The exploitation of Global Navigation Satellite System (GNSS) observations for monitoring the troposphere in support of meteorology has been proposed by [1
]. The initial parameter of interest was the Zenith Total Delay (ZTD), which represents a GNSS signal path delay due to transecting the neutral atmosphere in a vertical column above the station. The ZTD depends mainly on atmospheric pressure and partial water vapor pressure [2
]. First computation of ground-based GNSS ZTDs in near-real time (NRT) were demonstrated in 2000 [3
], i.e., shortly after establishing GNSS hourly data flow in 1999 and after providing precise orbits in ultra-rapid mode by the International GNSS Service (IGS, http://www.igs.org
) in 2000 [5
]. The COST Action 716 [6
] then played an important role in developing and evaluating methods of GNSS NRT troposphere monitoring and defining the standard format (COST-716) for the product dissemination. The operational production of GNSS ZTDs was organized within the COST-716 Demonstration Project [7
], and it has never been closed because its coordination was handed over to the newly established EUMETNET EIG GNSS Water Vapour Programme (E-GVAP, http://egvap.dmi.dk
) in 2004. Operational assimilations of ZTDs from E-GVAP into mesoscale Numerical Weather Models (NWM) for weather forecasting have been performed by Météo-France [8
] and UK Met Office [10
] since 2007. Both meteorological agencies started also an assimilation of the first E-GVAP global NRT ZTD product [11
]. Until present days, the standard GNSS ZTD monitoring in E-GVAP is performed in hourly update rate with a maximum product delay of 45 min with respect to the last GNSS processed observation. These requirements are a prerequisite of the ZTD assimilation performed usually with a 3–6 h time resolution.
Until now, a majority of E-GVAP analysis centers (ACs) uses a double-difference observation processing in a network solution. This strategy eliminates clock errors at GNSS receiver and satellite and was compulsory while public products were not available in NRT. The situation has changed in 2013, when the IGS introduced the Real-Time Service (RTS, http://rts.igs.org
) providing GPS and GLONASS orbit and clock corrections by combining contributions from several IGS real-time analysis centers [12
]. The IGS RTS aims at supporting real-time (RT) analyses with the Precise Point Positioning (PPP) method [13
]. The PPP is based on original observations or their linear combination without differencing between receivers or satellites. Though German Research Centre for Geosciences (GFZ) has provided a NRT PPP ZTD product [14
] since 2001, it was possible only thanks to their two-step processing approach consisting of (1) a global NRT solution for determining consistent satellite clock and orbit products and (2) a distributed PPP processing for ZTD estimated at each station individually. With the availability of global real-time data flow, software and standards specified for precise product dissemination, the PPP is becoming more popular for the troposphere monitoring. Compared to the traditional approach in E-GVAP dominated by the double-difference network processing, the PPP offers several advantages: (a) an easy production in real-time or NRT fashion, (b) flexible use of central or distributed processing scheme including a receiver built-in solution, (c) an estimation of tropospheric parameters in the absolute sense with a high spatio-temporal resolution, and (d) an optimal support of all satellite constellations and new signals including multiple frequencies; all profiting from a highly efficient and autonomous processing approach. The price for mentioned advantages is however paid by several disadvantages. Compared to the strategy using double differences, all observation models need to be carefully applied to reach the best accuracy. In addition, integer ambiguity resolution is possible only if precise observation phase biases are available, thus often non-integer-fixed ambiguities are usually estimated.
First ZTD products using the PPP method and new IGS RTS products were demonstrated in 2014 [16
]. Though the ZTD estimates already reached an internal precision below 10 mm when compared to GNSS final products, the studies also revealed significant station-/product-specific systematic errors attributed to inconsistencies in precise models applied in PPP and to the precise product generation. Fortunately, a large portion of the systematic errors are changing slowly over time and are thus not critical for an assimilation into NWMs which were designed to identify and remove biases on monthly basis by comparing station/product-specific GNSS ZTDs with their counterparts from the model background information [10
]. However, the precision of the real-time ZTDs obtained using the Kalman filter still remained worse by a factor of 1.5 compared to the E-GVAP standard NRT ZTD products. The reasons were twofold: (1) dependence on the quality of the RT products, and (2) use of the strategy for real-time data analysis. The E-GVAP solutions utilizing the batch processing and data files can be characterized with a standard deviation of 3–6 mm and 3–8 mm, for regional and global products, respectively, and a systematic error within 1–3 mm [19
]. The accuracy evaluated with respect to external data, such as radiosonde, corresponds to 1–2 mm in the precipitable water vapor [15
]. Anyway, real-time tropospheric products are ready to support assimilation in the rapid update cycle of NWM prediction or nowcasting, and target a short-term weather prediction or severe weather events monitoring [20
Nowadays, multi-GNSS offers many satellites and signals that are expected to strengthen all estimated parameters, in particular the ZTD and horizontal tropospheric gradients. For the E-GVAP production, data from the US NASTAR Global Positioning System (GPS) have been initially used. In 2008, the IGS started a provision of ultra-rapid precise orbit products for the Russian GLONASS satellites too. However, GLONASS data could have been included in NRT analyses only after resolving a 1.5 mm systematic difference in the ZTD when estimated independently from GPS and GLONASS data [21
]. The bias was caused by using the IGS05 model of phase center offsets (PCO) for GLONASS satellite antennas [22
], and the problem has been eliminated in 2012 [19
] by adopting consistent PCO models for both constellations in new IGS08 realization [23
]. Besides others, general limitations for the use of GNSS data from other global systems, European Galileo and Chinese BeiDou, persist mainly in (1) incompleteness of the constellations, (2) lack of precise models and calibrations for new signals, receiver and satellite instrumentations, and (3) lack of precise orbit and clock products supporting the ultra-fast processing mode. The situation will change soon as both global systems will become operational in next years. Since 2012, the IGS Multi-GNSS Experiment (MGEX, http://mgex.igs.org
) has been successfully filling the gaps in data, metadata, models, formats, standards and products for an optimal exploitation of all global satellite constellations and other regional augmentations. Although Galileo and BeiDou systems have not been completed yet, several groups reported an initial positive impact when using multi-constellation for estimating ZTDs and horizontal tropospheric gradients [24
The GNSS ZTD estimated from a single hour of data is not stable and accurate enough for the E-GVAP usage due to high correlations with coordinates and initial phase ambiguities. At least 12-h data batch is usually used for the NRT processing in E-GVAP, which means a high degree of redundancy in data processing updated on an hourly basis. In past, the issue was usually solved by reducing the data processing window to 1–6 h and by combining normal equations for the tropospheric parameter estimation [11
]. It was also reported, that ZTD values at the edges of the processing interval are about 20% less accurate compared to those in the middle of the interval and that the main impact is actually observed during the first and last hours of the interval [29
]. The E-GVAP expects estimated ZTDs within the last (production) hour, thus a trade-off between the accuracy and the latency is important.
A piece-wise linear function for modeling the tropospheric parameters within each hour is used by a majority of E-GVAP contributors using the Bernese GNSS Software V52 [30
]. In addition, horizontal tropospheric gradients are usually not estimated in operational solutions because of two reasons. First, there is presently no operational assimilation of gradients into NWM. Second, the estimation of high-resolution horizontal gradients reflecting a high spatio-temporal variability of local humidity increases the number of estimated parameters in the network and, consequently, the computation time by a factor of 2–3 at least. The PPP with an epoch-wise filtering supported by the IGS RTS products is an optimal strategy for generating advanced GNSS tropospheric products for future meteorological applications [20
] such as high-resolution ZTDs, horizontal tropospheric gradients, and slant tropospheric delays.
In this paper, we present a new processing strategy adaptable in the operational mode when prioritizing product latency (RT) or product accuracy (NRT). Additionally, all the advanced tropospheric parameters for RT and NRT products are derived within a single continuously operating RT PPP engine supported by the IGS RTS orbit and clock corrections. We believe that such unified processing system will replace soon our existing NRT contribution to E-GVAP. The strategy exploits all abovementioned advantages including benefits of full multi-constellation in near future.
The paper is organized as follows: after the introduction we assess real-time precise orbit and clock products available from the IGS RTS, evaluate an operational prototype of GNSS ZTD real-time production, both as pre-requisites for the new strategy. In Section 3
, we introduce the new strategy for an adaptable PPP processing solution for estimating ZTDs, horizontal tropospheric gradients and tropospheric slant delays using precise real-time products. In Section 4
and Section 5
, we assess the results of estimated tropospheric parameters driven by the backward data smoothing including study of impacts of IGS real-time products and multiple constellations. In the last section, we close the paper with conclusions.
4. Assessment of New Method Compared to the Existing E-GVAP Processing
The new strategy has been initially developed and assessed using the GNSS4SWEC Benchmark dataset [31
] and firstly compared to the GOP NRT tropospheric solution contributing operationally to E-GVAP. Though the new strategy can provide RT and NRT products in high temporal resolution, we compared only the ZTD as a product of HH:00 and HH:59 time stamps in every hour representing the standard NRT E-GVAP product, see Figure 7
summarizes results of three strategies and six ZTD solutions using 13 EUREF stations selected from the benchmark campaign and exploiting the EUREF combined ZTD product as a reference for all comparisons [39
]. The table show summary statistics indicating similar improvements in terms of the standard deviation over all ZTDs estimated at HH:00 in NRT independently of applied products (IGS RTS vs. IGS final products), processing strategies and software (G-Nut/Tefnut PPP vs. Bernese V52 DD). Compared to the Kalman filter ZTD estimated at the last epoch (HH:59), the backward smoothing running on hourly basis showed the improvement of 20% and 24% for the IGS RTS and the IGS final product, respectively. The E-GVAP/GOP product demonstrates a similar improvement (24%) comparing ZTD from HH:00 against HH:59, which corresponds to our previous results [29
]. The new adaptable PPP solution using the IGS final orbits reached the same accuracy as the E-GVAP/GOP product using the IGS ultra-rapid orbits and NRT DD network solution from the Bernese GNSS Software. On the other hand, the use of IGS RTS products instead of IGS final products in PPP indicates a degradation of 18% in ZTD SDEV and a 2.5 mm bias, which agrees with the summary in Section 2.2
and Section 2.3
. It should be finally noted, that the E-GVAP/GOP solution and the reference EUREF solution are based on a similar processing strategy and the software, while the new strategy is significantly different.
shows standard deviations and biases individually for all stations comparing the first ZTDs (HH:00) and the last ZTDs (HH:59). The statistics of ZTDs from the E-GVAP/GOP solution (PPP:DD-ultra) are plotted in red and pink for HH:00 and HH:59, respectively. The results from the new strategy using the PPP with IGS final orbit and clock products (PPP:IGS_final) are shown in dark and light blue and, using IGS RTS (PPP:IGS03) in black and grey. Standard deviations for all the stations show a similar improvement in the ZTD SDEV over all the strategies, software and precise products. However, there is no significant impact of the strategy on systematic errors, and we can observe only a common positive bias attributed to the use of the IGS RTS products.
We described a new strategy for generating advanced RT and NRT tropospheric products exploiting the PPP method and, simultaneously, the Kalman filter and the backward smoothing supported with the IGS real-time products. The strategy can be used to provide RT and NRT products synchronously, while it can be further optimized either for the latency or the accuracy of the NRT product. Although the NRT solution is generated as a side product of the real-time analysis, in terms of the precision, it is comparable to the traditional NRT ZTDs using the LSQ adjustment and batches of double-difference observations in a network mode, still commonly used within the E-GVAP. In addition, both the RT and NRT products can produce a consistent set of all parameters, namely the ZTD, horizontal tropospheric gradients and slant delays, all provided in high resolution and using optimally all observations from multi-constellations if precise products are available.
A long-term assessment of the IGS RTS in terms of the quality demonstrated that the products were available over 90% during the period 2013–2017 with 2–3 longer gaps only, and reached the 3D orbit accuracy of 4–6 cm and the clock precision of 2–4 cm. All the products were stable and were usable for the purpose of the troposphere monitoring, though some temporal variability was observed in the quality. An assessment of 1-year ZTDs generated routinely in the real-time PPP solution reached a stable precision of 6–8 mm over 18 global stations with about 10% improvements when additionally including GLONASS observations.
A simulated analysis of the impact of the backward smoothing on the ZTD and horizontal tropospheric gradients in NRT were evaluated within the GNSS4SWEC Benchmark campaign. Improvements of 20% for NRT ZTDs compared to RT ZTDs were demonstrated using the new strategy, i.e., the same improvement as observed within the piece-wise linear NRT ZTD estimation using the LSQ data processing. The overall improvements of the backward smoothing algorithm reached over 20%, independently if the final or real-time precise products are used. The impact of RT precise products on ZTD estimates indicates a systematic error of 2.4–2.8 mm and a degradation of 13–17% when compared with the use of final products, however, up to a factor of 2 worse in the accuracy when comparing parameters from collocated stations.
We further evaluated differences of ZTDs and horizontal tropospheric gradients from two collocated GNSS stations. We observed a significant impact of the backward smoothing on estimated tropospheric parameters when applied with 30-min latency in near-real time. A similar effect was observed for single- and multi-constellation solutions. An improvement of about 25% was then achieved when using multi-constellation compared to the standalone GPS, though observations from GLONASS and Galileo systems were down-weighted and their precise products and models are less accurate. Models were verified by visualizing carrier-phase post-fit residuals at individual stations. It showed that the backward smoothing improves mainly adjusted parameters, but does not affect essentially the distribution of post-fit residuals, however, retrieved slant delays still benefit from improved estimated parameters.