A drop in upper tropospheric humidity in autumn 2001, as derived from radiosonde measurements at Uccle, Belgium



[1] Simulations of climate models predict a doubling of the amount of upper tropospheric water vapor by the end of this century, caused by the increasing concentrations of greenhouse gases. Observations indicate that the tropopause height has increased by several hundred meters since 1979. In this paper, we verify and link these two results by carrying out a time series analysis on a uniform database of corrected radiosonde vertical profiles gathered at Uccle, Belgium, and covering the 1990–2007 time period. The most remarkable finding of this trend analysis is a significant drop in upper tropospheric humidity (UTH) around autumn 2001, which marks an end to the upper tropospheric moistening of the precedent decade. This UTH drop in autumn 2001 coexists with a sudden lifting and cooling of the tropopause and with a significant stretch-out of the free troposphere. Therefore, we conclude that these autumn 2001 trends are certainly associated with the dynamical behavior of the troposphere, triggered by the surface warming. Links with the solar variability and the lower stratosphere were investigated but could not be established definitely.

1. Introduction

[2] Water vapor is a key variable for climate research. It is the dominant greenhouse gas in the atmosphere and provides the largest known feedback mechanism for amplifying climate change [Bony et al., 2006; Soden and Held, 2006]. While the total water vapor amount is the first-order quantity determining the water vapor greenhouse effect, the upper tropospheric water vapor content has an especially strong influence on the amount of outgoing long-wave radiation [see, e.g., Solomon et al., 2010]. The total amount of water vapor is expected to increase due to global warming (and is confirmed from an observational point of view) [see, e.g., Durre et al., 2009; McCarthy et al., 2009], and simulations with coupled ocean-atmosphere models and satellite measurements both point to an increase in upper tropospheric specific humidity q with a near-constant upper tropospheric relative humidity (RH) [Minschwaner and Dessler, 2004; Minschwaner et al., 2006; Soden et al., 2005; Gettelman and Fu, 2008], despite the facts that (1) the different models have very different humidity mean states and (2) systematic specific humidity biases arise between the global climate model (GCM) results and satellite measurements [Soden and Held, 2006; John and Soden, 2007].

[3] Owing to their global coverage, the already mentioned satellite measurements for upper tropospheric humidity have a great potential for climate studies [Buehler et al., 2008]. Their major drawback is the very coarse vertical resolution they provide (measurements over layers that are typically 1–3 km thick). For instance, Gettelman et al. [2006] used the data from the Atmospheric Infrared Sounder on the NASA Aqua satellite to develop a climatology of upper tropospheric relative humidity. They found that the highest variances in humidity are seen around the midlatitude tropopause. Other research satellite instruments measure upper tropospheric water vapor (e.g., the Microwave Limb Sounder), but the data from these instruments are only available for relatively short time periods (see, e.g., Fetzer et al. [2008] for a comparison of the upper tropospheric humidity (UTH) observations from those two instruments). On the other hand, there exist some instruments measuring water vapor on operational meteorological satellites which span longer time periods. The longest available UTH record from satellite measurements dates back to 1979 and comes from the High-Resolution Infrared Sounder instrument. On the basis of these observations, Bates and Jackson [2001] described decadal trends in upper tropospheric (relative) humidity which are strongly positive in the deep tropics, negative in the Southern Hemisphere subtropics and midlatitudes, and of mixed sign in the Northern Hemisphere subtropics and midlatitudes. Since 1993, continuous humidity measurements have also been available from instruments on operational satellites using the microwave range, especially a prominent water vapor line at 183.31 GHz [Milz et al., 2009]. The advantage of using the microwave range instead of the IR (and most notably the spectral region near 6.7 μm) is that the data are much less affected by clouds.

[4] In addition to satellite measurements, other data sources for UTH are available. These are generally characterized by a higher vertical resolution. In situ aircraft observations are of high quality but of limited extent [see, e.g., Gettelman et al., 2004], as is the case for the sensitive balloon-borne instruments described by Vömel et al. [2007]. The longest available data record is from operational balloon radiosondes. Unfortunately, the spatial coverage of radiosonde measurements is poor and radiosonde UTH measurements can suffer from significant dry biases [Soden et al., 2004; Ferrare et al., 2004; Vaughan et al., 2005; Miloshevich et al., 2006]. Moreover, previous studies have documented substantial spatial [Soden and Lanzante, 1996] and temporal [Elliott and Gaffen, 1991] discontinuities in their historical records that are frequently related to differences in radiosonde instrumentation. In this paper, we focus on the time series analysis of radiosonde UTH observations of a single station, Uccle (Belgium, 50°48′N, 4°21′E, 100 m above sea level). The major advantage of this approach is that we know the exact metadata of these measurements, so we can avoid inhomogeneities due to instrumental changes. The use of these high-resolution profiles also enables us to link the UTH time variability to tropopause properties, tropospheric variability, surface warming, and lower stratospheric temperature changes at the same site. We also take advantage of this fact by defining the UTH relative to the tropopause, and not only in absolute terms (as was done, e.g., by Milz et al. [2009]). Of course, with this case study, we do not aim at finding and explaining trends or changes which are representative for the whole globe, not even for the northern midlatitudes or Europe, although we extend our analysis to some radiosonde stations in the surroundings.

[5] The paper is organized as follows. In section 2, we describe in detail the radiosonde observations on which the present study is based. The methodology of the time series analysis is covered in section 3. In section 4, the results of a uniform, but short-term (17-year), time series analysis are presented. The findings of this analysis are then put into a larger perspective, both in space and time (section 5). Section 6 is reserved for a discussion of the results and drawing conclusions.

2. Data

[6] The observations which form the basis of the research described in this paper are radiosonde vertical profiles of temperature, pressure, altitude, and relative humidity, gathered at Uccle. Uccle is located in the residential part of Brussels, 6 km south of the city center. For a detailed analysis of the intensifying urban heat island effects on the surface air temperature time series, we refer to the work of Hamdi et al. [2009]. At Uccle, the radiosonde data are available in digital form since September 1963 for the standard levels and since January 1968 for the tropopause identification. However, this data set, covering more than 40 years, is gathered by different types of radiosondes and, hence, different types of pressure and humidity sensors, different thermometers, etc. This lack of uniformity of instruments most seriously affects the humidity measurements, as large differences exist between the response of different humidity sensors. Moreover, the humidity sensors of the early years were simply not sensitive enough at lower temperatures, so the resulting measurements cannot be trusted. Another source of inhomogeneity is the difference in launch times throughout the 40 years of observations, which of course also has an effect on the temperature measurements, next to the humidity measurements. Therefore, to build up a homogeneous time series of temperature profiles, we restrict ourselves to radiosondes launched at a fixed time; for a homogeneous time series of humidity profiles, the radiosonde type is additionally imposed.

[7] As a consequence, the longest possible homogeneous time series of radiosonde humidity measurements available at Uccle spans a period of more than 17 years, from January 1990 until August 2007. During this period, the type of radiosonde used was Vaisala's RS80-A. Until November 2001, the RS80 sondes were launched twice a day, at 0000 and 1200 UTC; after November 2001, this type of sonde was only launched in the context of ozone soundings, with a reduced frequency of three times a week (Monday, Wednesday, and Friday) at 1200 UTC. For this 17 year time period, we dispose of data points every 10 s, so that the theoretical vertical resolution is about 100 m on average, neglecting at the moment the time lag in the humidity sensor response. Throughout this paper, the 1990–2007 Uccle time series of RS80-A radiosonde data, launched at 1200 UTC, with data points every 10 s, is designated as the uniform time series. The total Uccle time series is defined as all available radiosonde observations at 1200 UTC, starting in 1963.

[8] The used Vaisala RS80-A sondes contain the A-Humicap sensor, a planar thin-film capacitive sensor using a highly porous polymer electrode, whose capacity depends on the amount of water vapor and the air temperature [Verver et al., 2006]. The RS80-A sonde was introduced in the early 1980s and has been the most frequently used radiosonde in the world for more than a decade, despite its reported dry bias, especially at low temperatures [e.g., Miloshevich et al., 2001; Wang et al., 2002; John and Buehler, 2005; Leiterer et al., 2005; Nash et al., 2005; Sapucci et al., 2005; Vaughan et al., 2005; Häberli, 2006; Suortti et al., 2008; and references therein]. A number of error sources have been identified:

[9] 1. Temperature-dependence error occurs when an inaccurate calibration model is used for the temperature dependence of the sensor response at low temperatures (dominant error source for T ≪ −20°C [Suortti et al., 2008]).

[10] 2. Chemical contamination error occurs when nonwater molecules (e.g., from packaging material) occupy binding sites in the sensor polymer. This contamination dry bias was corrected by Vaisala by a change in the packaging (absorption material, from September 1998, and a removable boom cover, from June 2000 onward). This adaptation clearly reduces the RS80-A dry bias; however, a dry bias still remains in the data [Wang et al., 2002; Wang and Zhang, 2008].

[11] 3. Sensor aging, or long-term instability of the sensor material, can cause a dry bias. The sensor's drift is mainly caused by reduced polymer sensitivity to water vapor and is therefore seen more clearly at high humidities. The A-Humicap drift at saturation is approximately −5% RH after a 2 year storage time, and less than −0.5% RH per year thereafter [Wang et al., 2002].

[12] 4. Despite its aluminized protective cap, the humidity sensor remains susceptible to solar heating.

[13] 5. A thin ice layer can form around the humidity sensor, so it behaves more like a (bad) thermometer.

[14] 6. A time-lag error can cause the humidity sensor to have a too-slow response at low temperatures.

[15] To retrieve unbiased and improved relative humidity data, correction algorithms have independently been developed by Miloshevich et al. [2004] and Leiterer et al. [2005] for the RS80-A sondes. Both methods include an additional improved temperature-dependence correction and time lag corrections. In addition to the temperature-dependence error, the method described by Leiterer et al. [2005] also addresses the RH-dependent part of the RS80-A error, which is a factor in temperatures above −40°C. This is provided by a modeled ground check correction at 100% RH [Suortti et al., 2008]. The correction method of Miloshevich et al. [2004], in collaboration with Vaisala, is based on laboratory tests and comparisons with fast-response hygrometer measurements. Leiterer et al. [2005] developed their own research sondes (“FN sondes,” modified Vaisala RS90-H sondes) and calibration method, and the RS80-A correction scheme is based on the comparison with those reference sondes.

[16] Both time lag correction schemes produce very similar results and are able to recover more vertical structure in the upper tropospheric humidity field. The differences between these methods arise mainly from the temperature-dependence correction. On the basis of tropospheric comparisons of Vaisala radiosondes with balloon-borne frost-point and Lyman-α hygrometers during a dedicated experiment, Suortti et al. [2008] classified the correction algorithm by Leiterer et al. as the best available because the RS80-A dry bias can be almost totally removed. The case study comparing radiosonde humidity data to advanced microwave sounding unit (AMSU) satellite humidity observations described by Buehler et al. [2004] demonstrated that the radiosonde correction performed at Lindenberg by Leiterer et al. significantly reduces the bias between simulated and measured AMSU radiances, particularly in the upper troposphere. As a result, the overall agreement is very good, with radiance biases below 1.5 K (which translates to about 15% relative error in relative humidity), but the corrected radiosondes still underestimate the relative humidity under extremely dry conditions, showing 0% RH when the true value is 2%–4% RH.

[17] The other approach [Miloshevich et al., 2004] tends to overcorrect in high RH conditions when T < −50°C. For T > −30°C, it is ineffective and does not correct the RS80-A dry bias in high ambient RH [Suortti et al., 2008]. During a small intercomparison campaign in Uccle with three types of Vaisala radiosondes (RS80-A, RS90-H, and RS92-H), the overcorrection of the method of Miloshevich et al. [2004] also came out. Another result from our intercomparison campaign is that, even after correction with either method, the RS80-A humidity profiles show, on average, an absolute dry bias of more than 10% at the surface, which decreases with increasing height, with regard to simultaneous RS9x-H profiles (with x = 0 or 2). This discrepancy is probably caused by the deterioration of the RS80-A humidity sensors (sensor aging plus chemical contamination) owing to their long storage time (about 2 years in the case of the RS80-A sondes used in the intercomparison campaign).

[18] Our uniform time series of RS80-A radiosonde measurements at Uccle is corrected by the schemes proposed by Leiterer et al. [2005]. Additionally, we also used their icing recognition algorithm as a guidance, next to the scrutinous visual inspection of the humidity profile for sensor icing. It turns out that about one fifth of the observations are retrieved by an ice-contaminated sensor. About the same amounts are found in the Finnish RS80-A radiosonde records [Suortti et al., 2008] and in the Lindenberg soundings [Leiterer et al., 2005]. Humidity profiles resulting from an iced sensor are discarded from our data set of radiosonde relative humidity observations. There exist some interannual variations in the relative frequency of measurements retrieved by iced humidity sensors, but without a distinct relation with the variability discussed in this paper.

3. Methodology

3.1. Statistics

[19] As stated in section 2, homogeneity of climate data is indispensable for many aspects of climate research, especially for a realistic and reliable assessment of historical climate trends and variability, and also for the calculation of related statistics that are needed and used to define the state of climate and climate changes. On the other hand, autocorrelation and periodicity are inherent in most climate time series; for example, successive observations are unlikely to be independent of one another and there is a clear seasonal cycle in climate time series. Unfortunately, most statistical analysis tests (linear regression, change-point tests, and correlation) rely on independent and identically distributed time series. The rationale and theoretical basis, examples, and technical details of such statistical tests can be found in the work of Lanzante [1996]. Therefore, all data used in this study are in anomaly form: the seasonal cycle has been removed by subtracting from all monthly means the long-term monthly means.

[20] Despite our precautions to construct a uniform time series (see section 2), variations between different RS80-A production batches, the introduction of new packaging material and a new sensor boom cover, and changes in measurement techniques can lead to artificial (nonclimatic) discontinuities in our data set. On the other hand, natural variability can cause a change (discontinuity) in the level (mean) of the time series. Hence, as a first step we apply a change-point test to identify a possible stepwise shift in the mean of the time series. We chose to use the Pettitt-Mann-Whitney (PMW) test, which is a nonparametric test based on the ranks of the values of the sequence. It seeks to find a single change point; to find more change points, the sequence should be cut into parts. Because it is based on ranks, the test is not adversely affected by outliers and can be used when the time series has gaps [Lanzante, 1996]. A drawback of the test is the sensitivity to breaks in the middle of the time series. Change points that exceeded the 90% confidence level were retained. For comparison, we also applied the more commonly used Wilcoxon rank-sum test and the cumulative sum test to detect change points in the mean in a time series. If not otherwise stated, the mentioned change points are always detected by all three change point tests.

[21] Next to a stepwise shift in the mean of a time series, a change point might also be responsible for a change in a trend slope. Or, in case of a slowing of the rate of increase, the change point might represent only a change in the trend slope. Therefore, we also applied in some cases the change-point detection method of Lund and Reeves [2002] which is designed to detect both step- and trend-type change points. This test is based on a classic simple linear regression model that allows for two phases, one before and one after the change point.

[22] A last tool or technique to evaluate the time behavior of a data series is to visualize the (monthly) cumulative deviations: these are acquired by calculating for each monthly value (mean, anomaly) the sum of the deviations of the preceding monthly values with the global mean. By definition, the monthly cumulative deviations of a time series start and end at zero. Examples are given later in this paper (see, e.g., Figure 1).

Figure 1.

(top) Uccle time series of monthly anomalies of the integrated specific humidity for (left) the upper tropospheric layers between 500 and 200 hPa and (right) a layer that extends from the tropopause to 3 km below the tropopause. The mean before and after the detected change point are shown in grey. The linear regression lines for (left) the entire time period and (right) the time periods before and after the detected change point are also drawn. Red lines are used for positive trends, blue lines for negative trends. A solid line denotes a statistically significant trend, and a dashed line denotes a statistically insignificant trend. The statistical significance of the trends is investigated by Spearman's test. Green lines denote the zero anomaly lines. (bottom) Respective monthly cumulative deviations of the anomalies shown (Figure 1, top) from the mean of the anomalies. Green lines denote the zero lines.

[23] Single linear trends are calculated by minimizing the least squares. The slope of the linear regression line was used as a quantitative indication of the rate of change over the data period. The standard error of the linear regression slope was also computed as an estimate of the uncertainty in the slopes [Ross and Elliott, 1996]. Additionally, to test the statistical significance of this trend, we applied Spearman's test of trend. This is a nonparametric measure of linear association based on correlation of ranks. In this study, values that rejected the null hypothesis of randomness at the 95% confidence level were considered statistically significant [see also Ross and Elliott, 2001]. If a real climatological change point was detected by the PMW test, we calculated trends for the two parts of the time series (before and after the change point) to check if the change point is also a trend turning point. These last two steps are immediately combined in the two-phase linear regression scheme developed by Lund and Reeves [2002].

[24] If a change point is present in a time series of climate variables, the long-term behavior might better not be quantified using the linear slope of a single straight line. In such a case, an increments study should give some added value. Seidel and Lanzante [2004], for instance, explored this idea and assessed three alternatives to linear trends for characterizing global atmospheric temperature changes. Here, as in the work of Añel et al. [2006], in the case of a statistically significant change point, we split the initial series at the change point and we compute for each segment its increment by multiplying the linear slope with the period covered by the segment. The “increment” of the entire time series is then defined as the value obtained by adding the increments of the two segments, divided by the total length of the initial time series.

[25] Finally, to determine the correspondence between (1) different atmospheric variables measured at Uccle or (2) different stations for a given atmospheric variable, linear Pearson correlation coefficients were calculated and scatterplots were constructed. Correlations were taken for monthly anomaly time series that are not detrended. However, the correlation coefficients calculated for time series first detrended with single linear regression lines.

3.2. Tropopause Identification

[26] In this paper, we introduce an upper tropospheric humidity relative to the tropopause. Therefore, the identification of the tropopause is of major importance. The tropopause used here is the standard (first) thermal tropopause, defined as “the lowest level at which the lapse rate decreases to 2 K km−1 or less, provided also the average lapse rate between this level and all higher levels within 2 km does not exceed 2 K km−1” [WMO, 1957]. In case of the uniform 1990–2007 time series of radiosonde data, we dispose of profiles with a vertical resolution of about 100 m (data points every 10 s), so that the tropopause is calculated, after the sounding, using all these levels. Therefore, we used the algorithm described in the appendix of Zängl and Hoinka [2001] but also carefully checked the vertical (T, RH, and ozone) profiles to avoid erroneous identifications of the thermal tropopause. The sounding profiles of the uniform 1990–2007 time series were also used to determine the location of a second tropopause, if present. Only soundings that reached an altitude of 20 km were considered and the definition of the World Meteorological Organization's (WMO's) Commission for Aerology is applied [WMO, 1957]: “if above the first tropopause the average lapse rate between any level and all higher levels within 1 km exceeds 3 K km−1, then a second tropopause is defined by the same criterion as for the first tropopause. This tropopause may be either within or above the 1 km layer.”

[27] For the total Uccle time series of radiosonde measurements, starting in 1968, the first tropopause is identified in a similar way. However, in the early years, only the mandatory levels of the sounding are available. The tropopause determination of the other European radiosonde stations considered further in this paper is based on the identification (code 22) in the Integrated Global Radiosonde Archive (IGRA). Antuña et al. [2006] showed in a case study that, although there is a high amount of missing data in the IGRA data set for a given station, the existing data set is statistically representative of the complete data set for the tropopause features.

[28] Finally, for the uniform Uccle time series, we also calculated the ozone tropopause, as defined by Bethan et al. [1996] using the following three criteria:

[29] 1. The vertical gradient (evaluated over a depth of ≈200 m) in the ozone mixing ratio exceeds 60 ppbv km−1 (values of this gradient are generally in the range 50–70 ppbv km−1 near the tropopause).

[30] 2. The ozone mixing ratio is larger than 80 ppbv.

[31] 3. The mixing ratio immediately above the tropopause exceeds 110 ppbv. This criterion rejects layers of stratospheric air in the troposphere where the maximum mixing ratio is less than 110 ppbv.

[32] Clearly, this definition of the border between the troposphere and the stratosphere takes advantage of the fact that ozone has very different tropospheric and stratospheric concentrations, with a very sharp gradient at the tropopause. Although there are clear differences between the thermal tropopause and the ozone tropopause in individual soundings, we want to stress that the time behavior of the properties (height, pressure, and temperature) of both tropopause definitions is identical. This concerns both the trends as the change points in the time series. Therefore, for the remainder of the paper, we stick to the thermal tropopause.

3.3. UTH Definition

[33] Contrary to the thermal tropopause, there is no precise, well-established definition for the upper troposphere. And there are also numerous parameters describing “humidity.” Consequently, there is also a variety of definitions for the upper tropospheric humidity. First, we prefer to work with the (integrated) specific humidity as it represents the actual amount of water vapor in the atmosphere. The relative humidity, on the other hand, also depends on the temperature of the atmosphere through the saturation vapor pressure (see, e.g., Peixoto and Oort [1996] for more theoretical considerations on the relations between relative humidity, other moisture parameters, and temperature). It is important to note that the same trends occur when the UTH is defined in terms of relative humidity instead of specific humidity.

[34] Second, when using satellite humidity observations, the UTH is defined in absolute terms, that is, as the integrated specific humidity between fixed, absolute levels: for example, between 500 and 300 hPa in the work of Soden et al. [2002], and between 500 and 200 hPa in the work of Milz et al. [2009]. Of course, satellite observations usually do not provide direct or precise information about the tropopause location, so there is no other option for defining the upper troposphere. However, if information about the tropopause location is present, as is the case for radiosonde observations, the upper troposphere can be defined with the tropopause as an upper limit and a height, temperature, and pressure relative to the tropopause as a lower limit [e.g., tropopause height minus 3 km in Figure 1 (right)]. Defining the UTH relative to the tropopause has many advantages with respect to the UTH defined in absolute terms:

[35] 1. We prevent the potential mixing of moist upper tropospheric air with dry lower stratospheric air (beyond the detection limit of the radiosonde humidity sensor) in our analysis. The mean tropopause pressure at Uccle, calculated for the uniform time series, is 229.34 ± 45.18 hPa (1σ), so we would include for some soundings lower stratospheric air in the UTH value, if the upper troposphere were to be defined between 500 and 200 hPa.

[36] 2. The upper troposphere always has the same thickness (in km, °C, or hPa, with only the last related to the mass), so we are sure that the UTH is not affected by the stretching or shrinking of the upper tropospheric layers.

[37] 3. Related to the previous point, we are independent of the dynamics of the troposphere (and stratosphere) in general and, for example, the lifting or descending of the tropopause.

[38] In Figure 1, the UTH time series, in monthly anomaly form, calculated from two different upper troposphere definitions, one in absolute terms and one relative to the tropopause, are presented. It should be obvious that the delimitation of the upper troposphere plays an important role in assessing the time behavior of the UTH, as large differences exist between both UTH time series. The time variation before 2001 is very similar [and nicely demonstrated by the monthly cumulative deviations in Figure 1 (bottom)]: the UTH decreased until around the year 1995 and then recovered until the years 2000–2001. On the other hand, the behavior after 2000 is very distinct: the “absolute” UTH more or less remains constant, whereas the “relative” UTH drops down after autumn 2001. In this paper, we first focus on this large discontinuity around autumn 2001 in the relative UTH time series. Exploring the nature of this apparent and rather sudden drying out of the upper troposphere, defined relative to the tropopause, might also illustrate why the same phenomenon does not arise in the upper troposphere, defined in absolute terms. We nevertheless stress again that, in our opinion, the UTH should be defined relative to the tropopause and, consequently, it is especially this time series of UTH that is studied in great detail.

3.4. Homogeneity Checks

[39] In this section, we check if the discontinuity in the UTH time series around autumn 2001 can be ascribed to an inhomogeneity in the uniform time series of radiosonde observations. In any case, we underline that the occurrence of this change point is independent of (1) the definition and the extension of the upper troposphere (ranging from 1 to 4 km below the tropopause, or from 100 to 300 hPa below the tropopause), (2) the used change-point test (all mentioned change-point tests found a statistically significant change point in autumn 2001), (3) the correction method used (the Miloshevich correction also resulted in an autumn 2001 change point), and (4) the used subset of the uniform database (allowing all launch times or considering only ozone soundings, three times a week, does not affect the presence of the autumn 2001 change point).

[40] We first investigate a nonphysical (e.g., changes in equipment, measurement technique) origin of this change point. As mentioned in section 2, Vaisala made changes in the sonde packaging in September 1998 and June 2000. Because we made an inventory of all metadata of the radiosondes launched since 1990, we dispose of the production dates of the individual sondes and we can calculate their ages at launch. The mentioned packaging changes might have led to a discontinuity (moist bias) in our time series in October 1999 and in November 2000, respectively. Even if we take the uncertainties of the change-point detection into account, these packaging changes cannot account for the autumn 2001 change point in the UTH time series. Moreover, after autumn 2001, a “dry bias” instead of a moist bias is observed. On the other hand, our PWM test statistic does not reveal any traces of change points around October 1999 and November 2000. Furthermore, there is no metadata event around autumn 2001 for the uniform radiosonde database; the storage conditions of the sondes or the technicians performing the radiosoundings did not change during the 1990–2007 time period.

[41] However, when we apply the PMW test to the time series of the sonde ages at launch, a change point in autumn 2001 is detected. On average, the sondes are older at launch after autumn 2001 than before. This can be explained by the fact that the operational synoptic radiosoundings (performed daily by our weather office, twice a day, except when an ozone sounding took place) switched from Vaisala's RS80-A to RS90-H in autumn 2001. As we only consider RS80-A radiosondes launched at 1200 UTC in this study, this radiosonde type change gives rise to a reduced frequency of radiosonde launches (three times a week, at 1200 UTC; see also section 2), next to the even more important increase of sonde ages after autumn 2001. Indeed, our radiosonde launch practice was such that radiosondes used for ozone soundings were on average older than the synoptic radiosondes. As noted in section 2, the dry bias in the UTH after autumn 2001 might be caused by the older radiosondes used due to the sensor aging and the chemical contamination errors. This would also mean that the proposed correction schemes for these errors by Wang et al. [2002], which we included in the so-called Miloshevich correction, are insufficient, because, also in the Miloshevich-corrected UTH time series, a dry bias is noted after autumn 2001. But we believe that there exist even stronger arguments which rule out the autumn 2001 change point in the sonde's ages as the major contributor to the UTH change point. First of all, the autumn 2001 change point is also present in the UTH database formed by only considering ozone radiosoundings. This subset of the uniform radiosonde database has no change point in the sonde ages time series, so also not around autumn 2001. Second, there is no autumn 2001 change point in the precipitable water time series. Nevertheless, this variable is most susceptible to the aging and the chemical contamination of the humidity sensor. So the UTH time series should also not be so strongly affected by the radiosonde age time series. Third, the autumn 2001 change point also arises in the time series of other atmospheric variables, like the tropopause temperature, pressure and height, and tropospheric temperatures, although it is not statistically significant in all of these cases. As far as we know, radiosonde temperature and pressure sensor aging or contamination are not as much of a major issue as they are for the humidity sensor, so the sonde age change point is likely not responsible for the temperature and pressure change points around autumn 2001. Moreover, to our knowledge, Vaisala has never made modifications in their radiosonde temperature, pressure, and humidity sensors all at once within a given radiosonde type.

[42] It is important to make another consideration. As we mentioned in section 2, serious doubts exist about the quality of RS80-A humidity measurements at low temperatures. Therefore, we removed humidity profiles resulting from an iced humidity sensor and we applied the best available correction method, the one developed by Leiterer et al. [2005]. Suortti et al. [2008] pointed out that after this correction, for T < −50°C and at high RH, the dry bias did diminish considerably, but on average there still remained an ≈5% RH dry bias in the upper troposphere. However, we are convinced that the UTH trend captured by the Uccle RS80-A radiosondes (see Figure 1) reflects a real UTH climatology. Indeed, when descending through the upper troposphere, and hence enhancing the reliability and data quality of the radiosonde humidity measurements, the described humidity trend persists. This is obvious from Figure 2, in which the (specific) humidity trends are shown for layers of 10°C thickness and with top temperature equal to the tropopause temperature +n*10° C (n = 0,1,2,3).

Figure 2.

Uccle time series of monthly anomalies of the integrated specific humidity for upper troposphere layers of 10°C thickness and with temperatures in the intervals (a) tropopause temperature (Ttropo) < T < Ttropo + 10°C, (b) Ttropo + 10°C < T < Ttropo + 20°C, (c) Ttropo + 20°C < T < Ttropo + 30°C, and (d) Ttropo + 30° C < T < Ttropo + 40°C for the different subplots. Each time series has a statistically significant change point in autumn 2001, and the linear regression lines before and after these change points are shown. Red lines are used for positive trends, and blue lines for negative trends. Solid lines denote statistically significant trends, and dashed lines are statistically insignificant trends. The different layers, from top to bottom, have the following mean pressures and thickness of geopotential height (where minima and maxima are denoted in parentheses): 262 hPa (204–319), 328 hPa (263–395), 397 hPa (319–478), 482 hPa (384–587), and 1.522 km (1.224–1.884), 1.205 km (1.088–1.420), 1.218 km (1.115–1.447), 1.330 km (1.139–1.538).

[43] To summarize, we rule out any instrumental cause for the drop in the UTH around autumn 2001, and we ascribe it to natural variability. The main issue in the remainder of the paper is then to find the origin of this UTH decrease, given the information available from the radiosonde data.

4. Analysis of the Uniform Uccle Time Series

[44] In this section, we describe the time variations (trends and change points) present in the uniform Uccle 1990–2007 database of the UTH (section 4.1) and related properties such as the tropopause (section 4.2), double tropopause occurrences (section 4.3), and tropospheric (section 4.4) and lower stratospheric (section 4.5) variables.

4.1. UTH

[45] The similar time behavior given by both UTH definitions before 2001 (see Figure 1) is also in agreement with the observations of the upper tropospheric water vapor between 300 and 500 hPa from the NASA Water Vapor Project and from TIROS Operational Vertical Sounder products [see Soden et al., 2002, Figure 2]. In their paper, they explain the upper tropospheric drying in the early 1990s as a response to the Mount Pinatubo eruption in June 1991. Owing to the volcanic aerosols, the reduced solar heating led to a global cooling of the lower troposphere. Associated with this cooling was a reduction in the global water vapor concentrations, which closely tracked the decrease in temperature [see Soden et al., 2002, and references therein]. The authors demonstrated that their GCM reproduces the observed temperature profile changes only if the water vapor feedback (through the radiative calculations scheme) is turned on. The time variability of the UTH in the second half of the 1990s can then be interpreted as the return to normal values before the eruption.

[46] In contrast to the UTH time variability before ∼2000, the behavior after 2000 is not understood. However, not only in the upper troposphere but also in the lower stratosphere a substantial, persistent drop in humidity since 2001 is found in both global (60°N–60°S) satellite observations from the Halogen Occultation Experiment (HALOE, at 82 hPa) and balloon observations at Boulder, Colorado (40°N) [Randel et al., 2006]. This feature is complemented by an anomalously cold tropical tropopause and a decrease of ozone near the tropical tropopause during this period and is also believed to be associated with enhanced deep convection between 20°N and 20°S [Rosenlof and Reid, 2008] and with an increase of total stratospheric NO2 in the tropics after 2001 [Pastel et al., 2009]. These phenomena are ascribed to an enhanced tropical upwelling (Brewer-Dobson) circulation after 2001 [Randel et al., 2006; Garcia and Randel, 2008], caused by an enhanced planetary wave driving [Dhomse et al., 2008]. This is suggested to be a result of enhanced mixing in the extratropics, leading to additional air being drawn from the lower stratospheric tropics and causing cooling in the tropical tropopause region due to adiabatic expansion and thus reducing water vapor values [Jones et al., 2009].

[47] Of course, if there is a link between the UTH drop at Uccle in autumn 2001 and the lower stratospheric humidity decrease after 2000 and, consequently, this event is a larger-scale climate event, it would be very hard to infer any causes or mechanisms from just one (or even a few) station's data. Therefore, first we focus on the trend analysis of some local, related variables and try to find a mechanism for the UTH drop at Uccle on a local (or regional) scale. As the humidity drop in autumn 2001 is only present when the upper troposphere is defined relative to the tropopause, we first concentrate on the time behavior of the tropopause itself to find the physical phenomenon that drives the UTH discontinuity.

4.2. Tropopause Properties

[48] First, we want to mention that the tropopause at Uccle has a mean height of 11.02 ± 1.36 km and a mean temperature of −59.00 ± 6.34°C for the uniform time series. The close connection between the 1990–2007 time series of the tropopause and the relative UTH is nicely demonstrated by the correlation analysis. The UTH, defined in a layer extending from the tropopause to 3 km below it, for instance, is positively correlated with the tropopause temperature (the linear Pearson correlation coefficient, R2, equals +0.68) and negatively correlated with the tropopause height (R2 = −0.72). As already marginally mentioned, the autumn 2001 change point is also present in the time series of the tropopause temperature, pressure, and height, though not statistically significant at the 90% confidence level in the case of the temperature and height. Before autumn 2001, the tropopause height was descending (especially due to a dip in the year 2001; see Figure 3) and the pressure increasing, both significantly, and heating up (not significantly). Around November 2001, the tropopause started to lift up (see again Figure 3) and cooled down. Hence, from then on, the opposite trends (not significant) or no trends are observed. Clearly, the tropopause properties (height and pressure, temperature) are highly intercorrelated.

Figure 3.

(top) Uccle time series of monthly anomalies of (left) the tropopause height and (right) the relative frequency of double tropopause events. The mean before and after the detected change point are shown in gray. The linear regression lines for (left) the time periods before and after the detected change point and (right) the entire time period are also shown. Red lines are used for positive trends, and blue lines for negative trends. A solid line denotes a statistically significant trend, and a dashed line a statistically insignificant trend. The statistical significance of the trends is investigated by Spearman's test. Green lines denote the zero anomaly lines. Over the entire time period, the tropopause height is increasing at a rate of 30 m/decade (single linear regression slope) or decreasing at a rate of 177 m/decade (incremental). (bottom) Respective monthly cumulative deviations of the anomalies shown (Figure 3, top) from the mean of the anomalies. Green lines denote the zero lines.

[49] In general, the long-term variability of the tropopause can be of stratospheric and/or tropospheric origin [e.g., Santer et al., 2003; Seidel and Randel, 2006]. Indeed, the tropopause height is negatively correlated with lower stratospheric temperatures and positively correlated with tropospheric temperatures (see Figure 4, which is a remake of Figures 6 and 7 from the work of Seidel and Randel [2006] for the Uccle station and only for the 1990–2007 time period): the tropopause rises (descends) with a warming (cooling) troposphere and cooling (warming) stratosphere, which is in midlatitudes due to balanced dynamical structure in cyclones and anticyclones [Seidel and Randel, 2006, and references therein]. Additionally, there are other possible causes of the tropopause long-term variability (rising) closely related to the tropospheric and stratospheric variability but nevertheless with their own featuring, such as the broadening of the tropical belt and the poleward movement of tropospheric jet streams [Seidel et al., 2008], the acceleration of the Brewer-Dobson circulation under rising concentrations of greenhouse gases [Garcia and Randel, 2008], and the strengthening of the meridional thermal gradients in the upper troposphere/lower stratosphere (UTLS) at the subtropics and midlatitudes and consequently an increase of the UTLS wave baroclinicity in these regions [Castanheira et al., 2009]. This latter phenomenon is also associated with an increase in the frequency of double tropopause events.

Figure 4.

Vertical profile of the correlation between temperature anomalies at a given level and tropopause height anomalies, calculated for the uniform database at Uccle. The different symbols indicate the presence of a (significant) change point in the temperature anomaly time series around autumn 2001.

4.3. Double Tropopauses

[50] Before going into more detail about the possible causes for the time behavior of the tropopause, we take some time to describe the occurrence of double tropopauses above Uccle. Double tropopauses are associated with a characteristic break in the thermal tropopause near the subtropical jet, wherein the low-latitude (tropical) tropopause extends to higher latitudes, overlying the lower tropopause [Randel et al., 2007]. At Uccle, for the uniform time series, a double tropopause is located on average at about 17.05 ± 3.37 km high (which is derived through the atmospheric hydrostatic equation which uses the temperature, pressure, and humidity as data inputs, with a mean pressure of 97.21 ± 44.36 hPa) and has a mean temperature of about −60.34 ± 6.13°C. The frequency of the double tropopause events in Uccle is about 57% in winter, 26% in spring, 18% in summer, and 35% in autumn. Both these numbers and their seasonality agree with the results mentioned in the studies by Añel et al. [2007] and Randel et al. [2007]. In this paper, we are mostly interested in the time variability of the frequency of double tropopause occurrences, shown in Figure 3 in terms of their monthly anomalies and cumulative deviations. It is striking that the monthly cumulative deviations reveal a rather similar time behavior as the UTHs before the year 2000 (see Figure 1) but do not show a drop around autumn 2001. Indeed, the frequency of double tropopause events decreased in the early 1990s, started to increase in the second half of the 1990s, and was followed by a leveling off after the year 2002. The net result over the entire 1990–2007 time period is no significant trend in the double tropopause occurrences.

[51] The analysis of the variation in time of the height, pressure, and temperature of the second tropopause is hampered by the reduced frequency of radiosonde launches, and hence an increased scatter, after November 2001. However, we are still able to detect an increase in the second tropopause height at the end of 2001, in agreement with the lifting of the first tropopause. On the other hand, we can study the effect of the presence of a second tropopause on the properties of the first tropopause. Therefore, we split our uniform database into (1) a subset containing all measurements in which a single tropopause is detected and (2) a subset of measurements characterized by the occurrence of at least a double tropopause. Although there is an increased scatter in the monthly anomaly time series of (first) tropopause properties of these two subsets with regard to the entire database of measurements, the general tropopause time variability of both subsets is very similar and is also nearly identical to the tropopause behavior described so far.

4.4. Tropospheric Variables

[52] We now come back to Figure 4. Compared to Figures 6 and 7 in the work of Seidel and Randel [2006], we add an extra feature by marking if the temperature time series exhibit a (statistically significant) change point around autumn 2001. It turns out that this change is present in the time series of all tropospheric temperatures from 850 to 350 hPa. In these cases, the change point is also a trend change point (see Figure 5 for the temperature at 500 hPa): before autumn 2001, these tropospheric temperatures have a tendency to decrease; around autumn 2001, there is an increase in the tropospheric temperatures and, afterward, a tendency to increase or to remain constant. These trends are opposite to the tropopause temperature trends. Overall, the troposphere warms by, for example, 0.28°C/decade at 500 hPa if quantified by the single linear regression slope or cools at −0.19°C/decade if expressed in terms of the total incremental change (see section 3.1).

Figure 5.

Uccle time series of monthly anomalies of (top left) the surface temperature, (top right) the temperature at 500 hPa, (bottom left) the thickness of the free troposphere, and (bottom right) the temperature at 70 hPa. In case of a relevant and statistically significant change point, the means and the linear regression lines for the time periods before and after this change point are shown in gray (means) and in color (linear regression). When there is no change point, the linear regression lines for the entire time period are shown. Red lines are used for positive trends, and blue lines are used for negative trends. A solid line denotes a statistically significant trend, and a dashed line denotes a statistically insignificant trend. The statistical significance of the trends is investigated by Spearman's test. Green lines denote the zero anomaly lines. For the entire time period, the temperature at 500 hPa and the thickness of the free troposphere increase at a rate of 0.28°C/decade and 4.84 m/decade, respectively, or decrease at incremental rates of −0.19°C/decade and −6.41 m/decade, respectively.

[53] Another interesting tropospheric parameter for which we analyzed the time behavior is the thickness of the free troposphere. This variable is defined here as the difference between the geopotential heights at 300 and 700 hPa. As could be expected from Figure 4, a (significant) change point in the time series occurs in autumn 2001. This is also a trend change point: prior to this date, the free troposphere shrinks (not significantly), followed by a (nonsignificant) stretching afterward (see Figure 5). Around autumn 2001, the free troposphere is stretched out rather suddenly. If the thickness of the geopotential height is computed for the 1000–400 hPa pressure interval, as in the work of Añel et al. [2006], a similar time variability is found, with a (trend) change point around autumn 2001. A last point to consider is the strong positive correlation between the thickness of the free troposphere and the surface temperature (R2 = 0.68), although the surface temperature does not exhibit a similar trend (see again Figure 5).

4.5. Lower Stratospheric Variables

[54] We now study the time behavior of some lower stratospheric parameters. As pointed out again recently by Son et al. [2009], the cooling of the lower stratosphere associated with ozone depletion has a large impact on the tropopause height (e.g., next to the tropospheric warming due to greenhouse gas increases [see, e.g., Santer et al., 2003]). The temperatures observed in the lower stratosphere with the uniform Uccle radiosonde data set decrease in time in the period 1990–2007: the higher up in the stratosphere, the more significant the cooling becomes (see Figure 5 for the temperature time series at 70 hPa). The different change-point tests do not agree about one significant change point, but from our analysis it stands out that the early years of the time period (before the year 1998) contribute most to the overall cooling. This is also the period in which the stratospheric ozone above Uccle has not started its recovery yet. Additionally, the thickness of the geopotential height for the lower stratospheric 100–50 hPa layer (introduced by Añel et al. [2006]) also remains more or less constant during the last two decades. So, the lower stratospheric temperatures and thickness do not reveal any event or change at the end of 2001. Although the study of the stratospheric ozone time series of Uccle is far beyond the scope of this paper (it is the subject of a forthcoming paper), we want to mention here that we also do not find any discontinuity or trend change around autumn 2001 in the (lower) stratospheric ozone content, even after detrending the data with the dominant component of the ozone decadal variability. Apparently, at Uccle, this is the equivalent effective stratospheric chlorine (EESC) content.

5. Relevance of the Autumn 2001 Change

[55] The existence of a change point in autumn 2001 in the Uccle tropospheric temperatures, tropopause properties, UTH, and (free) tropospheric thickness is a very interesting feature on its own, but is it also present in the time series of observations at other midlatitudes? And how exceptional is it in a larger time perspective? We try to answer these two questions in sections 5.1 and 5.2.

5.1. Spatial Uniformity

[56] In this section, we examine to which extent the studied Uccle time series trends are representative for the European midlatitudes. We especially focus on the presence of a change point around autumn 2001 in the time series of other European radiosonde stations. Therefore, we selected radiosonde stations within 10° latitude from Uccle and with a longitude range of −10° to +30° and downloaded their data provided in the Integrated Global Radiosonde Archive (IGRA) described by Durre et al. [2006]. In this section, we deal with the tropopause temperature time series of these stations, rather than with (upper tropospheric) humidity data records, as the data homogeneity is not guaranteed for an individual IGRA station. However, at least for the Uccle data, there was a clear correlation between the (relative) UTH and the tropopause temperature trends.

[57] Certainly, there is extensive evidence that the long-term variation in radiosonde temperatures as well is affected critically by inhomogeneities introduced through changes in instruments and measurements practices [Lanzante, 2009, and references therein). But, as the existing “homogenized” data sets [e.g., Haimberger et al., 2008; McCarthy et al., 2008; Sherwood et al., 2008] apply only to the mandatory pressure levels of radiosonde analysis, these data sets do not provide enough vertical resolution for a trend analysis of tropopause temperatures [Rosenlof and Reid, 2009]. Recently, McCarthy et al. [2009] extended their homogenization procedure to tropospheric humidities (standard levels up to 300 hPa) and introduced as such a greater consistency between temperature and specific humidity trends from day and night observations. However, as this homogenization system is neighbor-based, we chose not to use such a database to detect a single change point in different neighboring stations.

[58] The autumn 2001 change does occurs not only at Uccle, but also at other European stations. Examples for the tropopause temperature are given in Figure 6. The other three European radiosonde stations shown in Figure 6 lie in the vicinity of Uccle, and especially the overall resemblance between Uccle and De Bilt is very striking. As the autumn 2001 change is also a distinct feature in other European radiosonde stations (with different equipment, and instrumental changes occurring at other periods), the physical nature of the tropopause temperature drop at the end of the year 2001 stands out without any doubt. Furthermore, we also calculated the correlation coefficients between the monthly anomalies (“correlation between months”) of the entire time series of tropopause temperatures of the selected IGRA stations with the corresponding Uccle data. A contour plot in the latitude-longitude field of these correlation coefficients is shown in Figure 7. This figure shows a rather concentric contour pattern around Uccle, meaning that, also for the entire time series of tropopause temperatures (typically starting in the late 1960s), (1) the Uccle data series (trends) are typical for Western Europe and (2) deviations from the Uccle data series (trends) occur rather smoothly, both in latitude and longitude.

Figure 6.

Time series of moving averages of the monthly anomalies of tropopause temperatures for Uccle, De Bilt (WMO code 06260, NL, 52°6′N, 5°11′E), Larkhill (WMO wode 03743, UK, 51°12′N, 1°48′W), and Meppen (WMO code 10304, D, 52°44′N, 7°20′E). The data of the last three stations are taken from the IGRA database. Autumn 2001 is marked with a vertical dashed line.

Figure 7.

Latitude-longitude contour plot of the correlation coefficients between the tropopause temperature monthly anomalies of selected IGRA radiosonde stations (red crosses) and the Uccle station (red asterisk) for the longest time period possible.

[59] However, the contours in Figure 7 have a more longitudinal structure which could resemble the fact that the features of the global tropopause are more homogeneous in longitude than in latitude. Moreover, the contour plot might also suggest a discrimination of trends between the western, more maritime, and eastern (continental) part of Europe. This might point to the fact that the autumn 2001 event is not a large-scale natural variability event.

5.2. Extended Study

[60] To put the autumn 2001 change in a larger time perspective (“how exceptional is this change in tropospheric and tropopause properties”), we now consider the entire database of radiosonde measurements at Uccle (at 1200 UTC), starting in 1963 (1968 for the tropopause identification). But we have another reason to deal with a longer time series. The autumn 2001 change point is, for a number of tropospheric and tropopause variables, a trend turning point, or at least it breaks the preceding trend. Therefore, the extension of our Uccle data set helps to study the relevance of either time period. This entire database of vertical profiles is gathered by different radiosonde types, so we do not investigate the humidity field trends but limit ourselves to the temperature, height, and pressure measurements. These are estimated to suffer less from sensor changes than the humidity measurements, although their long-term variation is also affected by inhomogeneities [Lanzante, 2009, and references therein], as already discussed at the beginning of section 5.1.

[61] Generally, it comes out that the surface temperature at Uccle increases at a rate of 0.50 ± 0.06°C/decade, the lower tropospheric temperature (at 500 hPa) increases 0.18 ± 0.06°C/decade, the tropopause cools by −0.33 ± 0.08°C/decade (see Figure 8, for the period 1968–2009), and the lower stratospheric temperature (at 100 hPa) decreases by −0.66 ± 0.06°C/decade. The tropopause height increased 36 ± 19 m/decade. All these trends are statistically significant. The numbers are also in line with the literature reports for Northern Hemisphere midlatitudes [see, e.g., Seidel and Randel, 2006; Sherwood et al., 2008; Schmidt et al., 2008]. So, the aforementioned opposite (but not always statistically significant) trends in the 1990–2001 decade for the restricted RS80-A database represent a temporary disruption of the general trends.

Figure 8.

Uccle time series of the tropopause temperature monthly anomalies for the 1968–2009 period. A change point is found around October 1988; the autumn 2001 change point is indicated by the vertical dashed line.

[62] In the entire radiosonde database, the autumn 2001 change point is less pronounced: the major features of the entire time series are the strong overall decreasing or increasing trends (leading to an artificial change-point detection in the middle of the time series). After detrending the time series, no significant change points are identified in the tropospheric temperatures or the tropopause features. For the surface and the lower stratospheric temperatures as well as the thickness of the free troposphere, a change point at the end of the 1980s is present. Only in the case of the surface temperature a trend reversal does occur in this period (i.e., the surface temperature decreases significantly before November 1988 and increases significantly thereafter, but this increase is mostly due to the high values during the last decade). The same change point is present in the time series of surface temperatures retrieved by the weather station at Uccle. For the lower stratospheric temperatures, the end of the 1980s (and beginning of the 1990s) marks the end of the strong cooling, followed by a slight increase of temperatures. Here, the lower stratospheric ozone recovery since the mid 1990s certainly plays a role, but the change from VIZ to RS80-A radiosondes in January 1990 has without doubt a strong effect, too.

[63] So, to conclude this section, we mention that the autumn 2001 change point in temperatures and tropopause properties is less pronounced when a longer time series of radiosonde measurements at Uccle is considered. This is an important consideration, as the HALOE stratospheric water vapor database, in which a prominent drop was observed in 2001, only spans from 1992 onward.

6. Discussion and Conclusions

[64] On the basis of a uniform (one radiosonde type, identical vertical resolution) data set of corrected radiosonde vertical profiles, we investigated the properties of the upper troposphere above Uccle, Belgium, in the 1990–2007 time period. Over the entire period, no significant trends in upper tropospheric humidity are detected, which is in line with climate model predictions of a constant relative humidity but opposes the predicted specific humidity increase in the upper tropospheric levels (see section 1). As a matter of fact, up to the year 2000, the most prominent feature is a period of negative UTH anomalies in the early 1990s, which is a response to the Pinatubo eruption: volcanic aerosols led to a global cooling of the lower troposphere and reduced the global water vapor concentrations. In the late 1990s, the UTH then slowly recovered from the Pinatubo response. However, the most curious finding of this study is the drop in upper tropospheric humidity in autumn 2001, which marks an end to this significant moistening of the upper troposphere. We argued exhaustively that we do not find any instrumental or environmental cause in our data set and we ascribe it to natural variability.

[65] As this drop is prominently present in the time series of UTH defined relative to the tropopause (with the tropopause as the upper limit and height, temperature, and pressure relative to the tropopause) and absent in the absolute UTH time series (e.g., between 500 and 200 hPa), the cause of the UTH decrease is without doubt associated with the variability of the tropopause itself. As a matter of fact, the tropopause underwent a lifting and cooling around autumn 2001, with the opposite behavior before autumn 2001. A similar change in the tropopause properties around autumn 2001 is detected in the time series of other European radiosonde stations. This variability of the tropopause can be of tropospheric and/or stratospheric origin.

[66] In the time series of the Uccle tropospheric temperatures from 850 to 350 hPa, a change point around autumn 2001 also exists. Before this change point, the troposphere has a tendency to cool down; afterward, it has a tendency to warm up. Hence, only for the short time period of almost two decades, the association between a cooling (warming) troposphere and descending (ascending) tropopause is well established. In this context, we also mention the change in the thickness of tropospheric geopotential heights around autumn 2001: the tropospheric shrinking trend of the 1990s is interrupted by a stretching of the troposphere in 2001, which can be expected from the temperature records. As such, the tropospheric vertical movement (stretching and shrinking) might provide the link between the tropospheric temperature changes and the tropopause properties. The tropospheric dynamics (circulations) contributing to the change around autumn 2001 can then be summarized as follows: in autumn 2001, the tropospheric temperature rose significantly, so vertical turbulent motions or convection led to a considerable stretching out of the free troposphere. As a consequence, the tropopause is also lifted up and cooled down, and the upper tropospheric layers started to freeze-dry in the fall of 2001.

[67] This hypothesis puts all the pieces of the tropospheric puzzle together. However, the tropopause behavior is also known to be affected by stratospheric processes such as ozone depletion, which cools the lower stratosphere. In any case, in the time series of lower stratospheric temperatures, thickness of geopotential height, and (EESC-detrended) ozone amounts, we do not find any change around autumn 2001. The lower stratospheric temperatures hardly decrease during the studied time period of 17 years, possibly because of the onset of the ozone recovery in the second half of the 1990s. The time variability of height and pressure of the second tropopause also exhibits a steplike change in autumn 2001 and hence seems to follow the vertical motion of the first tropopause. The frequency of double tropopause occurrences varies in time with the UTH humidity, with the exception that the relative frequency of double tropopauses only started to level off after 2002. Pan et al. [2009] showed that the occurrence of double tropopauses is, at least frequently, associated with tropospheric intrusions of subtropical air into the extratropical lower stratosphere. As the frequency of double tropopause events can be interpreted as an indication for the strength of the UTLS wave baroclinicity [Castanheira et al., 2009], there is also a leveling of the increase in UTLS wave baroclinicity at Uccle after 2002.

[68] We elaborate more on the (meridional) cross-tropopause transport of air and we want to discuss the possible link between the autumn 2001 drop in UTH and tropopause temperature and the enhanced tropical upwelling after 2001, leading to a substantial, persistent, global decrease in stratospheric water vapor since the end of 2000 [Randel et al., 2006]. Climate model simulations indicate that a strengthening of the tropical upwelling, and hence an intensification of the Brewer-Dobson circulation, in response to climate change will lead to an enhanced downwelling at the (northern) midlatitudes [Li et al., 2008; McLandress and Shepherd, 2009]. This increased downwelling could inject dry stratospheric air, which enters the upper troposphere across the tropopause. This might be a possible contributing cause to the observed drop in UTH in autumn 2001 above Uccle, taking some time lag with the increased tropical upwelling into account. Jones et al. [2009] noted a time lag of about 6 months between the drop in the stratospheric water vapor in the tropics and in the 20–25 km altitude midlatitude bins. On the other hand, increased downwelling would also lead to a downward shift of the tropopause, which is likely to warm up then. Exactly the opposite tropopause behavior is observed around autumn 2001 in the Uccle data set. The consequences of the strengthening tropical upwelling on the midlatitude troposphere and stratosphere certainly deserve some additional research, from both modeling and observational points of view.

[69] However, in a more indirect way, the lower stratospheric water vapor might trigger the UTH variability. Increases in stratospheric water vapor act to cool the stratosphere but to warm the troposphere, whereas the reverse is true for stratospheric water vapor decreases [Solomon et al., 2010]. In our uniform time series, after 2001, we find higher mean tropospheric temperatures and similar mean stratospheric temperatures compared to the decade before, so the association of these temperatures with a possible drop in stratospheric water vapor above Uccle is not straightforward. But Solomon et al. [2010] showed also that the decrease in stratospheric water vapor concentrations after 2000 acted to slow the rate of increase in global surface temperature over 2000–2009 by about 25% compared to that which would have occurred due only to carbon dioxide and other greenhouse gases. In this sense, the stratospheric variability might also act to slow down (or spin up) the tropospheric circulation discussed earlier in this section. To conclude, we do not find any observational evidence of a direct link of the UTH drop around autumn 2001 with stratospheric processes. This is not very surprising, however, because if the autumn 2001 decrease in UTH is a large-scale climate event, a more global observational record at midlatitudes (and also of photolytic tracers), next to climate model simulation studies, are needed to establish and test mechanisms that give rise to abrupt and unexpected trend changes.

[70] For the last consideration we want to make, we return to Figure 8. The tropopause temperature time series seems to have some periodic behavior, under the form of a decadal oscillation. An obvious candidate to account for this cyclic variability in the tropopause and tropospheric properties is the 11 year solar cycle. Indeed, a number of independent analyzes reported on a relatively robust but modest influence of solar cycle forcing on the stratosphere-troposphere system. The lower stratosphere warms at solar maximum due to the solar UV effect on increased stratospheric ozone; the troposphere generally appears to warm and moisten during solar maximum conditions (see Gleisner et al. [2005], Salby and Callaghan [2006], van Loon et al. [2007], and many more references therein), but there are disagreements as to whether the tropical region warms, or primarily the subtropics through midlatitudes [Rind et al., 2008]. Two mechanisms have been suggested by modeling studies to explain the solar influence on the troposphere: the top-down dynamical response to the stratospheric variations and the bottom-up coupled ocean-atmosphere surface response [Rind et al., 2008; Meehl et al., 2009].

[71] However, when looking at Figure 9, in which the time series of the 10.7 cm solar flux, the Uccle tropopause temperature (1968–2008), and UTH are shown, only during the last considered solar cycle (roughly from about 1995 on), both the tropopause temperature and the UTH variability are in phase with the solar cycle. The variability of the entire tropopause temperature time series (starting in 1968) seems to be less coupled with the solar cycle. So, the solar cycle maximum in 2001 could be partly responsible for the autumn 2001 change in tropopause temperature and UTH, but the solar cycle variability alone cannot account for the observed variations, on more or less a decadal timescale, in the tropopause temperature. This correspondence in variability between the solar cycle and the tropopause temperature in the 1990–2007 time period is not restricted to the Uccle radiosonde database alone but also applies to other IGRA radiosonde stations (see section 5.1) located in the western part of Europe. As a matter of fact, there is a clear discrimination between the west (strong) and the east (weak) of Europe with regard to the correlation of the tropopause temperature (and temperature at 100 hPa) with the solar cycle during 1990–2007. This again underlines the representativeness of Uccle for only the western, maritime, part of Europe. However, for all chosen European IGRA stations, the entire time series of tropopause and lower stratospheric temperatures shows decadal variations which are not directly linked to the solar cycle.

Figure 9.

Time series of normalized, 2 year running means of the monthly anomalies of the 10.7 cm solar flux, the Uccle tropopause temperatures, and UTH. The autumn 2001 change is marked by a vertical dashed line. We consider 2 year running means in order to compare the variability at timescales larger than 2 years.

[72] The discussion about the possible influence of the solar cycle on the tropopause and UTH variability once more underlines the need for a long-term time series. Indeed, when considering the entire, although inhomogeneous, time series of radiosonde measurements, the change in tropopause properties around autumn 2001 becomes less pronounced and is not unique in time at all. These examples underscore the bottom line for any climate change research: there is a strong need for long-term, homogeneous time series of observations.


[73] This research was supported by the AGACC project (contract SD/AT/01A) funded by the Belgian Federal Science Policy Office. R. Van Malderen is now a fellow of the Solar-Terrestrial Center of Excellence (STCE), also funded by the Belgian Federal Science Policy Office. We thank U. Leiterer, H. Dier, and L. Miloshevich for providing their code to correct the humidity profiles and for their help with implementing it. This research was not possible without the commitment of the technical staff that performed the radiosoundings at Uccle throughout the years. We also want to thank the three anonymous reviewers for their in-depth comments which substantially improved the manuscript.