COSMIC is a highly successful RO mission that has demonstrated wide scientific applications in operational weather forecasts see, e.g., and global atmospheric studies see, e.g.,. In the RO data retrieval process, the bending angle (α) derived from Doppler shift measurements onboard low earth orbiting (LEO) satellites can be inverted to recover refractivity (N) based on the Abel transform, which is related to total pressure (P), temperature (T), and water vapor pressure (Pw) . In a dry atmosphere (with Pw=0), density profiles are obtained from the known relationship between refractivity and density, while pressure and dry temperature can be derived using the hydrostatic equation and equation of state for ideal gas see. In the presence of water vapor (especially in the lower troposphere), humidity and temperature profiles should be complemented with a priori information (e.g., numerical weather forecasts).

Alternatively, wet profiles can be generated using a one-dimensional variational (1-D-Var) method implemented at the COSMIC Data Analysis and Archive Center (CDAAC) at the University Corporation for Atmospheric Research (UCAR). These atmospheric profiles are provided as Level 2 RO data by various data retrieval centers including CDAAC (see http://cdaac-www.cosmic.ucar.edu/cdaac/status.html). In this study, COSMIC Level 2 RO data (both wet and dry profiles) covering the GBM river basin between April 2006 and December 2013 are used. The wet and dry profiles mainly differ in the lower troposphere due to presence of water vapor but are similar and highly accurate between 8 and 20 km see. The wet profiles are used for evaluating the accuracy of radiosonde observations (see details in Supplement) while the dry profiles (i.e., temperature only) are used for examining the interannual variations of UTLS over the GBM river basin. The GBM river basin received 59 419 COSMIC profiles from April 2006 to December 2013, out of which ∼14 % were found to be of bad quality. Figure a shows the number of monthly accumulated COSMIC RO data retrieved during the period, which indicated an average of ∼576 profiles per month. The number of profiles decreased considerably between late 2010 and 2012 (Fig. a) due to problems in some COSMIC satellites (see http://cdaac-www.cosmic.ucar.edu/cdaac/status.html). Figure b shows the distribution of RO data points at various pressure (altitude) levels indicating that many COSMIC data penetrate deep into the lower troposphere with more than 56 % of the profiles reaching at least 850 hPa (∼1.5 km a.m.s.l.). The geographical distribution of COSMIC profiles at 850 hPa (∼1.5 km), 700 hPa (∼3.1 km), 500 hPa (∼6.0 km), and 400 hPa (∼7.5 km) levels in Fig. a–d indicates the effect of high topography over the region (Fig. ), which blocks the GNSS radio waves. A near-complete coverage of the RO data can be seen at 400 hPa (∼7.5 km), corresponding to the highest altitude of the Himalayas.

COSMIC temperature profiles over the GBM region are compared with two high-resolution modern reanalysis products, (a) MERRA and (b) ERA-Interim . MERRA is produced by the state-of-the-art Goddard Earth Observing System Data Assimilation System version 5 (GOES-5), at National Aeronautic and Space Administration (NASA), USA. GEOS-5 assimilates data from a wide variety of observing systems (e.g., in situ, satellites) to produce a consistent set of spatio-temporal meteorological and climatic variables since the start of the satellite era (i.e., 1979). GEOS-5 is run on a 1/2∘×2/3∘ (or ∼50×70 km) grid with 72 vertical layers extending from the surface through to the stratosphere. Atmospheric variables (e.g., temperature, humidity) are produced at various temporal scales ranging from 3 hourly at 1.5∘×1.5∘ (or ∼150×150 km grid) to monthly scales at the nominal horizontal resolution.

ERA-Interim is the latest global atmospheric reanalysis produced by ECMWF covering the period 1979 to present . ERA-Interim builds on the previous generation of reanalyses (e.g., ERA-15 and ERA-40) with improved model aspects, more advanced assimilation techniques (e.g., 4-D Variational Schemes) and better land surface model, and assimilates atmospheric profiles retrieved from the GNSS RO data. The atmospheric variables (e.g., temperature and water vapor) are simulated at 6-hourly timescales over 60 vertical levels at a ∼ 79×∼ 79 km grid. Because ERA-Interim assimilates GNSS RO data, ERA-Interim and COSMIC RO data are not completely independent. found that GNSS RO data help to reduce temperature bias of ERA-Interim in the UTLS but are found to produce drying effects in the tropics. Monthly mean temperatures at 14 pressure levels from 500 to 10 hPa are obtained for both MERRA (see http://disc.sci.gsfc.nasa.gov/daac-bin/FTPSubset.pl) and ERA-Interim (see http://apps.ecmwf.int/datasets/data/interim-full-daily/levtype=sfc/) to assess UTLS temperature over the GBM river basin.

Three ocean–atmospheric indices are used in this study, representing the (a) ENSO, (b) IOD, and (c) QBO, which are commonly associated with significant fluctuations in UTLS temperatures. ENSO is commonly defined by sea surface temperature (SST) anomalies in the equatorial Pacific ocean, typically over 5∘ N–5∘ S, 120–170∘ W, which is also known as Niño3.4 region see. ENSO events are said to occur if SST anomalies exceed 4 ∘C for 6 months or more. Warm and cold ENSO phases are referred to as El Niño and La Niña events, respectively, which are represented by anomalous warming of the central and eastern tropical Pacific (warm phase) and vice versa. Niño3.4 index is obtained from the National Oceanic and Atmospheric Administration (NOAA, see http://www.esrl.noaa.gov/psd/data/climateindices/list/).

IOD is measured by difference of SST anomalies between the western (50–70∘ E and 10∘ S–10∘ N) and eastern (90–110∘ E and 10–0∘ S) equatorial Indian ocean, also referred to as Dipole Mode Index (DMI). Positive IOD events are identified by cooler than normal water in the tropical eastern Indian Ocean and warmer than normal water in the tropical western Indian Ocean and are associated with a shift of active convection from eastern Indian Ocean to the west, leading to potentially higher than normal rainfall over parts of the Indian subcontinent. DMI is obtained from http://www.jamstec.go.jp/frsgc/research/d1/iod/. QBO is stratospheric phenomenon characterized by an east–west oscillation in stratospheric winds over a period of approximately 28 months . The QBO dominates variability of the equatorial stratosphere and is easily identified as downward propagating easterly (negative) and westerly (negative) wind regimes. It is commonly characterized by an index derived based on zonal winds at 30 or 50 hPa. Here, the QBO index at 30 hPa level covering the period 2006–2013 is obtained from NOAA (http://www.esrl.noaa.gov/psd/data/correlation/qbo.data).

Various methods have been used to define the tropopause, such as lapse-rate tropopause (LRT), cold point tropopause, dynamical tropopause (isentropic potential vorticity), ozone tropopause, and 100 hPa pressure level see. This study focuses on the changes and interannual variations of the LRT given that it is an important indicator of climate change see. Based on , the LRT (hereinafter as tropopause) is defined as “the lowest level at which the lapse rate decreases to 2 ∘ km-1 or less, provided also the average lapse rate between this level and all higher levels within 2 km does not exceed 2 ∘C km-1”. COSMIC RO data obtained from CDAAC already contain the derived tropopause parameters (heights and temperatures), while MERRA also provides tropopause temperatures and pressures based on the definition. The tropopause heights, hLRT (in km), for MERRA was approximated from the tropopause pressures, P (in Pa), using the following relationship : hLRT≡44 330.8-4946.54×P0.1902632.

Monthly COSMIC RO data are interpolated to a 0.5∘×0.5∘ grid over 14 standard pressure levels from 500 to 10 hPa (or 7.5–31 km) and the tropopause using the kriging technique (see details in the Supplement). Geostatistical kriging methods have been shown to be more robust and spatially more reliable than other existing methods such as inverse distance weighting, Thiessen polygons see, e.g.,. To study the interannual variations of temperature at various pressure levels, PCA is applied to the deseasonalized (annual cycle removed) tropopause heights and temperatures. PCA is a well-known data exploratory tool used in atmospheric/oceanic science since it allows for a space–time display of geophysical data (e.g., temperature), in very few modes see. The idea of PCA is to find a set of orthogonal spatial patterns (or empirical orthogonal functions, EOFs) along with a set of associated uncorrelated time series or principal components (PCs) that captures most of the observed variance (expressed in percent) from the available spatio-temporal data (e.g., temperature). In summary, the EOF decomposition can be written as X(t,s)=P(t,n)E(s,n)T, where X(t,s) is the space (s)–time (t) data with time mean or annual cycle removed, E(s,n) contains the EOFs with n number of retained modes, and P(t,n) are the PCs obtained by projecting the original data (X(t,s)) on the orthogonal base functions E(s,n), i.e., P(t,n)=X(t,s)E(s,n).

First, we compared COSMIC profiles (both dry and wet profiles) with temperature, water vapor pressure, and refractivity profiles from 24 radiosonde stations across the GBM river basin from August 2006 to December 2013. The results (see Supplement) confirmed those of the previous studies e.g., with radiosondes over India (referred to as IMD/MK4) and Bangladesh indicating substantial bias in the UTLS. The results from three recently upgraded radiosondes (over India) show significantly reduced bias in the UTLS with respect to the COSMIC RO data, suggesting that incorporating GPS receivers in conventional radiosondes helps to provide better estimates of temperature and water vapor profiles through more accurate measurement of pressure at various altitude levels (see Supplement). This highlights the importance of RO data in providing state-of-the-art data for calibrating existing radiosondes.

Figure  shows the regional mean temporal evolution of UTLS (dry) temperature anomalies (time mean removed). For COSMIC, dry temperature anomalies are plotted in Fig. a. The temperature anomalies range between ±6 ∘C, indicating largest values above 50 hPa level (lower stratosphere) and below 200 hPa (troposphere). A strong seasonal cycle is evident in the troposphere below 200 hPa level and the stratosphere (above 70 hPa level). The three data sets (COSMIC, MERRA, and ERA-Interim) agree very well above 200 hPa where water vapor is negligible. Below 200 hPa level, however, COSMIC data (Fig. a) are found to be colder than the two reanalyses (Fig. b–c) as the effect of water vapor becomes more significant. Both MERRA and ERA-Interim show quantitatively similar biases with respect to the COSMIC RO data, indicating a bias of 1.23 and 1.22 ∘C, respectively, when averaged over 200 to 70 hPa level, whereas the difference between MERRA and ERA-Interim was found to be ±0.5 ∘C over the same layer during the last 89 months. The annual cycle of temperature at (a) 200 hPa, (b) 100 hPa, (c) 70 hPa, and (d) 50 hPa levels are shown in Fig.  to estimate the absolute bias (against COSMIC RO data) in reanalysis temperature data. Both MERRA and ERA-Interim indicate warm bias at all the levels with varying magnitudes over different seasons. Both reanalysis products are found to be warmer by ∼1 ∘C in June at 200 hPa level (Fig. a), ∼1.5 ∘C at 100 hPa level (Fig. b), and up to 2 ∘C at 70 hPa level from November to May (Fig. c).

Figure  shows the detrended (as well as deseasonalized) time series of temperature anomalies at four pressure levels mentioned above. The temperature anomalies show considerable interannual variability from 2006 to 2013, indicating large negative anomalies in the troposphere during 2009/2010 winter and early 2013, and the low stratosphere during 2007/2008, 2008/2009, and 2012/2013 winters. The 100 hPa level was warmer by ∼1.5 ∘C during the period 2006/2007, 2009/2010, and 2012/2013, and this is consistent in all the three data sets (Fig. a–c). The warm anomalies at the 100 hPa level for all three periods coincide with warm (i.e., El Niño) ENSO phase, while anomalously cold temperatures at 50 hPa level during 2009/2010 and 2010/2011 coincides with the recent stratospheric sudden warming (SSW) events. The stratospheric planetary waves in the winter Northern Hemisphere can become so intense that they can rapidly disrupt the northern polar vortex, replacing the westerly winds with easterly winds at high latitudes, leading to a dramatically warm polar stratosphere. This phenomenon is called SSW and has a tendency to cool the stratosphere in the tropics and subtropics (e.g., 50 hPa level in Fig. ). The temperature decreased by about 5 ∘C during the 2008/2009 SSW event at 50 hPa level see also. The SSW events can occur during both westerly and easterly phase of the QBO, in which both 2008/2009 and 2010/2011 events occurred during the westerly phase. The warm temperatures during 2006–2007 at the UTLS are associated with the combined impacts of ENSO and IOD, whereas positive temperature anomalies above 50 hPa possibly indicate a weak SSW.

In order to relate them to the three ocean–atmospheric indices described in Sect. , the temperature anomalies at (a) 200 hPa, (b) 100 hPa, (c) 70 hPa, and (d) 50 hPa are plotted in Fig. a–d together with the three indices in Fig. e. The 200 hPa level temperature anomalies clearly indicate the influence of 2009/2010 El Niño event (Fig. a) where temperature decreased by ∼1.5 ∘C when El Niño was at its peak in January 2010 (Fig. a). The 200 hPa level temperatures are negatively correlated with ENSO (Fig. e) and 100 hPa level temperature (Fig. b). The 100 hPa level, whose temperatures are highly correlated with ENSO (see Table ), also appears to be closely associated with the QBO anomalies especially during 2008/2009 and 2010/2011 (Fig. b). The temperature anomalies at 70 and 50 hPa levels (Fig. c–d) primarily depict the structure of recent major SSW events (see also, Fig. ), both of which occurred during the westerly phase of the QBO cycle. Correlation coefficients between temperature anomalies at these four pressure levels and three atmospheric/ocean indices are given in Table . Significance of the correlations is tested at 95 % confidence level using a reduced degree of freedom, which was obtained by dividing the total number of months (n=89) by 4 months that are used to smooth the time series.

As shown in Fig. , ENSO is highly correlated (0.82 at 1-month time lag) with temperatures at 100 hPa level and tend to be insignificant above 70 hPa level. Warmer (colder) SST leads to stronger (weaker) convection resulting in colder (warmer) tropopause temperatures and are also negatively correlated with temperatures at 200 hPa level. Both MERRA and ERA-Interim show similar correlation coefficients (results not shown). Since the IOD phenomenon is closely associated with the ENSO events between 2006 and 2013 (with a correlation coefficient of 0.42), it is found to be significantly correlated with temperatures in the lower troposphere with a correlation of -0.53 at 400 hPa level (see Table ). At the 200 hPa level, its correlation decreased to -0.42 (see Table ) but is still significant at 95 % confidence level. Temperatures at 100 hPa level over the tropics have been shown as an approximation of the QBO signal in due their very high correlation (0.86) between 2004 and 2010. However, their relationship did not hold steady (with a correlation of 0.47 at 100 hPa level) as the QBO westerly phase slowed dramatically lasting for about 21 months (i.e., 1 and 1/2 cycle) from June 2008 to January 2010, followed by a step easterly phase in June 2010.

Temperature changes estimated at the four pressure levels (Fig. ) from 2006 to 2013 are given in Table . The linear trends are estimated from the deseasonalized temperature anomalies using non-parametric Sen's slope estimator . Significance of trends are tested at 95 % confidence level based on the Mann–Kendall non-parametric test . Consistent with the time series shown in Fig. a, there is a slight increase (but not significant) in temperature (0.02±0.02 ∘C year-1 based on COSMIC RO) at 200 hPa level and a decrease in temperature (-0.04±0.05 ∘C year-1 based on COSMIC RO) at the 100 hPa level. It also confirms the recent stratospheric cooling trends , indicating a temperature decrease at the rate of 0.07±0.05 ∘C year-1 based on COSMIC RO at 70 hPa level during the past 8 years. These trends are also consistently estimated by the two reanalysis products (MERRA and ERA-Interim) except at 100 hPa level where MERRA data did not show any trend (see Table ). The uncertainties in trend estimates were relatively larger than the trend themselves due to the short time span but nevertheless the trends are clearly visible at different levels (see Fig. ).

The annual mean and standard deviation of tropopause temperatures and heights are plotted in Figs.  and . The tropopause is generally colder (higher) in south (closer to the equator), reaching a minimum (maximum) temperature (height) of -81.5 ∘C (16.9 km) over southern Myanmar (Figs. a and a). While the temperature gradually increases from south to north (from -81.5 to -69.5 ∘C, based on COSMIC RO in Fig. a), its heights are more or less homogenous at around 16.8 km below 29∘ N with its boundary roughly falling on the northern boundaries of Bhutan. However, its height changes steeply by around 2 km from 29 to 35∘ N, which also shows the largest standard deviation (∼1.8 km, Fig. b). The standard deviations of temperatures reaches up to ±6 ∘C based on COSMIC RO data (Fig. b) over the same region.

The tropopause over the GBM river basin often reaches as high as 18 km in response to the Indian summer monsoon, when intense convective activities occur, and as low as 10 km during winter. The spatial patterns of tropopause shown by MERRA are consistent with those from COSMIC RO but are found to be more zonally homogenous, warmer (by up to 4 ∘C in the north) and lower (by ∼1 km) over the region. This warm bias (against COSMIC RO data) is also observed in ERA-Interim at 100 hPa level (see Fig. b) but is relatively lower than MERRA, especially during the monsoon. The annual cycle of area-averaged (over the spatial domain covering 71.5–99.5∘ E, 16.5–34.5∘ N) tropopause temperatures and heights of COSMIC RO and MERRA are shown in Fig. . The area-averaged temperatures (heights) of COSMIC RO reach minimum (maximum) in June but are warmer (lower) in MERRA by 1.0–2.5 ∘C (∼1.2 km from May to December). MERRA also shows the tropopause temperature minimum in July instead of June (Fig. a). The large differences in MERRA could partly be related to the approximation in Eq. (), but it should be noted that errors in tropopause heights may cause large errors in temperature due to the lapse-rate criterion. The warm bias (against COSMIC RO data) observed in reanalysis products was thought to mainly stem from assimilation of radiance observations from aircrafts and satellites see, e.g.,and references therein. Large variations in tropopause temperatures and heights in Fig.  (indicated by error bars) during winter and spring could be related to high diurnal temperature variations .

The annual cycle was removed from each grid cell to examine the interannual variability of tropopause temperatures and heights and to estimate linear trends over the period August 2006 to December 2013. The area-averaged linear trends and their uncertainties are given in Table . In general, based on the COSMIC RO data, the tropopause appears to be cooling (increasing in height) at a rate of -0.039±0.05 ∘C year-1 (6.01±5.02 m year-1) during the period (see Table ), which to some degree is also estimated by MERRA. However, MERRA shows negligible cooling compared to COSMIC RO while its height increase is not consistent with its temperature decline. The increasing (decreasing) tropopause heights (temperatures) has been consistently observed in GNSS RO data over the years at both global and regional scale e.g.,, which is evidently in response to enhanced warming in the upper troposphere and substantial cooling in the lower stratosphere.

Next, PCA is applied on the deseasonalized time series of tropopause temperatures and heights in order to study the spatio-temporal characteristics of tropopause over the GBM river basin. PCA is particularly relevant here because tropopause is a transitional layer that responds to perturbations from both the troposphere and stratosphere, which makes it difficult to understand their variability modes. Figure  shows the EOFs (or spatial maps) of the first three leading modes of variability. The first EOF accounts for a variability of ∼73 % (COSMIC RO) and ∼63 % (MERRA) indicating positive anomalies (up to 1.1 ∘C) across the GBM river basin. EOF 1 appears to be rather symmetric around 29∘ N in COSMIC RO but seems to be shifted slightly southwards in MERRA (Fig. a and d). Their corresponding PCs are shown in Fig. . PC 1 (Fig. a) is found to be highly correlated with Niño3.4 index with a correlation of 0.77 (COSMIC RO) and 0.78 (MERRA) with a time lag of 1 month, indicating that ENSO dominates tropopause variability over the region (see Table ). DMI is also moderately correlated (0.35) with PC 1 of both COSMIC RO and MERRA (with a lag of 2 months), indicating that warmer SSTs in the equatorial Indian ocean might be having some influence on the tropopause variability.

The second EOF shown in Fig. b and e indicates a diagonal (dipole) pattern with positive (negative) anomalies in the northwest (southeast), accounting for variance of ∼10 % (COSMIC) and ∼18 % (MERRA). Its corresponding PC is found to be significantly correlated with the QBO index with a correlation coefficient of 0.40 (COSMIC RO) and 0.53 (MERRA) (at zero lag). It is not surprising that the relationship between PC 2 and QBO is relatively low compared to the equatorial (or tropical) tropopause since QBO is a tropical phenomenon e.g.,. It also stems from the fact that QBO westerly phase prolonged for an extended period of 21 months before changing to a westerly phase in January 2010. The third EOF (Fig. c and f) explains about 5 % (COSMIC RO) and 10 % (MERRA) of the variability and shows positive (negative) anomalies below (above) 30∘ N, although MERRA shows a diagonal dipole pattern similar to EOF 2. Their corresponding PCs are found to be moderately correlated with ENSO and IOD. The tropopause heights are negatively correlated with their temperatures and, therefore, vary inversely with its temperature, i.e., increase in tropopause height with decrease in temperature (figures not shown). The correlation coefficient between the PCs of three leading modes of tropopause heights and the ocean–atmospheric indices indicate similar magnitudes of correlations but with opposite signs (see Table ).

To show the influence of ENSO mode on the tropopause, the seasonal mean area-averaged anomalies of tropopause temperatures and heights are plotted in Fig. . Seasonal mean anomalies are obtained by averaging the products of EOF 1 and PC 1 (of the COSMIC RO data). The ENSO effect is found to be maximum during the winter (e.g., 2009–2010, 2012–2013) when ENSO was at its peak. Its effects are also felt during autumn (in 2007 and 2011) and spring (in 2008) and during the El Niño and La Niña periods. The largest tropopause anomaly occurred during the major El Niño event of 2009/2010 when tropopause temperature (height) increased (decreased) by about 1.5 ∘C (300 m) in the winter (Fig. a–b). La Niña periods (e.g., 2007/2008, 2010/2011) are mainly associated with deep convections in the troposphere leading to a wider troposphere or higher tropopause height (see Fig. b).