic et al., 2007]. While the residual optical distortion in the cross-track direction was corrected in vF08_0017, the distortion in the along-track direction was neglected. The resulting along-track coregistration error was found to vary across the swath with a magnitude of ±0.4 pixel for the most oblique 70° views, and ±0.1 pixel for the 46° views. The vF01_0001 algorithm applies subpixel registration corrections in the along-track direction as well to further mitigate the effect of optical distortions. These corrections were derived from cloud-free, terrain-referenced imagery, and showed no variation throughout the mission time line.

[59] The effect of vF01_0001 upgrades on the cross-swath bias is depicted in Figure 12. The 70.4 km MISR domains were numbered 1 through 8 from west to east across the swath; however, the edge domains were usually not observed by all nine cameras and, thus, contained no retrievals. Note that registration corrections were actually applied at the native 17.6 km resolution in the new algorithm, but here I only show the 70.4 km average results.

Figure 12.

Cross-swath dependence of (a) MISR ground retrievals, (b) SMV-CMV differences, and (c) SMV-CMV differences with MISR ground retrieval biases subtracted, separately for E-W wind (asterisks), N-S wind (diamonds), and height (triangles), and, vF08_0017 (grey, orange, green) and vF01_0001 (black, red, blue) MISR data. The vF08_0017 results are from Figure 12 in Lonitz and Horváth [2011] and correspond to SMVs with WQ ≥ 3, while vF01_0001 results refer to 70.4 km averages of SMVs with QI ≥ 80.

[60] The cross-swath dependence of MISR ground retrieval biases is plotted in Figure 12a. The magnitude of variation in vF08_0017 E-W wind, N-S wind, and height was 0.6 m s−1, 2.6 m s−1, and 210 m, respectively. The vF01_0001 registration corrections had a relatively small effect in the western half of the swath (domains 2, 3, 4), but significantly improved biases in the eastern half (domains 5, 6, 7), especially for N-S wind and height. As a result, the amplitude of E-W wind variations reduced slightly to 0.4 m s−1, and that of N-S wind and height variations decreased by half to 1.3 m s−1 and 105 m in the new retrievals. In both data sets, the cross-swath variation of height bias mirrored that of N-S wind bias due to coupling between these two parameters with a sensitivity of −70 to −80 m per m s−1.

[61] The cross-swath dependencies of SMV-CMV mean differences given in Figure 12b were very similar to those of MISR ground retrieval biases. The vF01_0001 registration corrections had again the largest effect in the eastern half of the swath, yielding similar bias improvements as in ground retrievals. For the new SMVs, the E-W wind and N-S wind mean difference had reduced west-to-east variation from −0.4 m s−1 to +0.1 m s−1, and from −0.9 m s−1 to +0.5 m s−1. Height differences generally decreased by 200–250 m due to a decrease in vF01_0001 SMV heights, and cross-swath variations also reduced to between 140–260 m.

[62] As shown in Figure 12c, the SMV-CMV cross-swath variations could be almost completely eliminated in both data sets by subtracting the corresponding MISR ground retrieval biases from Figure 12a. Swath-mean differences, however, were still smaller in vF01_0001 data, especially for N-S wind and height. In addition, the ground-bias correction could only be applied in a statistical sense and, thus, for instantaneous retrievals vF01_0001 was clearly superior to vF08_0017, as demonstrated in Figure 12b.

3.8 Vortex Street Case Study: 4.4 km SMVs

[63] The operational vF01_0001 SMVs are generated at an improved resolution of 17.6 km; however, the new stereo algorithm is capable of retrieving wind vectors at even finer scales. This is demonstrated in Figure 13 for a von Kármán vortex street that formed in the wake of Jan Mayen Island, located in the Norwegian Sea. The ice-capped 2277 m Beerenberg volcano often produces mesoscale vortices in stratocumulus-topped flow downwind of the island. The vortex street imaged by MISR on 6 June 2001 (path 217, orbit 7808, blocks 33–35), was particularly impressive, comprising eight pairs of alternately rotating vortices extending 350 km southward of the volcano. Animating the nine separate views clearly showed the eastern train of vortices rotating clockwise and the western train of vortices rotating counter-clockwise. This highly textured cloud field was well suited to test the new algorithm's ability to retrieve the fine structure of complex atmospheric flows.

Figure 13.

Experimental MISR retrievals for the von Kármán vortex street off Jan Mayen Island at 1252 UTC on 6 June 2001: (a) 4.4 km residual winds after mean wind removal, (b) relative vorticity field, and (c) wind-corrected 1.1 km cloud top heights. The 70.4 km wide and 348 km long image segment is from path 217, orbit 7808, blocks 33–35. Cyan and orange stars mark the clockwise and counter-clockwise rotating vortex centers, and the yellow centerline is a third-order polynomial fit to the intervortex midpoints. The cross-street distance (H), the along-street distance (L), and the cross-wind island width (D) are also indicated.

[64] Height-resolved stereo winds obtained at 4.4 km resolution are overlaid on the MISR nadir image in Figure 13a. Here I plotted residual winds after subtracting upstream wind; that is, I used a coordinate system moving with the mean flow. Retrievals outside the island wake indicated a mean upstream wind direction and speed of ~25° (north-northeast) and ~12 m s−1, respectively. The residual winds were clearly not random and captured the expected flow pattern, including the counter-rotating vortices, reasonably well. Deviations from a textbook vortex wind structure were partly caused by assuming a spatially and temporally constant upstream wind estimated at the satellite overpass time of 1252UTC, although the large-scale wind slightly changed during vortex shedding, as indicated by the curvature of the vortex street.

[65] The corresponding relative vorticity field is plotted in Figure 13b. The clockwise and counter-clockwise rotating vortices were nicely collocated with local minima and maxima in vorticity. The relative vorticity averaged over the vortices varied between 5 and 15 × 10−4 s−1, tending to decrease with distance from the island, as shown in Figure 14. The MISR-observed magnitude and downwind trend of vorticity were in good agreement with the idealized large eddy simulations of Heinze et al. [2012], which showed vortex size increasing and mean vorticity decreasing downstream due to diffusion (see their Figure 6). These simulations also indicated the vortex core is ~0.2 K warmer than its environment, and features a continuous updraft fed by a convergent near-surface inflow of warm air. The updraft is associated with a divergent outflow at vortex top, which leads to a sinking inversion and additional temperature increase. Heinze et al. [2012] hypothesized this local lowering of the capping inversion to be responsible for the typically cloud-free vortex eye; the gradual filling of vortex centers, also visible in Figure 13a, could then be explained by downstream decrease in vortex core temperature.

Figure 14.

Mean relative vorticity of the vortices marked in Figure 13b versus distance from the island center. The downwind distance x is normalized by the estimated island base diameter D1 = 14.6 km. Blue and red stars indicate clockwise and counter-clockwise rotating vortices, and the solid lines are the corresponding linear regression lines.

[66] From the MISR nadir image I also calculated two different dimensionless ratios describing vortex geometry: (i) the aspect ratio of cross-street distance between two halves of the vortex street H to the along-street distance between two adjacent, like-rotating vortices L, and, (ii) the ratio of cross-street distance H to crosswind width of the island D, also known as dimensionless width. These quantities were computed following the methodology of Young and Zawislak [2006], who analyzed island wake vortex streets in Moderate Resolution Imaging Spectroradiometer imagery. First, the vortex center positions were determined as centroids of clear pixels within each vortex; the first three vortices in the immediate lee of the island, and the very last vortex were excluded because they lacked well-defined cloud-free cores, yielding 12 vortex centers in total (cyan and orange stars for clockwise and counter-clockwise rotation). Cross-street and along-street distances were then calculated relative to a curved centerline obtained by fitting a third-order polynomial to the intervortex midpoints (yellow line). Finally, I computed two estimates for the crosswind width of the island: (i) at sea level from two coastal points (D1 = 14.6 km), and, (ii) at inversion level from the clearing in the cloud field around the volcano (D2 = 10.2 km), which might have been the more relevant value.

[67] From similarity theory and laboratory measurements, von Kármán and Rubach [1912] derived an aspect ratio of 0.28, while Tyler [1930] reported a dimensionless width of 1.2 for two-dimensional, inviscid, neutrally stratified flow around a cylinder. Island wake vortices, in contrast, usually form in a highly viscous, well mixed, stratified boundary layer capped by an inversion. For such three-dimensional atmospheric flows, earlier satellite observations yielded aspect ratios of 0.33 to 0.60, and dimensionless widths of ~1; however, the estimates were rather sensitive to view geometry and pixel size. In a recent study based on high-resolution Moderate Resolution Imaging Spectroradiometer imagery, Young and Zawislak [2006] found that atmospheric vortex streets do follow geometric similarity theories, but with larger values for the dimensionless ratios than those predicted for inviscid flow around a bluff body. From a sample of 30 cases, they derived for the aspect ratio a 95% confidence interval of 0.36 to 0.47 with a mean value of 0.42, when assuming a straight centerline. Using a curved centerline, which might be more relevant for this case, the observed confidence interval and mean value were 0.30 to 0.43 and 0.37, respectively. Similarly, the confidence interval and mean value of dimensionless width were found to be 1.23 to 2.00 and 1.62 without curvature, and 1.40 to 2.25 and 1.83 with curvature.

[68] In the current vortex street, cross-street distance showed only small variation between 13.5 and 16.4 km with a mean of 14.8 km. The along-street distance, however, systematically increased downstream of the island, from 32.0 to 49.3 km, with a mean of 38.7 km. As a result, the aspect ratio varied between 0.31 and 0.42, having a mean value of 0.38, in good agreement with the Young and Zawislak [2006] estimates for curved centerlines. My overall estimate for the dimensionless width was 1.02 or 1.45, depending on crosswind island diameter used (D1 or D2). These values were considerably lower than those of Young and Zawislak [2006]; however, the dimensionless width ratio is a more uncertain quantity due to difficulty determining the elevation at which island diameter is relevant to the flow.

[69] Finally, the wind-corrected, 1.1 km resolution MISR heights are given in Figure 13c. The satellite-measured peak elevation of 2023 m represented a 254 m underestimation of Beerenberg, well within the ±400 m uncertainty of instantaneous retrievals. The southern/southwestern slopes of the volcano were also well captured. Cloud-top heights were typically between 600 and 1500 m, and showed reasonably coherent, although hard-to-interpret, fluctuations. This case study hopefully demonstrated that high-resolution stereo wind and height retrievals can help describe the fine structure of atmospheric vortices, and might even aid the evaluation of numerical simulations.

4 Summary and Concluding Remarks

[70] The MISR stereo algorithm, which simultaneously retrieves cloud motion and associated height, recently underwent a major overhaul, and the Level 2 cloud product was reprocessed to the new standard for the entire Terra mission starting from 1 March 2000. Upgrades included an area-based stereo matcher, improved quality control modeled on the EUMETSAT scheme, subpixel registration corrections to mitigate cross-swath biases arising from focal plane distortions, and an increased retrieval resolution of 17.6 km. This paper documented the resulting improvements to winds in the new TC_CLOUD vF01_0001 product compared to the previous TC_STEREO vF08_0017 product. I repeated the MISR SMV – Meteosat-9 CMV comparison study of Lonitz and Horváth [2011], replacing vF08_0017 SMVs with their vF01_0001 counterparts in the ~200,000 collocated wind pairs obtained from one year's worth of data. To facilitate a direct comparison with earlier results and reduce representativeness errors, the new MISR winds were averaged down from their native resolution to the previous standard of 70.4 km.

[71] The quality of the new winds was significantly better than that of the old winds at all levels, although still tended to decrease with height. The coverage of high-quality retrievals, defined as the fraction of QI ≥ 80 for compatibility with Meteosat-9, increased by 40% at low level, and a factor of 2 to 3 at middle and high levels compared to vF08_0017. As a result, vF01_0001 SMVs achieved better coverage than Meteosat-9 CMVs at low, and, especially middle level. High-level geostationary winds, however, retained their superior coverage, probably due to the inherent difficulties in stereo matching oblique and nadir MISR views of cirrus and deep convective clouds. I also found that the new quality control scheme best captured retrieval uncertainties in high-level clouds, while the choice of QI threshold had relatively little effect on SMV-CMV comparison statistics in low- and middle-level clouds.

[72] Upgrades had an overall neutral impact on E-W wind comparison statistics, which were already quite good in vF08_0017. N-S wind comparison statistics, however, improved significantly on global, zonal, and regional scales, and throughout the entire atmosphere. The negative MISR N-S wind bias was practically eliminated at low level and reduced by 2.5 to 3.0 m s−1 at middle and high levels. These bias improvements were accompanied by RMSD reductions of similar magnitude, and correlation increases of 10 to 20% or more, especially at middle and high levels and in tropics and subtropics (30°S–30°N). The vF01_0001 subpixel registration corrections had their biggest positive impact in the eastern half of the MISR swath, where biases were largest in vF08_0017 data, reducing amplitude of N-S wind and height cross-track variations by half, to 1.4 m s−1 and 120 m. As the net effect of these improvements, N-S wind errors became comparable to E-W wind errors at low and middle levels, and zonal wind remained slightly more accurate than meridional wind only at high level.

[73] Although the vF01_0001 algorithm upgrades had an overwhelmingly positive effect on N-S wind retrievals, the low-level N-S component bias slightly increased in two specific regions. For the oceanic paths 207–213, bias increase was likely due to sampling issues arising from more frequent orbit-level geolocation uncertainties; it remains to be seen if paths with few land tie points for georegistration are similarly affected over the Pacific. In the Gulf of Guinea this explanation does not hold; however, low-level Meteosat-9 winds also show large observation-model differences, and, as a result, are often blacklisted in this region. Therefore, increased N-S wind difference here might actually have been the result of comparing improved SMVs to biased CMVs.

[74] Despite substantial reduction in N-S wind errors, which are highly correlated with stereo height errors, the MISR – Meteosat-9 height comparison did not improve. Heights were, on average, 200 to 300 m lower in vF01_0001 than in vF08_0017, due to much-reduced negative N-S wind biases, but otherwise height differences remained similar to previous results, pointing to CMV height assignment errors as the primary driver of discrepancy. Attribution of the remaining wind and height differences to either MISR or Meteosat-9 retrieval errors will ultimately require a triple collocation approach [Stoffelen, 1998], adding a third independent wind data set to the comparison.

[75] The new algorithm's potential for retrieving wind fluctuations at scales even finer than the operational 17.6 km was also demonstrated. Experimental 4.4 km SMVs allowed a reasonable description of the fine structure of an island wake vortex street including its vorticity field, in good quantitative agreement with idealized large eddy simulations. Such high-resolution stereo wind and height retrievals might be useful for in-depth dynamical case studies of complex atmospheric flows.

[76] The MISR instrument has been retrieving height-resolved winds continuously for over a decade and is still in excellent health. Its carrier, the Terra satellite, has sufficient propellant to keep a nominal 705 km 10:30 A.M. equator crossing time orbit through spring 2018 (E. Moyer and D. Diner, personal communication, 2013). Science operations can continue at 705 km altitude through June 2020 by allowing a drift in equator crossing from 10:30 A.M. to 10:15 A.M. Options are also being evaluated to extend operations even beyond 2020, when Terra could exit the morning satellite constellation and settle at a lower orbit. Thus, the MISR mission is expected to continue for several more years, eventually producing a 20-plus year climate data record. The current latency between data acquisition and product availability is 12 h, too long for routine data assimilation purposes. Experiments have, however, shown the possibility of a 5 h latency or better, which might be adequate for certain NWP applications. Studies investigating the impact of stereo winds on forecast skill have already started.


[77] This work was partially funded by the Hans Ertel Centre for Weather Research (HErZ) initiative of the German Weather Service (DWD). I am indebted to Kevin Mueller, Catherine Moroney, and Veljko Jovanovic of the Jet Propulsion Laboratory, California Institute of Technology, for help with the operational MISR data set as well as producing the experimental high-resolution retrievals of the vortex street case study. I also thank Katrin Lonitz of the Max Planck Institute for Meteorology, Hamburg, for computational support. Finally, the suggestions of three anonymous reviewers greatly improved the paper.