Datasets
This work uses the MISR Cloud Motion Vector (CMV) product (version F02_0002)10 over the period March 2000 to December 2020. The product is not a usual gridded, monthly mean product normally used in climatological studies. Instead, the product contains a simple list of all CMV retrievals for a given month, with each retrieval tagged by latitude, longitude and time. The height-resolved CMVs are obtained through stereoscopic means by tracking the progression of features in the MISR 275-m resolution red-band imagery (380-km swath) over a 3.5-min period between the initial 70° forward view and the nadir view, and again for the 3.5-min period between the nadir view and 70° aft view10. The resolution of the MISR CMV product is 17.6 km × 17.6 km. Our ***ysis uses only the daytime descending node of the MISR orbit to keep local time consistent within high-latitude grids. The latest versions of MISR cloud-top heights and CMVs have been extensively validated24,25,26. The near-global validation of cloud-top height has been validated against a space-based lidar26, showing a bias ± precision of −280 m ± 370 m. The precision in cloud motion speed is 3.7 m s−1, with biases in U = 0.0 m s−1 and V = 0.3 m s−1 relative to static ground targets and with biases in U and V relative to geostationary-derived CMVs <±0.5 m s−1 (for cloud-top heights at which they have moderate agreement) and possibly up to −1.5 m s−1 for the V component, depending on the method of blockessment24,25. The stability of the product is also relevant for trend ***ysis. Although MISR geometric telemetry needed for stereoscopic retrievals indicate no trends over the operation of the mission (Veljko Jovanovic, personal communication), we nonetheless perform here the first ***ysis to quantify its stability. The MISR stereographic retrievals are agnostic to the texture being observed, be it from cloud or land surfaces, receiving no prior. Therefore, we use the surface as a stable target for measuring the stability of the MISR TC_Cloud_F0_0001 product22, which is the main input to the CMV product. We ***ysed cloud-top height and wind retrievals from data flagged as ‘high-confidence near-surface’ by the Stereoscopically-Derived Cloud Mask in the TC_Cloud product, which typically indicate clear sky or the occasional near-surface cloud. Our ***ysis encompblockes 20 years of global land data between 50° N and 50° S as in ref. 26 We conducted a trend ***ysis on the modes (rather than mean to avoid any possible trend in near-surface clouds) of annual histograms of these retrievals. For the surface heights, the trend is small at 0.54 ± 2.5 m decade−1 (95% CI) and insignificant (P = 0.94). Near-surface wind retrievals also exhibit negligible trends: the U component shows a trend of 0.00 ± 0.01 m s−1 decade−1 (P = 0.94) and the V component indicates a trend of 0.02 ± 0.05 m s−1 decade−1 (P = 0.51). These results confirm the long-term stability and reliability of MISR stereo measurements for climate research.
For the re***ysis model dataset, this study uses hourly data of the fifth-generation European Centre for Medium-Range Weather Forecasts (ECMWF) Atmospheric Re***ysis (ERA5)12. We use the U and V components of ERA5 wind, as well as the geopotential, on all of the available 37 pressure levels ranging from 1,000 hpa to 1 hpa. These hourly data are downloaded at a global 0.25° × 0.25° latitude–longitude grid, the highest spatial and temporal resolutions available for the ERA5 archive. Although the quality of ERA5 winds is evaluated against the MISR in the main text, we provide further discussion on ERA5 wind evaluation against other independent datasets below in ‘Extended data ***yses and discussion’, showing excellent agreement with the MISR in the very limited regions on which the other datasets report.
Sampled ERA5 at the MISR CMV record (ERA5_MS)
Because the ERA5 data have a spatial-temporal resolution that is comparable with the MISR CMV, we use a nearest-neighbour approach to sample the ERA5 U and V components at the time, location and altitude of each CMV retrieval. The time, latitude and longitude for each MISR CMV retrieval at a 17.6 × 17.6-km resolution are used to find the closest hour and the nearest grid point of the ERA5 data. For a CMV retrieval at a specific height, we locate the nearest ERA5 data point using the geopotential information at pressure levels. Specifically, geopotential heights are calculated by dividing the geopotential values by the Earth’s gravitational acceleration, given by 9.80665 m s−2 (constant). Hence, the sampled ERA5 data (ERA5_MS) have the exact same record length as the MISR CMV data. Wind speed is calculated from the U and V components for each record of CMV and ERA5_MS.
Trend ***ysis
Before trend ***yses, the U, V and wind speed data of MISR CMV and ERA5_MS are first aggregated into monthly, 0.25° × 0.25° latitude–longitude grid boxes. The aggregated data are then sorted into 20 height bins ranging from 0 to 20 km with a bin width of 1 km (with closed left side and opened right side). The mean of all of the 17.6-km retrievals in each grid box and height bin is calculated and stored in an intermediate file, along with the number of the retrievals for each bin. Hence, in one monthly intermediate file, U and V are stored in 720 (latitude) × 1,440 (longitude) × 20 (altitude) bins. For the zonal ***ysis (Fig. 1), the total number of bins is further reduced to 720 (latitude) × 20 (altitude) bins by averaging the data along the longitudinal dimension excluding the bins with no valid retrievals (for example, owing to high-altitude terrain lying above, say, the 0–1-km altitude bin). The zonal map of the total number of CMV retrievals is given in Extended Data Fig. 2.
To ensure a large sample size, only bins that have a total of more than 5,000 CMV retrievals over the 2000–2020 period are used in the zonal ***yses. This effectively removed the low-sample observations of the stratospheric clouds and the blockociated wind speed, thus keeping the focus of our discussion to the troposphere. As a reference for readers, the mean tropopause heights are plotted in Fig. 1. The mean tropopause heights were derived for the period 2000 and 2020 using the tavgM_2d_slv_Nx product of the Modern-Era Retrospective ***ysis for Research and Applications, Version 2 (MERRA-2)38, as tropopause heights are not directly available in ERA5.
The deseasonalized monthly anomalies for each bin are calculated as the deviation from the monthly means averaged over the 2000–2020 period. Trends ***ysis of the deseasonalized anomalies is conducted by initially applying the nonparametric Mann–Kendall test for the trend and the nonparametric Sen’s method for the magnitude of the trend using the Python package pyMannKendall39. In all ***yses and figures involving trend ***ysis, after performing local significance tests, we applied a tightened false discovery rate (FDR) correction, following the same procedure described in ref. 40. We aimed to control the FDR at a nominal level of 5%; however, the significance threshold was adjusted on the basis of the estimated proportion (50%) of true null hypotheses to maintain this control, as recommended in ref. 40. This adjustment often resulted in a higher nominal significance threshold, increasing the power to detect true effects while ensuring that the expected proportion of false discoveries remained at or below 5%. Grid points with adjusted P values below the adjusted FDR threshold were considered statistically significant.
Calculation of Lat_U0 and Lat_Umax
This study evaluates the tropical width and position of the polar jets using the zonal averages of the U data in the CMV, ERA5_MS and ERA5_AW. We use two metrics from the Tropical-width Diagnostics (TropD) software package41. Using TropD allows for consistency with other studies. Lat_U0 is the latitude at which the zonal-mean U component of the wind in the 0–1-km altitude bin equals zero after linear interpolation between two neighbouring latitude bins. This marks the latitude in the subtropics at which U switches sign from negative (easterly) to positive (westerly). It is calculated using the TropD_Metric_UAS module in TropD with default settings. Lat_Umax is the latitude of maximum zonal-mean U in the 1–2-km altitude bin. It is calculated using the TropD_Metric_EDJ module in TropD. We use the ‘peak’ method (weak smoothing) with the smoothing parameter of n = 6 as recommended in other studies42,43. The other parameters in the module are set with default settings.
Extended data ***yses and discussion
Examination of confounding factors
There are potential confounding factors at play when interpreting trends in MISR CMV as trends in speed, namely trends in the number of MISR CMV samples, their within-bin heights in the presence of within-bin vertical gradients in wind speed and their within-bin longitudes in the presence of within-bin horizontal gradients in wind speed.
Extended Data Figure 2 shows the number of CMV samples and their trends in terms of the within-bin percentage change. We see that the observed statistically significant trends in CMV samples are very small, mostly ranging from −0.6 to +0.2% decade−1. There is a decreasing fraction of CMV samples in the upper troposphere and an increasing fraction in the lower troposphere. These should not be compared with cloud-cover changes because a CMV retrieval is not sensitive to the underlying cloud fraction (that is, whether the 17.6 × 17.6-km area is 100% cloudy or 5% cloudy, we still get a CMV sample). Moreover, the positive trend in the lower troposphere is confounded by the decreasing trend in the upper troposphere, as less clouds above leads to more opportunity to retrieve clouds below. The decreasing trend in the upper atmosphere may be related to decreases in the frequency of occurrence of optically thin cirrus that reside near the detectability threshold of MISR stereo26,44. Note that the spatial patterns in the small trends shown in Extended Data Fig. 2b do not match the spatial patterns we see in the trends in Fig. 1a,d,g, which does not support the notion that sample trends alone can explain the trends seen in Fig. 1. Moreover, these small trends in sample numbers would have no impact on trends in cloud-top-conditioned winds without a corresponding shift in the CMV height and longitudinal distributions within the 1-km bin, which we examine next.
The relative change in the CMV height distribution within a 1-km altitude bin and how the heights and winds covary within an altitude bin can produce confounding effects in interpreting MISR CMV trends reported in Fig. 1 as trends in speed. Extended Data Fig. 3 shows the within-bin mean CMV height trend. We see some statistically significant trends that are small, mostly in the 0 to ±40 m decade−1 range. If we consider a moderately large gradient in wind speed with altitude of 5 m s−1 km−1 in the free troposphere (see Fig. 1 mean values), then we estimate 5/1,000 m s−1 m−1 × ±40 m decade−1 = ±0.2 m s−1 decade−1 as an extreme influence of this effect on CMV trends. This is small relative to the wind speed trends discussed with reference to Fig. 1. Moreover, the CMV heights and winds within a 1-km bin are not well correlated, with correlation coefficients <|0.2| for all bins (figure not shown). The poor correlation is as expected because (1) the uncertainty in MISR heights is only about twice as small as the bin width and (2) the uncertainty in the MISR winds is about the same value as we would expect in wind speed changes over a 1-km depth. These two facts were the primary motivators for choosing the 1-km vertical bin width to begin with for our ***yses. Also, the spatial patterns in Extended Data Fig. 3 do not match the spatial patterns we see in the trends in Fig. 1a,d,g. Therefore, there is no support that the large MISR CMV trends in Fig. 1 are greatly affected by the confounding effects of changing cloud heights and their covariability with wind within a 1-km altitude bin.
Finally, a longitudinal shift of the MISR CMV samples to a region of different large-scale circulation (for example, a shift from the jet entrance towards the jet core) may also be a confounding factor, even if the large-scale atmospheric circulation does not have a notable trend. We examined whether there are any substantial trends in the centroid of the longitudinal distributions of the CMV samples for each latitude/altitude bin and found that few regions have notable trends (Extended Data Fig. 4), and when they did, these regions do not completely overlap with those shown in Fig. 1. Hence, the trends shown in Fig. 1 cannot be simply attributed to longitudinal shifts in CMV samples.
In summary, the confounding factors discussed above are small or cannot be used to explain the CMV changes in Fig. 1a,d,g. Therefore we cannot reject the notion that the observed MISR CMV changes are mostly attributed to changes in the cloud-top-conditioned atmospheric circulation.
An ERA5 cloud conditional ***ysis
A non-random sample of the true wind field and its comparison with the same samples reported in ERA5 is sufficient to indicate uncertainty in ERA5 winds, but not a full characterization of the ERA wind uncertainty, as the samples are non-random. This statement is true regardless of the conditioning (for example, true cloud tops only) placed on these non-random samples. These samples could be further examined to help diagnose problems within ERA5 (for example, whether ERA5 placed a cloud top in the right spot). Similarly, MISR CMVs are non-random samples conditioned to observed cloud tops. Differences between MISR CMV and ERA5_MS winds (that is, Fig. 1) would indicate uncertainty in ERA5_MS winds in regions in which differences are much larger than the uncertainty in MISR CMVs, as quantified in Methods. This is true regardless of the cloud-conditioned nature of MISR CMV samples. As discussed in the main text, such notable differences were only observed in certain regions of the upper troposphere.
As a diagnostic, the reader may be curious as to whether these ERA5_MS samples are also ERA5 samples of cloud top. We extract the ERA5 Fraction of Cloud Cover parameter blockociated with each ERA5_MS wind sample. In a sample-by-sample comparison, we find that 71.2% of the total ERA5_MS samples have a cloud cover > 0 at the altitude of the ERA5_MS sample; the remaining 28.8% are clear (that is, cloud cover = 0). We also use more strict criteria for the ERA5_MS to contain a cloud top: (1) ERA5_MS cloud cover > 0 at the altitude of the ERA5_MS sample and (2) there are no ERA5 clouds above this altitude. Using these criteria, we find that only 10.5% of the total ERA5_MS samples have a cloud top at the same altitude as the MISR CMV. This is stricter than it needs to be because MISR stereo can see through optically thin clouds to retrieve a lower cloud without any degradation in the quality of the retrieval26. Still, the difference between 10.5% and 71.2% is much more than can be explained by the frequency of observed thin high cloud over thicker lower cloud45. Regardless, when we recreated the Fig. 1 ERA5_MS ***ysis separately using the 10.5% cloud top, 89.5% non-cloud top, 71.2% cloud and 28.8% clear ERA5_MS samples, we found that their differences are not statistically different (95% CI) between each other or against Fig. 1b,e,h.
These results provide strong evidence that MISR CMVs can be used to evaluate ERA5 winds at the times and locations of MISR CMV sampling, regardless of whether ERA5 indicates the presence of a cloud (or cloud top) or not. The results are symptomatic of a large uncertainty in the ERA5 parameterization of cloud physics, particularly in how it relates to the coupling of clouds and circulation. Global models used to generate ERA5 have limitations in fully resolving the fast, fine-scale dynamics in observed moist convection46, even though the blockimilation of vast amounts of data effectively constrains the state of slow-varying, large-scale circulation. Although MISR CMVs are inherently tied to observed clouds and could reflect changes in moist convective systems (for example, shifts in intensity or organization affecting cloud motion), ERA5’s parameterized convection and its representation of blockociated dynamical features such as divergence/convergence may not fully capture these nuances. This could mean that climate-driven changes in moist convection, if present in MISR CMVs, are challenging to isolate or validate using present global re***yses.
Comparison with other works
It is instructive to compare differences in ERA5_MS wind and MISR CMV reported here to differences in ERA winds against other observations reported in other studies. This is done to gain confidence in our ***yses and those reported in other studies.
In one study47, satellite altimeter and scatterometer data were used to validate ERA5 surface winds (10 m) over the Atlantic between 60° N and 60° S. Over this region, they show that ERA5 has zonal surface wind speed relative biases that vary latitudinally between 0 and 0.8 m s−1. Other studies48,49 have compared ERA5 surface winds to land surface station data, most of which were equatorward of 60° latitude. These land station comparisons indicated mean absolute difference with ERA5 surface winds <0.4 m s−1. These results are in line with ERA5_MS biases relative to MISR CMV for the lowest 1-km bin, with results varying latitudinally (and averaged over all longitudes) within the range −0.2 to +0.8 m s−1 between 60° S and 60° N. If we restrict ourselves to 35 to 60° N, at which we have a dense network of land surface stations49, then the latitudinally varying surface wind speed relative biases in this latitude band between MISR CMV and ERA5_MS range from −0.1 to +0.1 m s−1. This improvement is expected given: (1) the dense global network of station data that is blockimilated in ERA5 over land within this latitude range and (2) the high accuracy of the MISR CMV product. Over ocean, however, few surface stations data are blockimilated into ERA5, so the larger ERA5 wind biases relative to altimeter, scatterometer and MISR data makes sense.
For winds above the surface, this study is the first validation of ERA5 tropospheric winds (cloud-top-conditioned or otherwise) over the globe based on independent observations. However, one study50 using Aeolus51 data over one rawinsonde station in Singapore also evaluated the ERA5 winds. Aeolus is a Doppler wind lidar capable of deriving vertically resolved, zonal winds (that is, U). Using data between 2019 and 2021, they show that the height-resolved, mean zonal winds measured by Aeolus is within ±1.5 m s−1 of ERA5 between the surface and 14 km. Above 14 km, ERA5 reaches a maximum bias relative to Aeolus of +3.5 m s−1 at an altitude of 16.5 km (that is, near the tropopause). We extracted 20 years of MISR CMV U component over Singapore and it showed very similar results, despite being cloud-conditional: within ±1.0 m s−1 of ERA5_MS between 0 and 14 km, with a maximum relative bias of +3.2 m s−1 also at 16.5 km. The similarities are remarkable, which speaks to the very high quality of both Aeolus Doppler winds and MISR CMVs, as well as to the high quality of ERA5 winds at altitudes in the lower to middle troposphere, at least at this tropical location. That study was able to attribute the large relative bias near the tropical tropopause to the poor representation of Kelvin wave dynamics in ERA5, in which re***yses are known to struggle52. The positive impact that the blockimilation of global Aeolus winds had on numerical weather prediction model forecasts, including ECMWF53,54, is further evidence that modelled winds still have room for improvements, particularly in the upper troposphere (that is, where the mean MISR CMV show the largest disagreement with ERA5_MS in Fig. 1).
A comprehensive comparison of the time series of ERA5 tropospheric winds against independent satellite data does not yet exist—the results here with MISR are a first. A time-series ***ysis with satellite scatterometers is challenging because of the different instruments with different orbit (and orbit drifts) that need to be stitched together. In one study55 that used a blended method with other data to help with some shortcomings in the satellite data, they show trends of surface winds over ocean between 60° N and 60° S between 1992 and 2012. Their results show latitudinal variability in zonal mean trends ranging between −0.2 and +0.2 m s−1 decade−1. In the case of MISR CMVs, ERA5_MS and ERA5_AW, few latitude bins show statistically significant trends in the surface (0–1-km) bin, and where they do, the trends range between −0.2 and +0.2 m s−1 decade−1 (Fig. 1). This is similar to the scatterometer study, recognizing the caveat in the comparison owing to differences in time periods and ocean only.
On the basis of the above comparisons with other studies, we find similar relative biases with ERA5 winds as those reported using the MISR for the very limited regions of the troposphere that these studies cover. These comparisons, along with extensive validation of MISR CMVs that show a highly accurate and stable dataset (see Methods), supports the conclusion that ERA5_MS winds are insignificantly different from MISR CMV in the lower to middle troposphere and have small but significant differences in the upper troposphere, as described in the main text. Discrepancies between MISR CMVs and ERA5_MS, particularly in the tropical upper troposphere, resonate with previously reported limitation in the ECMWF model and ERA5 performance. Independent blockessments have identified substantial biases in the ECMWF model’s representation of tropical tropopause wind shear relative to radiosondes56, errors in ERA5 tropical ocean surface winds57 and persistent errors in tropopause-level wind shear in re***yses58. These findings reinforce the MISR-based evidence that re***ysis wind remains uncertain in dynamically complex regions, particularly in the tropics and near the tropopause, and they highlight the value of using independent, high-quality observations such as MISR CMVs for further model refinements.
Seasonal variability of MISR CMV and ERA5 winds trends
Modelling studies have shown that the patterns and drivers of long-term circulation changes have some seasonality14,59. Hence, we have compared the decadal trends of seasonal means in height-resolved winds (Extended Data Figs. 5–8) against the trends in deseasonalized monthly anomalies (Fig. 1). There is general agreement between the two patterns of trends, except that the level of significance is reduced in seasonal trends, as each seasonal plot has only a quarter of the total data used in Fig. 1. There are only two notable exceptions to this broad agreement.
The first exception is in the strengthening of the U component of the winds along the polar front in the SH seen in Fig. 1—this feature largely disappears in boreal winter (December, January and February) for all three datasets. During December, January and February, stratospheric ozone depletion over the Antarctic regions has been attributed as a mechanism for enhanced poleward shifts in the eddy-driven jet and the SH Hadley cell edge in climate models. This enhanced poleward movement would result in a more meridional flow of wind than zonal in the polar jet and could probably explain the lack of strengthening in the U component over these months.
The second is in the presence of substantial strengthening of the U component in the subtropical jet of the SH seen in the ERA5_MS data but not in the CMV—it is largely absent in the ERA5_MS in the boreal winter (December, January and February), weakened in boreal summer (June, July and August) and very strong in the boreal spring and fall seasons (March, April and May and September, October and November, respectively). Apart from these two exceptions, the lack of strong seasonality in the trends in Fig. 1 implies that whatever is driving the trends is doing so regardless of seasonal forcing.
