Three years of COSMIC RO soundings from both JPL and UCAR over the Arctic (65–90∘ N) were analyzed, with a special emphasis on the troposphere (e.g., below ∼ 10 km). The post-processed version of the JPL retrievals and the UCAR retrievals from 2008 to 2010 were used, where a consistent retrieval algorithm has been implemented throughout the data processing period. The COSMIC RO data are grouped into four seasons: winter – DJF (December–January–February); spring – MAM (March–April–May); summer – JJA (June–July–August); and fall – SON (September–October–November). Note that the month of December in the winter season is from the previous year (e.g., DJF 2008 denoted as December 2007, January 2008, and February 2008). The JPL retrieval algorithm has been detailed in Hajj et al. (2002) and Ao et al. (2012), whereas the UCAR retrieval algorithm can be found in Kuo et al. (2004). Different calibration (e.g., orbit and clock calibration; Wickert, 2002; Schreiner et al., 2009), retrieval algorithm, and quality control procedure (e.g., Ho et al., 2009) lead to some differences in the total throughput of COSMIC RO soundings from the two data centers. JPL retrievals generally yield slightly fewer soundings than UCAR retrievals. A total of 112 156 and 129 538 RO profiles are retrieved from JPL and UCAR, respectively. Additional discussions on the processing differences can also be found in Ho et al. (2009, 2012) and Steiner et al. (2013).

High-resolution radiosonde soundings from the ASCOS field campaign (red dots) throughout the cruise in the Atlantic sector of the Arctic from 3 August to 7 September 2008 (Tjernström et al., 2014) and the COSMIC RO (e.g., UCAR) soundings distribution (blue diamond) during the same time period over 75–90∘ N are shown in Fig. 1. Detailed description of ASCOS including the instrumentation and measurements can be found in Tjernström et al. (2014). Radiosondes were launched from the helipad of the icebreaker four times a day at approximately 00:00, 06:00, 12:00, and 18:00 Coordinated Universal Time (UTC). A total of 145 soundings were collected throughout the entire period of the cruise. Only one sounding was discarded due to missing latitude and longitude information. Each profile was interpolated onto uniform vertical levels with intervals increasing from 5 m in the lowest 1 km of the troposphere to 100 m in the stratosphere (Birch et al., 2012). In addition, the ERA-I reanalysis profiles over the Arctic were also analyzed. The ERA-I applied a T255 grid scheme with 0.75∘ horizontal grids (∼ 83 km near the Equator or ∼ 14 km near latitude ∼ 80∘ N), and 60 vertical layers (Dee et al., 2011).

One major limitation of using GPS RO sounding to study the lower troposphere is that not all the RO profiles reach the surface. Here we define the “penetration depth” as the minimum height of each individual RO sounding above the local surface. Note that the GPS height is the height above the mean sea level (MSL), which is referred to as the height above the geoid, i.e., the equipotential gravity surface height from a standard gravity model such as the Earth Gravity Model 1996 (EGM96). The high-resolution digital terrain elevation data (0.16∘ grid) used in this paper are also in reference to above the geoid. Thereafter, the penetration depth in this paper will be the GPS height above the local surface after subtracting the terrain elevation.

In the RO retrieval, the penetration depth of an individual occultation sounding is affected by the quality of the GPS RO signal at the receiver and the quality control criteria used. Several factors could result in the degradation of RO signal that leads to earlier termination of RO profiles before the sounding reaches the local surface, i.e., a positive penetration depth. For example, the topography along the GPS and the RO receiver line of sight could block the RO signals and lead to the early termination of the RO sounding. In addition, the RO signal could be degraded due to receiver signal tracking issues attributed to the presence of large vertical moisture variation in the lower troposphere (Ao et al., 2012).

In the JPL retrieval system, the ending of an RO sounding (i.e., the penetration depth) is determined as the transition point where the RO signal quickly degrades into an unusable noisy regime by fitting a step function to the transformed signal amplitude after the canonical transform inversion algorithm (Ao et al., 2012). A similar approach is applied in the UCAR retrieval as well but using a different transformed signal through the full-spectrum inversion (FSI) algorithm (Jensen et al., 2003; Kuo et al., 2004). With the implementation of open-loop tracking in the COSMIC mission, the receiver tracking errors in the lower troposphere have been significantly reduced (Ao et al., 2003; Sokolovskiy et al., 2006a, b). However, the non-uniform RO penetration depth across the globe remains and requires further investigation (Ao et al., 2012; Xie et al., 2012). It is important to note that the vertical resolution of the Level 2 RO refractivity retrievals in the lower troposphere is limited to be ∼ 200 m due to the vertical smoothing of the retrieval profiles. Therefore, the penetration depth at or below 100 m is essentially as good as reaching the surface.

The detailed description of the GPS RO technique has been covered extensively in several papers (Kursinski et al., 1997; Hajj et al., 2002; Anthes et al., 2008). Here we only summarize the key concepts of the retrieval processes. GPS RO senses the atmosphere by tracking the GPS radio signals that traverse the atmosphere as a moving receiver sets (or rises) behind the horizon relative to the transmitting satellite. The radio wave is refracted, and its travel time is delayed due to the variations of refractivity of the atmosphere. GPS RO precisely measures phase and amplitude of GPS signals that traverse the Earth's atmosphere. After the phases are calibrated by removing the GPS and LEO clock errors, a time series of excess phase at both GPS frequencies (e.g., L1 and L2) is derived. Then under the assumption of a local spherically symmetric atmosphere, the vertical profile of the bending angle (α) and the refractivity index (n) can be derived. In a neutral atmosphere, the refractivity (N=(n-1)×106) measured by GPS RO is related to pressure (P in mbar), temperature (T in K), and water vapor partial pressure (Pw in mbar) as the following Eq. (1) (Smith and Weintraub, 1953): N=77.6PT+3.73×105PwT2. Based on the RO refractivity, the temperature and humidity profiles can be derived with certain external information. In the upper troposphere and above, the second term (the so-called “wet term”) on the right-hand side of the refractivity Eq. (1) can be neglected. The so-called dry temperature (Tdry) can be derived as follows: Tdry=77.6PdryN. As the saturation vapor pressure decreases rapidly with decreasing temperature according to the Clausius–Clapeyron equation, the water vapor pressure Pw can be neglected in the upper troposphere, where temperature is low (e.g., T < 250 K; Kursinski et al., 2000; Hajj et al., 2002; Melbourne, 2004). Given the pressure (or more strictly, the dry pressure, Pdry, derived from the hydrostatic balance) and the RO refractivity profiles, the temperature can be retrieved. Accurate temperature profiles can be derived throughout the stratosphere down to the mid-troposphere or even lower altitudes depending on latitudes, where the water vapor is negligible.

In the middle or lower troposphere where the moisture is not negligible, the derivation of temperature and humidity becomes an underdetermined problem, which is generally referred to as the dry–wet ambiguity problem. To solve this issue, the JPL and UCAR data centers use different approaches.

In the JPL retrieval (e.g., Hajj et al., 2002), the moisture retrieval only starts when the dry temperature is over 250 K, i.e., when the moisture contribution to the refractivity becomes non-negligible (Kursinski et al., 1997). In the temperate and tropical regions where the water vapor is abundant with large uncertainty, the temperature profile from numerical weather prediction (NWP) model analysis usually is relatively better known and thus can be used to aid the water vapor retrieval. In practice, the nearest 6-hourly ECMWF global analysis temperature profiles are interpolated into each RO sounding location as a priori. Given the RO refractivity profile (Tdry > 250 K) and the a priori ECMWF temperature profile, the RO water vapor profile can be derived. On the other hand, given the RO refractivity and the a priori ECMWF water vapor profile, the RO temperature profile can also be derived. While the approach is relatively simple, the RO water vapor (or temperature) will contain both measurement uncertainty in RO refractivity and the uncertainty in the a priori ECMWF temperature (or water vapor).

The UCAR data center applied the optimal estimation of the water vapor, temperature, and pressure through a variational method (Kuo et al., 2004). The variational method combines the occultation measurements (e.g., refractivity) with the a priori (or background) atmospheric condition in a statistically optimal way (Zou et al., 1995; Healy and Eyre, 2000). For example, the optimal solution to the state vectors (e.g., T, Pw, and P) can be found by adjusting the state vector elements in a way that is consistent with the estimated background errors, to produce simulated measurement values that fit the observations to within their expected observational errors (Healy and Eyre, 2000).

Both JPL and UCAR temperature and humidity retrievals require the a priori information from models and thus are not independent measurements. Instead, measurements (e.g., refractivity, dry temperature) are model-independent observations. Note that the errors of the geophysical parameters derived from the 1-D-variational method (UCAR) could become more challenging to interpret because the errors are a combination of the a priori model background errors with the RO measurement errors.

Note that a fixed latitude–longitude grid (2.5∘ × 2.5∘) near the Equator has an area of ∼ 78 400 km2, which is equivalent to a square cell with sides of ∼ 280 km. However, the length of the fixed grid longitude is significantly reduced at higher latitude, especially near the polar region. To accommodate such a significant reduction of grid area at higher latitude resulting from the fixed latitude–longitude grid, the equal-area-grid mapping method is applied. A fixed 2.5∘ latitude interval (∼ 280 km) is chosen, and each latitude band is evenly divided to have each grid area close to 78 400 km2. Therefore, each grid will have a roughly equal area across the polar region with a fixed 2.5∘ latitude interval and a variable longitude interval (increasing at higher latitude or toward the poles). The mapping technique will result in only 3 grids within 87.5–90∘ N latitude bands; 9 grids within 85–87.5∘ N; and, increasing further south, the maximum of 58 grids within 65–67.5∘ N (Rossow and Schiffer, 1991).

COSMIC RO refractivity profiles over the Arctic (65–90∘ N) from both JPL and UCAR are analyzed. The 3-year (2008–2010) median seasonal variations of the penetration depth for both centers and their associated RO profile sample numbers are shown in Fig. 2. Both centers show rather homogeneous sampling with ∼ 60–120 soundings in each equal-area grid (equivalent to the area of a square cell of 280 km × 280 km). JPL shows slightly less sounding at each grid than UCAR, which is consistent with the overall smaller sample size.

A very similar geographical pattern of RO penetration depth is seen for both centers in all seasons, with rather uniform deep penetration (∼ 0–300 m) over the Arctic Ocean and generally poorer penetration over land and islands. The distinctly poorer penetration (i.e., larger penetration depth) in summer than other seasons indicates the degradation of RO retrieval due to the increase of the lower-tropospheric water vapor during the warmer summer season. In addition, central Greenland shows deep penetration (100–300 m) likely due to the relatively flat terrain. Over the land surrounding the Arctic Ocean, the penetration depth varies from 100 to 900 m with slightly poorer penetration in JPL data, over the northwestern edge of Greenland. It is worth noting that ∼ 100 m penetration depth is essentially as good as reaching the surface as both JPL and UCAR centers apply ∼ 200 m vertical smoothing on the RO bending/refractivity retrieval.

Figure 3 further illustrates the seasonal-mean percentage of RO profiles reaching different altitude (binned into 100 m vertical intervals) above the surface, with land and ocean separated. For both centers, a high percentage of COSMIC RO soundings (∼ 85–90 %) penetrates below 1 km in all seasons. But a sharp decrease by ∼ 8–20 % is seen in RO profiles reaching 0.5 km above the surface. Above 1 km, no significant difference in RO penetration is seen between the land and the ocean. However, a slight decrease in the percentage of RO profiles reaching below 1 km over land could be due to the topography blocking effect.

The UCAR retrieval shows generally a higher percentage of profiles (∼ 92–95 %) penetrating down to 1 km above the surface than the JPL (∼ 88 %) in all seasons except the summer, when the percentage drops to ∼ 84 % for both centers. UCAR also shows a generally higher percentage of RO soundings extending to the height levels within 0.5 to 1 km, except the summer season, when JPL shows a slightly higher percentage of profiles extending to all levels below 1 km. In addition, JPL also shows a slightly higher percentage of RO profiles extending below 0.3 km than UCAR in all seasons, with a maximum of ∼ 5 % difference in summer.

The much higher percentage of deeply penetrating RO profiles below 0.5 km over the very dry Arctic Ocean (65–80 %) than the moist tropics of 20–30 % (Ao et al., 2012; Xie et al., 2012) strongly indicates that the increase of lower-tropospheric moisture could result in larger receiver tracking errors and the early termination of the RO profile. The poorer penetration in the warmer and moister summer Arctic further confirms the moisture impact on the RO penetration issue.

To quantify the structural uncertainty of the COSMIC RO soundings over the Arctic (65–90∘ N), the RO retrieval difference between JPL and UCAR data centers is analyzed. The RO soundings in both winter (DJF) and summer (JJA) seasons of 2008 are investigated. Although both data centers start from the same Level 1 raw COSMIC measurements, their retrieval algorithms are slightly different in terms of calibration, retrieval, and quality control processes (Ho et al., 2012). A total of 4782 pairs of common soundings in winter (DJF) and 8375 pairs in summer (JJA) of 2008 have been identified.

All RO sounding refractivity, temperature, and specific humidity (e.g., N, T, q) from both centers are interpolated into 100 m vertical intervals. The ensemble mean, standard deviation and standard error of the inter-center differences, and the sampling number are shown in Fig. 4. The statistics at three selected altitude ranges (e.g., from the surface up to 3, 5, and 10 km) are summarized in Table 2. It is also worth noting that the significantly decreasing number of common refractivity soundings in the lowest 3 km is mainly due to the limited RO sounding penetration depth as discussed in Sect. 4.1. The statistics for the lowest 200 m were discarded due to the significant drop of the number of RO observations near the surface, limited by the ∼ 200 m vertical resolution of RO retrievals. For the RO temperature and specific humidity profiles, the variation of common sounding numbers at different altitudes is mainly limited by the availability of JPL retrievals. For example, UCAR retrieves both T and q at all altitude where refractivity retrievals are available. However, JPL only retrieves tropospheric temperature at altitudes cooler than a threshold of about 250 K (see Sect. 3.2), leading to a significant reduction in the number of JPL temperature retrievals in the lower troposphere, especially in the warmer summer season (Fig. 4d). To account for the number of pairs that change with the altitude, the standard error is also computed (i.e., the standard deviation divided by the square root of the sample size), which is an estimation of the difference of the sample mean from the population mean affected by the sample size.

In winter, small but persistent biases are seen in the JPL retrieval as compared to the UCAR retrieval below 3 km in refractivity (∼ 0.04 %), temperature (0.54 K), and specific humidity (0.03 g kg-1), with a standard deviation of 0.19 %, 0.45 K, and 0.33 g kg-1, respectively. In summer, the difference for refractivity and specific humidity slightly changes, with a mean difference of 0.02 % and -0.05 g kg-1, and a corresponding standard deviation of 0.44 % and 0.50 g kg-1, respectively.

The mean fractional refractivity difference between JPL and UCAR is ∼ 0.06 in winter and 0.07 % in summer below 10 km, which is consistent in magnitude with the inter-center comparison of CHAMP RO data within 8–12 km over 60–90∘ N (Ho et al., 2009). No significant difference is seen between the winter and the summer. The mean temperature difference is 0.38 in winter and 0.61 K in summer below 10 km, with a corresponding standard deviation of 0.40 and 0.62 K, respectively. The mean specific humidity difference is 0.03 in winter and 0.002 g kg-1 in summer below 5 km, with a corresponding standard deviation of 0.34 and 0.41 g kg-1, respectively. The standard errors are generally small, except when there is a sharp decrease in sample size resulting from fewer JPL temperature retrievals below 5 km in summer (Fig. 4d) and moisture retrieval above ∼ 3 km in winter (Fig. 4e). The generally small standard error indicates reliable statistics given the large sample size at most altitude levels.

To quantify the parametric uncertainty of the COSMIC RO soundings over the Arctic (65–90∘ N), both JPL and UCAR retrievals are collocated with a total of 144 radiosonde soundings collected from the ASCOS summer field campaign in 2008. A total of 65 JPL profiles and 68 UCAR profiles were found to be collocated with the radiosondes within 3 h and 300 km. The ensemble mean difference, standard deviation, and standard errors are shown in Fig. 5. The statistics of the comparisons at three selected altitude ranges (e.g., 3, 5, and 10 km from the surface) are listed in Table 3.

Compared with the radiosonde from the surface up to 10 km, the COSMIC RO refractivity shows an overall median bias of 0.10 % in JPL retrievals (Fig. 5a1), -0.20 % median bias in UCAR retrievals (Fig. 5a2), and -0.21 % overall median bias in ERA-I refractivity retrievals (Fig. 5a3). However, the overall JPL median bias (0.10 %) is comprised of both positive bias (e.g., ∼ 0.2 % in 6–8 km) and negative bias (e.g., ∼ -0.5 % below 4 km). Most of the negative bias is coming from the lower troposphere in JPL and UCAR. For example, JPL refractivity retrievals show a negative bias below ∼ 2 km, which increases to a maximum of about -1 % in the lowest 1 km above the surface. Similarly, UCAR refractivity retrievals also show a negative bias, albeit starting from ∼ 5 km down to the surface, with a smaller maximum bias of about -0.5 %. On the other hand, the ERA-I shows a generally negative bias above ∼ 1 km but a small positive mean bias below ∼ 0.5 km.

Figure 5b1 shows a warm bias of 0.62 K in JPL temperature retrievals from 10 km down to ∼ 4.5 km, where JPL stops retrieving temperature. A sharp increase in standard error in JPL temperature below 5 km is due to the sharp drop in the available JPL temperature retrieval. On the other hand, UCAR temperature shows a smaller overall median bias of 0.53 K below 10 km, with an increasing bias below 2 km up to near 2 K (Fig. 5b2). ERA-I temperature shows a smallest overall median bias of 0.12 K below 10 km (Fig. 5b3), with a warm bias (∼ 1K) below 1 km and small cold bias above, which are consistent with Wesslén et al. (2014).

For specific humidity, the JPL retrievals exhibit negative biases from the surface up to ∼ 2 km and transition to positive biases above, with a median bias of about 0.07 g kg-1 from the surface up to 5 km (Fig. 5c1). The UCAR specific humidity has a negative overall bias of -0.03 g kg-1, mostly coming from below ∼ 5 km (Fig. 5c2). The ERA-I specific humidity has an overall bias of 0.004 g kg-1, mostly coming from below ∼ 0.6 km (Fig. 5c3). It is also worth noting that the increasing biases and variations of RO refractivity, temperature, and specific humidity retrievals below 5 km strongly indicate the impact of the increasing availability and variation of water vapor in the Arctic summer on the RO retrievals.

Overall, JPL and UCAR refractivity retrievals are consistent with temperature and humidity retrievals (e.g., negative refractivity bias corresponding to positive bias of temperature and negative bias of the specific humidity), except at certain altitudes, where the JPL and radiosonde comparison shows positive fractional refractivity bias corresponding to positive temperature bias (e.g., 6–8 km), which may be affected by pressure. In the JPL retrieval, the refractivity errors are directly mapped into both temperature and humidity errors. In the UCAR 1-D-variational retrieval, the refractivity errors can be mapped to both/either temperature and/or humidity errors. In addition, the errors in the a priori information from ECMWF model analysis used for both JPL and UCAR will also affect the RO temperature and humidity retrieval. It is worth noting that the errors of the geophysical parameters derived from the 1-D-variational method in the UCAR retrieval become more challenging to interpret as they include errors from both the model background and the RO measurement. For instance, the errors of UCAR specific humidity are very similar to the ERA-I but not consistent with the refractivity errors, which indicate the model a priori humidity might dominate the 1-D-variational UCAR humidity retrieval.

It is important to point out that the sharp drop in the number of near-coincidence profiles below ∼ 1 km (Fig. 5) is primarily due to the limited number of RO soundings penetrating deep into the PBL, especially the bottom ∼ 300 m above the surface.

To further quantify the parametric uncertainty, the COSMIC RO soundings from JPL and UCAR are also compared with the near-coincident ERA-I reanalysis profiles. Since the ERA-I assimilated COSMIC RO bending angles retrieved by the UCAR data center, they are not fully independent datasets. However, in the data assimilation, large RO measurement errors at lower altitudes (e.g., 20 % in bending angles errors near the surface) are normally applied, along with a limited number of available RO soundings; the impact of RO sounding on ERA-I in the lower troposphere remains limited (Poli et al., 2010).

The total numbers of common COSMIC RO soundings from both JPL and UCAR are 4782 pairs in winter and 8375 pairs in summer of 2008 over the Arctic (65–90∘ N). The comparison between COSMIC RO from the two centers and their near-coincident ERA-I profiles are presented in terms of fractional refractivity (Fig. 6), temperature (Fig. 7), and specific humidity (Fig. 8) differences. All profiles are interpolated into 100 m vertical intervals before the difference statistical calculation. Again, the statistics for the lowest 200m were discarded due to the significant drop of the number of RO observations near the surface. The statistical differences between COSMIC RO and ERA-I at three selected altitude ranges (e.g., 3, 5, and 10 km from the surface) are detailed in Table 4.

In Fig. 6, the fractional refractivity difference between RO (JPL and UCAR) profiles and the near-coincident ERA-I profiles shows small biases in winter. In the lowest 3 km, JPL refractivity exhibits a small positive bias of ∼ 0.12 %, whereas UCAR has a smaller positive bias of ∼ 0.07 %. In summer, an increase in positive bias above ∼ 4 km and negative bias below ∼ 2 km are seen. The maximum negative bias reaches about -0.7 % (-0.6 %) in the JPL (UCAR) retrieval near 0.5 km. An average negative refractivity bias of about -0.3 % below 3 km is seen in both centers.

Similarly, Fig. 7 shows the temperature difference between COSMIC RO and the near-coincident ERA-I profiles. Below 3 km, in winter, JPL shows a positive bias of 0.14 K, whereas UCAR has a negative bias of -0.12 K. In summer, UCAR has a warm bias of 0.58 K, whereas JPL does not have retrievals below 3 km. Below 10 km, temperature difference in winter (Fig. 7a, c) exhibits much smaller variation than that in summer. For instance, relatively large positive biases are seen in temperature retrievals from UCAR below 2 km (Fig. 7d) and from JPL within 6–10 km (Fig. 7b) as compared with the ERA-I profiles.

Moreover, Fig. 8 shows the difference between the specific humidity retrievals from COSMIC RO and the near-coincident ERA-I. Within 3 km, JPL and UCAR have a negligible bias of -0.002 and -0.02 g kg-1, respectively, in winter. In summer, much larger dry biases of about -0.12 in JPL and -0.08 g kg-1 in UCAR retrievals are detected. Overall, the RO specific humidity exhibits negative biases below 2 km, likely due to the abundant water vapor near the ocean surface.

In summary, the COSMIC RO (JPL or UCAR) refractivity difference from the near-coincident ERA-I retrievals is consistent with the temperature and humidity retrieval differences (e.g., negative refractivity bias corresponding to positive bias in temperature and negative bias in the specific humidity). However, positive fractional refractivity bias corresponds to positive temperature bias in JPL retrievals. Most of the mean biases have negligible standard errors due to large sample sizes.

To further study the impact of moisture on the RO soundings, the L1 signal-to-noise ratio (SNR) and excess Doppler for two typical JPL COSMIC RO soundings from summer (a, b) and one from winter (c) were presented along with the near-coincident ASCOS radiosonde (in summer) and the ERA-I profiles (in winter; Fig. 9). The two summer cases show nearly double the moisture with specific humidity (∼ 2 g kg-1) near the surface than that in the winter (less than ∼ 1 g kg-1). In the winter case, very smooth excess Doppler along with a relatively quiet SNR is shown. A sharp drop in SNR to the noise background is clearly seen around 39 s (Fig. 9c1), when tangent point descends to the smooth Arctic Ocean surface. In contrast, much larger variations in both the SNR and excess Doppler are seen in the two summer cases, especially below 5 km. The transition of the SNR to noise background near the surface is smeared due to more lower-troposphere moisture variations, which likely introduce multipath and SNR variations. Even though the lower-troposphere moisture in Arctic summer is still rather low (less than 2 g kg-1 near the surface) as compared to the low latitudes, a surprisingly large difference in refractivity (-7 % near surface) is seen (Fig. 9a2). The systematic negative RO N bias (-1 %) in summer season (e.g., Figs. 5 and 6) could be directly attributed to the lower-troposphere moisture variations. However, the impact of moisture and its variations on the RO signal dynamics and so the RO calibration and retrieval processes warrant a more comprehensive investigation.