2.1 Reanalysis data

In this study we use two modern reanalyses: the ECMWF reanalysis, ERA-Interim (Dee et al., 2011), and the NASA (National Aeronautics and Space Administration) reanalysis, MERRA-2 (Gelaro et al., 2017). ERA-Interim is the successor of ERA-40 and benefits from many improvements, especially a variational bias-correction scheme which reduces the impact of observing system changes. It is thus expected that trend estimates are more realistic. MERRA-2 is an update of MERRA which benefits from recent developments in NASA's Goddard Earth Observing System (GEOS) model suite, intended to address the impact of the changes in observing system (Gelaro et al., 2017). As a result, atmospheric water balance and variability in MERRA-2 are more realistic, though variations of IWV with temperature are weaker in the main satellite data reanalyses (namely ERA-Interim and MERRA-2) compared to microwave satellite observations over the oceans (Bosilovich et al., 2017).

Data from both reanalyses were extracted for the common 1980–2016 period, on regular latitude–longitude grids, at their highest horizontal resolution (0.75^∘ × 0.75^∘ for ERA-Interim and 0.625^∘ longitude × 0.5^∘ latitude for MERRA-2). In this work, the two-dimensional (2-D) distribution of IWV is investigated with reanalysis fields and with point observations from GPS data. Because GPS antenna heights and surface heights in the reanalyses are not perfectly matched (see the GPS coordinates and the heights for both reanalyses in the Supplement Table S1), the IWV estimates were adjusted for the height difference Δh based on an empirical formulation derived by Bock et al. (2005): $\mathrm{\Delta }\mathrm{IWV}=-\mathrm{4}\times {\mathrm{10}}^{-\mathrm{4}}\times \mathrm{IWV}\times \mathrm{\Delta }h$, with Δh in metres. The advantage of this formulation is that it can be applied directly to the IWV data without requiring any auxiliary data. Its accuracy was checked compared to a more elaborate formulation using ERAI pressure level data (see Appendix B). The corrections were thus computed for each reanalysis separately using the monthly IWV data from the nearest grid point to every GPS station.

2.2 GPS data

We used the reprocessed tropospheric delay data from GPS stations of the International GNSS (Global Navigation Satellite System) Service (IGS) network (Fig. 1). The data was reprocessed by the NASA Jet Propulsion Laboratory (JPL) in 2010–2011 and is referred to as IGS repro1. Basic details on the processing procedure are described by Byun and Bar-Server (2009) for the operational version at the time. The reprocessed data set was produced with more recent observation models (e.g. mapping functions, absolute antenna models) and consistently reprocessed satellite orbits and clocks (IGSMAIL-6298, 2012). Inspection of file headers revealed that the processing options were not updated for a small number of stations for a period of nearly 1 year between March 2008 and March 2009. The comparison of solutions with old and new processing options (available for the year 2007) showed that this inconsistency in the processing has negligible impact at most stations, except for stations at high southern latitudes (e.g. in Antarctica). The data set covers the period from January 1995 to December 2010 for 456 stations. Among these, 120 stations have time series with only small gaps over the 15-year period. However, the geographical distribution is quite unequal between hemispheres and even within a given hemisphere, with a cluster of 20 stations in the western USA with inter-station distance smaller than 0.75^∘. In order to avoid over-representation of this region, 16 out of these 20 stations have been discarded (the selection retained those with the longer time series). The final GPS IWV dataset used in this study is thus limited to the selected 104 stations.

The basic GPS tropospheric observables are the zenith tropospheric delay (ZTD) estimates, which are available at a 5 min sampling in the IGS repro1 dataset. The GPS ZTD data were screened using an adaptation of the methods described by Bock et al. (2014, 2016). First, we applied a range check on the ZTD and formal error values using fixed thresholds according to the spatial and temporal range of expected values: 1–3 m for ZTD and 0–6 mm for formal errors. Second, we applied an outlier check based on site-specific thresholds. For ZTD, values outside the median ±0.5 m were rejected and, for formal errors, values larger than 2.5 times the median were rejected. The thresholds were recomputed for every year. Using these thresholds, we detected no ZTD values outside the limits. This is because the limits were sufficiently large to accommodate for the natural variability of ZTD values (Bock et al., 2014). On the other hand, the formal error check rejected $\mathrm{8.8}\times {\mathrm{10}}^{-\mathrm{4}}$ (i.e. less than 0.1 %) of the data overall. After screening, the 5 min GPS ZTD data were averaged in 1-hourly bins.

The conversion of GPS ZTD to IWV was done using the following formula: IWV = ZWD × κ (T_m), where κ (Tm) is a function of weighted mean temperature Tm, and ZWD is the zenith wet delay, obtained from ZWD = ZTD − ZHD, and ZHD is the hydrostatic zenith delay (see Wang et al., 2005 or Bock et al., 2007 for further details). In this work, the surface pressure used to compute ZHD and the temperature and humidity profiles used to compute Tm were obtained from ERA-Interim pressure level data. The profile variables are first interpolated or extrapolated to the height of the GPS stations at the four surrounding grid points and then interpolated bi-linearly to the latitude and longitude of the GPS stations. At this stage, the 1-hourly ZTD GPS data and the 6-hourly ERA-Interim data (ZHD, Tm, and IWV) were time-matched within ±1 h. Afterwards, monthly means of the 6-hourly IWV estimates are computed and those months which have less than 60 values (i.e. at least half of the expected monthly values) are rejected. Seasonal means are computed from the monthly values when at least 2 out of 3 months are available. These selection criteria ensure that the computed values are representative of the monthly and seasonal means.

Figure 2(a) Mean IWV for DJF 1995–2010 from ERA-Interim (shading) and GPS (filled circles); (b) same as panel (a) for JJA. (c) Relative variability in percentage (%) (standard deviation of the IWV series divided by its mean) for DJF 1995–2010; (d) same as panel (c) for JJA.

Download

In this work, inhomogeneities in the GPS IWV time series due to equipment changes were not corrected a priori, as the existing metadata may not be complete, but were rather discussed based on the intercomparison with ERA-Interim (see Appendix B). Detection and correction of inhomogeneities in the GPS ZTD and IWV data is a subject of ongoing research.

2.3 Computation of trends

The linear trends were computed using the Theil–Sen method (Theil, 1950; Sen, 1968), a non-parametric statistic that computes the median slope of all pairwise combinations of points. This method is described in more detail in the Appendix A to this paper, where it is also compared with another commonly used method for trend estimation, the least squares method.

Figure 3(a) Difference of mean IWV estimates (MERRA-2 minus ERA-Interim) for DJF 1995–2010. The global mean difference is 0.35 kg m⁻² (0.94 %) and the standard deviation of the difference is 0.32 kg m⁻² (1.56 %). (b) Same as panel (a) for JJA. The global mean difference is 0.56 kg m⁻² (1.66 %) and the standard deviation of the difference is 0.37 kg m⁻² (1.67 %). (c) Difference of relative variability estimates (MERRA-2 variability minus ERA-Interim variability) for DJF 1995–2010. The global mean difference is −0.01 % and the standard deviation of the difference is 0.26 %. (d) Same as panel (c) for JJA. The global mean difference is 0.28 % and the standard deviation of the difference is 0.42 %.

Download

The Theil–Sen method was applied to the anomalies obtained by removing the monthly climatology from the monthly data. In the case of seasonal trends, the mean anomalies for the months of December, January, and February (DJF) and June, July, and August (JJA) were used. The statistical significance of the monthly and seasonal trends was assessed using a modified Mann–Kendall trend test (Hamed and Rao, 1998), which is suitable for autocorrelated data, at a 10 % significance level.

Figure 4(a, b) Difference of mean IWV estimates (ERA-Interim minus GPS) for DJF and JJA 1995–2010; (c, d) same as panels (a) and (b) for MERRA-2 minus GPS. The monthly data were time-matched before seasonal means were computed. (e, f) Difference of relative standard deviations of IWV estimates (ERA-Interim SD minus GPS SD.) for DJF and JJA 1995–2010; (g, h) same as panels (e) and (f) for MERRA-2 minus GPS. The monthly data were time-matched before seasonal means were computed.

Download

Table 1Statistics (median ± interquartile range) for the differences (ERA-Interim minus GPS and MERRA-2 minus GPS) at the 104 stations, divided by season and hemisphere, for the differences in IWV mean, relative standard deviation, and trends (ERA-Iterim).

Download Print Version | Download XLSX

3.1 Means and variability in IWV

The mean seasonal IWV distribution and its inter-annual variability for DJF and JJA are presented in Fig. 2. In the maps of the means (Fig. 2a, b), we can see that ERA-Interim reproduces the spatial variability well compared to GPS, including the sharper gradients in IWV, for instance, on the northern and southern flanks of the intertropical convergence zone (ITCZ) in both seasons, and in the regions of steep orography (for example, along the Andes region, in South America). Similar mean patterns are observed in MERRA-2 for both seasons, although maximum values over the ITCZ have different intensities (not shown). In order to better gauge the differences, mean difference fields between MERRA-2 and ERA-Interim are shown (Fig. 3a and b). It is observed that ERA-Interim is generally drier than MERRA-2 over the ocean and in the regions of maximum IWV, and moister in the southern part of South America, north and south of the ITCZ over Africa, and in southern (central) Asia in DJF (JJA). This result is consistent with differences between ERA-Interim and MERRA reported by Trenberth et al. (2011).

Comparing the reanalyses with GPS in Fig. 4a–d shows that both reanalyses are too moist in the Northern Hemisphere in both seasons. ERA-Interim has a dry bias at southern tropical sites in DJF and northern tropical and mid-latitude sites in JJA (especially over the USA). MERRA-2 has a moist bias overall but this is more pronounced in the Southern Hemisphere in DJF, especially over Australia, and in the Northern Hemisphere in JJA (see the increased bias over North America, northern Europe and Asia). Hemispheric and seasonal statistics can be found in Table 1 which show that in the Northern Hemisphere, the median biases of the reanalyses are very consistent in DJF (0.5–0.6 kg m⁻²) while in JJA the bias in MERRA-2 (1.1 kg m⁻²) is more than double the bias in ERA-Interim (0.48 kg m⁻²). In the Southern Hemisphere the main difference is for the DJF bias, which is nearly zero in ERA-Interim and 0.7 kg m⁻² in MERRA-2. Similar conclusions can be drawn from the relative biases. High consistency is also seen in dispersion of the biases (inter-quartile range) between both reanalyses, with larger absolute values in the summer hemisphere and larger relative values in the winter hemisphere.

Note that in Fig. 4a–d GPS estimates at a number of sites show large biases with respect to both reanalyses. These sites are generally located in coastal regions and/or regions with complex topography, where representativeness differences can be suspected (Lorenc, 1986). We applied a statistical test to detect stations with significant differences in the mean values (with 99 % confidence level). When GPS and ERA-Interim are compared, 21 stations are detected in DJF and 18 in JJA, while with MERRA-2, 26 stations are detected in DJF and 44 in JJA. These numbers confirm in a more objective way the visual interpretation of Fig. 4. The values of all statistics can be found in the Supplement (Table S2).

For the analysis of the inter-annual variability we computed the relative standard deviation of the seasonal IWV time series (i.e. standard deviation of seasonal time series divided by its mean value). The relative variability emphasizes regions where the mean IWV contents are small (e.g. cold dry polar and/or mountainous regions and warm dry desert areas). In DJF (Fig. 2c), strong inter-annual variability (>15 %) is found in ERA-Interim for northern high-latitude regions (north-eastern Canada and eastern Greenland, the polar Arctic area, and a large part of Russia and north-eastern Asia) and for the tropical arid regions (Sahara, Arabic peninsula, central Australia). There are similar patterns of variability in MERRA-2 (not shown), with slightly higher variability in MERRA-2 over India and Antarctica. The difference fields between MERRA-2 and ERA-Interim highlight these differences (Fig. 3c). In JJA, large inter-annual variability is observed in ERA-Interim mainly over Antarctica and Australia (Fig. 2d). Locally enhanced variability is also seen over the Andes cordillera, but this is mainly due to the very low IWV values at high altitudes. MERRA-2 has similar patterns of IWV variability, with lower variability over northern Africa and north of the Andes, and higher variability over Antarctica (Fig. 3d).

In general, most of the marked regional features of inter-annual variability are confirmed by GPS observations (Fig. 2c, d). However, a few stations show different values compared to the reanalyses, but their values do not impact the variability statistics shown in Table 1 thanks to the choice of median and inter-quartile range instead of mean and standard deviation. Based on the median values, Table 1 shows that in general ERA-Interim is in slightly better agreement with GPS than MERRA-2. It also shows that both reanalyses have lower or equal variability than GPS in both hemispheres and in both seasons (negative or zero median differences of relative standard deviation in Table 1). Figure 4e–h show the spatial distribution of the differences for both reanalyses and seasons. Again, there is quite a large spatial dispersion (also revealed by the inter-quartile range in Table 1) with a number of outlying sites discussed in Appendix B.

Representativeness errors due to large spatial variations in IWV and complex topography are suspected at 20 stations. They contribute likely both to differences in the mean IWV values and in the variability. Errors in the GPS data, e.g. due to instrumental malfunctioning, measurement interferences, or changes in equipment, are also suspected at a small number of sites. Such problems can be further confirmed by comparison with IWV measurements from nearby GPS receivers or from other collocated instruments such as DORIS or VLBI (Bock et al., 2014; Ning et al., 2016). Lastly, errors in the reanalyses are suspected in data-sparse regions and regions where the performance of model physics and dynamics are poor. They are more difficult to diagnose. The differences observed between ERA-Interim and MERRA-2 might be due to a mix of differences in model physics and data assimilation.

3.2 Trends in IWV

Trends from ERA-Interim and MERRA-2 based on the time series of monthly data (hereafter referred to as monthly trends) are shown in Fig. 5. In general, there is continuity between oceanic and continental trends, suggesting a trend in air mass advections.

Both reanalyses show overall positive trends (moistening) especially marked (statistically significant) in the tropics along the ITCZ, both over the oceans and continents (northern South America, central Africa, and Indonesia in ERA-Interim) and at middle and high latitudes in the Northern Hemisphere. The moistening trends over the Arctic are significant in both reanalyses. They are interleaved with extended regions of negative (drying) trends which are statistically significant in both reanalyses over Australia, northern–central Africa, over the south-tropical eastern Pacific region, and on the oceanic border of Antarctica. The moistening–drying dipole structure in the eastern equatorial Pacific has been observed by several authors using various satellite data and reanalyses over different periods (Trenberth et al., 2005; Mieruch et al., 2014; Schröder et al., 2016; Wang et al., 2016). Trenberth et al. (2005) explained it as being due to the influence of different ENSO phases over the trends. The monthly trends computed at the GPS stations are consistent in sign and magnitude with the reanalyses where the reanalyses agree, except at a small number of GPS sites discussed in Appendix B. The trend values for ERA-Interim, MERRA-2, and GPS at all stations can be found in the Supplement (Table S3). Statistics of differences in the trend estimates between the reanalyses and GPS are given in Table 1. The median differences are small (below ±1 % decade⁻¹) for both reanalyses and both hemispheres. The interquartile range of differences vary depending on the season and hemisphere but they are similar for both reanalyses. The monthly trends agree quite well in the Northern Hemisphere (∼2.6 % decade⁻¹ vs. ∼4 % decade⁻¹ in the Southern Hemisphere), while the seasonal trends have larger errors in the winter hemispheres (∼7% decade⁻¹).

Figure 5Relative IWV trends (in % decade⁻¹) for the 1995–2010 period from GPS (stations marked as circles), ERA-Interim (a), and MERRA-2 (b). The statistically significant trends from ERA-Interim and MERRA-2 are highlighted by stippling. Two regions discussed in Sect. 3.2 are outlined as box A (Maritime Continent) and box B (northern Africa).

Download

Significant trend differences between MERRA-2 and ERA-Interim are seen (Fig. 5) over several parts of the globe, in particular over the Maritime Continent, northern–central Africa, central Asia, and Antarctica. In these regions the reanalyses show opposite trends. Such a discrepancy can be due to different representations of large-scale moisture transport, surface–atmosphere processes, and data assimilation in the two reanalyses.

Over the Maritime Continent (box A, in Fig. 5), ERA-Interim trends are positive while MERRA-2 trends are negative over the period from 1995 to 2010. Trenberth et al. (2005) also reported strong moistening and warming trends over the region for the 1988–2003 period using SSM/I IWV data and National Oceanic and Atmospheric Administration (NOAA) sea surface temperature (SST) data. Wang et al. (2016) confirmed the moistening but found a near-zero–slightly negative temperature trend for the 1995–2011 period (they used more recent releases of the SSM/I and NOAA data). Schröder et al. (2016) also found moistening trends in all their datasets over the 1988–2008 period, including SSM/I data and the MERRA reanalysis. Although not all studies concern the same period, they confirm the ERA-Interim results. This conclusion points to an inconsistent drying trend in MERRA-2 over the Maritime Continent. Comparison to the GPS IWV trends in Fig. 5a confirms that ERA-Interim is in better agreement with the observations than MERRA-2, though the comparison is quite difficult because there are not many GPS stations available within the domain.

Over northern Africa (box B, in Fig. 5), the drying trend in ERA-Interim reaches an extremely large value of −3.5 kg m⁻² decade⁻¹ (−17 % decade⁻¹), which is questionable. Such a large trend would imply a significant change in the regional and global water cycle that can hardly be supported by physical arguments. Unfortunately, there are no long-term GPS data available in the region. The trends in MERRA-2 over the region are smaller and more realistic, but they also show a different spatial pattern which is questionable as well. The drying in MERRA-2 extends southward over equatorial Africa, a region where moistening is expected to follow the observed warming trends (see also Fig. 8). The difference between the reanalyses is further emphasized when comparing seasonal trends below. A major uncertainty in both reanalyses over this region is certainly due to the paucity in observations going into the assimilation. This statement is consistent with the results of Bauer (2009) and Karbou et al. (2010) who report a strong impact of humidity data from modern satellite instruments on the analysed moisture fields over the Sahara.

Figure 6Seasonal IWV trends for the 1995–2010 period from ERA-Interim (shading) and GPS (filled circles) for DJF (a). (b) Same as panel (a) but for MERRA-2. The statistically significant trends from the reanalyses are highlighted by stippling. (c) Same as panel (a) but for JJA. (d) Same as panel (b) but for JJA. The statistically significant trends from the reanalyses are highlighted by stippling.

Download

Over Antarctica, the monthly trends in MERRA-2 and GPS (Fig. 5b) are significantly positive, in opposition to what is seen in ERA-Interim (Fig. 5a), where the trends are mainly negative, especially in the interior of the continent. However, one can notice that ERA-Interim shows spotted areas of positive trends in the vicinity of the GPS stations. These locally positive trends in ERA-Interim might be explained by the influence of surface and/or upper air observations collected from these sites that are assimilated in the reanalysis. Comparing ERA-Interim and MERRA-2 IWV time series in the interior of the continent reveals that the reanalyses diverge mainly before year 2000, with a positive trend in MERRA-2 between 1995 and 2000 (not shown). This divergence might be explained by a combination of differences in the observations actually assimilated and differences in the assimilation systems. Observations in the interior of the continent are most likely from satellites only. General documentation indicates that both reanalyses use the same types of satellite observations globally (Dee et al., 2011; Gelaro et al., 2017). However, the amount of assimilated data over Antarctica's ice sheet may differ between the reanalyses, which is actually not documented.

Over central and east Asia and over northern South America, ERA-Interim and MERRA-2 trends also disagree. The GPS data are in better agreement with MERRA-2 in the former region and with ERA-Interim in the latter.

Figure 6 shows the seasonal trends. A striking feature seen in both reanalyses is their relatively larger magnitude compared to the monthly trends (Fig. 5), which could be due to the fact the monthly trends use one value per month, while the seasonal trends use only one value per year. Large changes in magnitude and/or sign are also noticeable in most regions between seasons. These features emphasize that atmospheric circulation (which is largely changing between seasons) plays an important role in IWV trends.

In DFJ (Fig. 6a, b), the agreement between reanalyses is surprisingly good, given the inconsistencies pointed out from the monthly trends. The agreement of the reanalysis with GPS is also quite good, though some GPS trends are in strong contradiction (e.g. at IRKT in Siberia, KIRU in Sweden, IISC in India, and COCO east of Indonesia, which have very large magnitudes). Over Antarctica, the drying–moistening east–west dipole is consistent in both reanalyses though they are of different magnitudes. A drying–moistening dipole is also seen across Australia, consistent with the theory that precipitation over north-western Australia (the part of Australia mostly influenced by the monsoon flow in DJF) is very sensitive to the SST pattern over the western central Pacific Ocean (10^∘ S–10^∘ N, 150–200^∘ E) (Brown et al., 2016). During the 1995–2010 period, this SST pattern has actually been warming and the atmosphere moistening, leading to a drying over north-western Australia (Wang et al., 2016).

In JJA (Fig. 6c, d), the conclusions are more contrasted. Though the reanalyses generally agree well over the oceans, except over the Maritime Continent, the trends over land are poorly consistent over most continents (Asia, Africa, South America, and Antarctica). Over these regions, the GPS trends are generally in better agreement with MERRA-2. Over northern Africa, the drying in ERA-Interim is in contradiction with the recent recovering of precipitation over western Africa (Sanogo et al., 2015). In this respect, the MERRA-2 trends are more realistic.

4.1 Global analysis

Interpretation of IWV trends of the previous section must be done cautiously as the trends have been estimated for a specific and rather short period of 15 years (from 1995–2010). They should thus not be considered as representative of a longer period and might also be impacted by decadal variability and/or large singular events such as El Niño. Trenberth et al. (2005) suggested that a longer time series may be required to obtain fully stable patterns of linear trends. In this section we will assess how consistent our trends obtained for the 1995–2010 period (when GPS data are available) are with longer-term trends computed for the full length common to ERA-Interim and MERRA-2.

Figure 7 shows the monthly trends for both reanalyses over the period 1980–2016. In both reanalyses most structures are similar to those seen for the short period (Fig. 5), although the intensities are generally weaker for the longer period (note that the colour bars are different for Figs. 5 and 7), but most of them are significant. In ERA-Interim, the main differences between periods (changes in sign of the trends) appear over the eastern tropical Pacific, Canada, the Arabic Peninsula, the region around Madagascar, Western Australia, and a small part of Antarctica. The drying trend over Australia observed for the 1995–2010 period is not observed in the long term, which suggests that there have been periods of moistening trends as part of decadal variability during the 1980–2016 period. These changes are seen in MERRA-2 as well, which gives good confidence that they are due to decadal variability in the global and regional climates. However, the main differences between reanalyses already highlighted for the short period remain (over Antarctica and the global southern oceans, northern Africa, and eastern Asia), except over the Maritime Continent where MERRA-2 now represents a moistening trend consistent with ERA-Interim and over Australia where the reanalyses now disagree.

Figure 7Monthly trends in IWV for the 1980–2016 period for (a) ERA-Interim and (b) MERRA-2. The statistically significant trends are highlighted by stippling.

Download

Figure 8Seasonal trends in Tm and IWV for the 1980 to 2016 period for (a, c, e, g) ERA-Interim and (b, d, f, h) MERRA-2.

Download

Figure 8 shows the seasonal IWV trends and temperature trends. In general, it is seen that over the oceans, the temperature trends generally have the same sign as the IWV trends (but opposite colours), as expected by Clausius–Clapeyron (C-C) theory, despite some small-scale differences. Over land, most areas show an increase in temperature, except the high latitudes of the Southern Hemisphere and large parts of the Asian continent. This means that, except over Antarctica and parts of Asia, the drying observed in the aforementioned areas does not follow C-C theory. When we consider each season more closely, some areas indicate a cooling (Fig. 8a, b, e, f) consistent with a drying (Fig. 8c, d, g, h). This is observed over Antarctica (especially in ERA-Interim) and over central Asia in DJF (especially in MERRA-2). For JJA, most continental areas show a significant warming, with the exception of parts of Antarctica, a small area over northern Australia, and regions in central Asia, where a cooling is also displayed. Thus the C-C scaling ratio is not a good proxy for humidity when considering seasonal and regional variabilities and trends due to the important role of dynamics which allow the advection of dry or wet air masses (e.g. over USA, South America, eastern Sahel, and southern Africa in JJA).

Figures 7 and 8 confirm that MERRA-2 presents a more general moistening trend than ERA-Interim (as already seen over the shorter period), especially in the Southern Hemisphere in DJF (Fig. 8c, d), and in both hemispheres in JJA (Fig. 8g, h). The main differences in the trends over the oceans appear all around Antarctica, and those over continental areas are observed over Africa (where trends are positive in the north and negative in central Africa in MERRA-2 and the opposite in ERA-Interim) and USA in JJA, over Australia in DJF, and over Antarctica in both JJA and DJF. Over Africa and Antarctica, the important differences which exist between ERA-Interim and MERRA2 for both long- and short-term periods suggest that the physical processes are not well represented. These areas correspond to areas with very few observations available for data assimilation, reducing the constraint on the models. A more detailed investigation of the dynamics over Africa and Australia is presented in the next subsections.

Figure 9Temperature and IWV anomaly time series for a box over Western Australia (see Fig. 14), using ERA-Interim (blue) and MERRA-2 (red) data, for (a, c) the 1980 to 2016 period, (b, d) the 1995 to 2010 period, and (a, b) the monthly time series and (c, d) the DJF season.

Download

Other regions, such as the Indo-Pacific region, have different trends over the shorter period, but are in better agreement over the longer period. This is more obvious during JJA and can be explained by the strong variability that requires longer time series in order to obtain meaningful trends. The good agreement between reanalyses over this area is an important result in view of the fact that CMIP5 (Coupled Model Intercomparison Project Phase 5) models have large biases over this region in present-day sea surface temperature, which has direct consequences on the future projection of precipitation over Australia (Brown et al., 2016; Grose et al., 2014) and more generally over tropical and subtropical climates. Two areas are investigated in more detail in the next subsections because of the disagreement between both reanalyses over them: Australia in DJF and Africa in JJA.

Figure 10Close-up of Western Australia showing the mean monthly and DJF fields and trends (contours) of the u and v wind components at 925 hPa (shaded). The area of focus (where IWV trends are most intense in ERA-Interim) is marked by a box.

Download

4.2 Analysis over Western Australia

Figure 9 displays the time series of IWV and temperature anomalies for a box over Western Australia (15–30^∘ S, 115–135^∘ E, as shown in Fig. 10) for both the short and long periods, for both the full time series and the DJF seasons.

As can be observed in Fig. 9a, b, the moisture trend is opposite for the long (moistening) and short (drying) periods, for both reanalyses, while the temperature trend is weakly warming. However, when focusing on the DJF period (Fig. 9c, d), the differences between reanalyses are enhanced when considering the long period. ERA-Interim IWV indeed starts with higher anomalies than MERRA-2 until 1990 and ends with lower anomalies after the late 2000s, so that the resulting trend is close to zero and not significant. The different IWV trend estimates between the two periods are due to the existence of extreme cold and humid periods in both reanalyses after 1992, with a strong occurrence around the 2000s, which impact the linear trend estimate over the short period (which starts in 1995) more strongly than over the long period.

The wetter and colder summers (correlation between T and IWV being around $-\mathrm{0.7}/-\mathrm{0.8}$ for both reanalyses over the short period) are associated with a dynamical anomaly (not shown), with a weaker wind and a switching direction on average in the box shown in Fig. 10. As can be seen in Fig. 10 over this box, in DJF, the southern part of the box is under the influence of south-easterly wind, while the northern part of the box indicates the penetration of maritime air mass coming from the north–northwest and corresponds to the monsoon flow. The trend of the wind components in this box (indicated by the contours) show a reinforcement of the south-easterly wind to the detriment of the northern–north-western flow. The cold and wet years occurring at the beginning of the period are thus associated with a stronger monsoon flow which attenuates at the end of the period. Hence, as already mentioned by several studies (e.g. Power et al., 1998; Hendon et al., 2007) dynamics mostly explain the variability and trends of temperature and humidity over this area.

Figure 11Temperature and IWV anomalies time series for a box over the eastern Sahel (see Fig. 12), using ERA-Interim (blue) and MERRA-2 (red) data, for (a, c) the 1980 to 2016 period, (b, d) the 1995 to 2010 period, and (a, b) the monthly time series and (c, d) the JJA season.

Download

Note that although the climatological means of zonal and meridional wind components are similar between ERA-Interim and MERRA-2, their trends over and around Australia present different patterns, especially in DJF (Fig. 10e–h), likely explaining the different IWV trends between both reanalyses.

4.3 Analysis over northern Africa–eastern Sahel

Here we focus on a box over the eastern Sahel (10–20^∘ N, 10–40^∘ E). The monthly trend in IWV is negative (drying) and significant in both reanalyses (except for MERRA-2 for the longer period), though it is twice as intense in ERA-Interim than in MERRA-2 (Fig. 11b). Similarly, the temperature trends are positive (warming) and significant in both reanalyses. Although the monthly anomalies show many similarities, their agreement is far from being perfect. The general strong negative IWV trend in ERA-Interim implies that IWV anomalies are higher in ERA-Interim at the beginning of the period and lower at the end of the period. However, both reanalyses present four different periods in the time IWV series: a drying trend at the very beginning (1980–1985) followed by a moistening trend until 1995, then followed by a new drying period lasting until around 2008 when the trend seems to stop (Fig. 11a, b).

Figure 12Close-up of northern Africa showing the mean monthly and JJA fields and trends (contours) of the u and v wind components at 925 hPa (shaded). The area of focus (where IWV trends are most intense in ERA-Interim) is marked by a box.

Download

The trend in T anomalies also stops at around 2008 (Fig. 11a). Before that period, the temperature anomaly is increasing significantly, despite strong month-to-month variability. However, there is low or negative correlation between IWV and T anomalies when considering the monthly time series. In JJA, the same periods of drying and moistening are observed (Fig. 11c). The correlation of anomalies for JJA between both reanalyses is quite good, both for IWV (around r=0.67 for the short period and r=0.63 for the longer period) and T (around r=0.69 for both periods), although their amplitudes and trends are quite different. MERRA-2 presents an overall moistening trend, while ERA-Interim shows a drying trend (Figs. 8g, h and 11c). Simultaneously, the temperature trends are both positive and significant, thus not explaining the IWV trends according to C-C.

Dynamics at 925 hPa are shown in Fig. 12. The mean states are plotted in colours over which the contours of the trends are superposed. The mean states in u925 and v925 are similar in both reanalyses, with a mean monthly north-easterly wind over the box (Fig. 12a–d) which is almost completely replaced with a south-westerly wind in JJA (Fig. 12e–h). This wind is slightly stronger in ERA-Interim than in MERRA-2. For both reanalyses, the trends in the mean flow indicate an increase in the zonal component (Fig. 12a, b). The trends in the meridional wind component show a dominant increase in the northerly wind from the Sahara. This trend may explain the general warming and drying in the eastern Sahel. The trends differ, however, with MERRA-2 showing a decrease in the northerly flow in upper-left angle of the box (Fig. 12d) while ERA-Interim shows an increase there and an increasing southerly inflow at the southern border of the box (Fig. 12c). This difference can explain the difference of intensity in these trends. In JJA, the trends in MERRA-2 are very weak (Fig. 12f, h) while in ERA-Interim there is a strong increase in the southerly flow from central Africa and in the north-easterly flow from the Sahara, explaining the net drying and warming (Fig. 11d). Figure 13 displays the inter-annual variability of JJA wind (monsoon flow) averaged over the box for both reanalyses.

It is clear that ERA-Interim has a stronger southerly flow in JJA and weaker northerly flow for the other months (Fig. 13a) with large inter-annual and decadal variability (Fig. 13b). The time series of wind in JJA in MERRA-2 clearly indicates the same four periods as for the IWV trends identified above, with a weakening of the south-westerly wind between 1980 and 1985, followed by an intensification of the monsoon flow arriving in this box between 1985 and 1995, and a wind decreasing and turning to the west until 2005 or 2006 and then becoming more stable on average. In ERA-Interim, we only observe two main periods: a weaker south–south-westerly wind at the beginning of the period followed by an intensification after 1990. The wind intensity is maximum between 1995 and 2000 but stays quite intense and with a south–south-westerly direction until the end of the period, being stronger and more southerly than in MERRA-2 after 2000. The different dynamics of the two reanalyses observed in this box partly explains the increasing deviation between both reanalyses at the end of the period.

Figure 13Time series of mean wind vectors for a box over the eastern Sahel (see Fig. 12), using ERA-Interim (blue) and MERRA-2 (red) data, for the 1980 2016 period: mean annual cycle (a) and the monthly time series for the JJA season (b).

Download