In general, since voxels located on the side of the tomographic area do not travel by any ray, the water vapor density of these voxels can be estimated based on only constraint equations, which reduces the accuracy of the tomography solutions. The primary contribution of the side signals is to ensure that voxels not passed in the traditional model can be crossed, which allows the accuracy and stability of tomographic results to be improved. To test the HFM and evaluate the efficiency of side signals, four special experimental schemes were performed separately.
Section 4.1 describes the implementation of the four experimental schemes; the contribution of the side signals to the tomography solutions is analyzed in
Section 4.2. On the other hand, the existing improved model combining the side observations was reestablished using the GNSS data [
14], and is compared to the proposed HFM in
Section 4.3.
It should be noted that when assessing the accuracy of tomography solutions, the vertical profiles obtained from radiosonde were used as reference values. Although the radiosonde data has been made use of to calculate the isotropic height factor, its main contribution is to provide the vertical distribution characteristics of atmospheric water vapor, not to participate in the tomographic processing. For instance, the tomographic time was set to August 2017, according to the HFM, and the optimal tomography top boundary was set to 11 km. Additionally, the height factor model was applied to calculate the isotropic height factor. Consequently, only GNSS observations went sent to the tomography system as input information for reconstructing the three-dimensional water vapor field, which shows that the proposed HFM does not interfere with the accuracy assessment of tomography results.
4.1. Experimental Schemes
The GNSS data with the sampling rate of 30 s from 17 stations (red dot in
Figure 7) provided by the Hong Kong Satellite Positioning Reference Station Network (SatRef) were processed using GAMIT/GLOBK (v.10.7). In the processing, the four International GNSS Service (IGS) stations (BJFS station, HKWS station, LHAZ station, and SHAO station) were introduced. The ZTD and the gradient delay were estimated with the time resolution of 5 min, which was sufficient to reveal the variation characteristics of water vapor. The Saastamoinen model was set as the expression for the ZHD [
45], and Vienna Mapping Function 1 (VMF1) was used for calculating the SWDs [
46]. In addition, the post-fit residuals calculated by GAMIT/GLOBK exceeding 2.5 times the standard deviation were removed [
47]. The 30-year radiosonde observation data was obtained from the King’s Park Meteorological Station (HKKP, the blue dot in
Figure 7a). The research area of the tomography model spanned from 113.82°E to 114.36°E and 22.16°N to 22.56°N. Moreover, the horizontal resolution was 0.09° in longitude and 0.08° in latitude, as shown in
Figure 7a. In the current tomography studies based on the Hong Kong SatRef, about a month of GNSS data and the radiosonde observations derived from the HKKP station were processed for modeling the tomography area [
14,
20,
36,
41]. Therefore, the experimentation time in this work was set as August 2017, i.e., day of year (DOY) 213–243, when Hong Kong was in the summer and there were more rainstorms. The tomography region was segmented into 15 nonuniform layers from 0 to 11 km (August top boundary determined in
Figure 2) [
14].
In the most common voxel-based tomography model, to assess the accuracy of tomographic results, water vapor profiles of the voxel columns (small blue rectangles in
Figure 7b) where the radiosonde is located are compared with radiosonde-derived vertical profiles. However, the empty voxels are mainly distributed on the side of the three-dimensional tomographic model. Consequently, as shown in
Figure 8, four special test schemes based on four different tomography areas (red boxes of
Figure 8a–d) were implemented. The RS-voxels (the blue voxels, shown in
Figure 7b, where the radiosonde is located) were placed on four boundaries of the tomographic region in the four schemes, for instance, RS-voxels on the east side (
Figure 8a), RS-voxels on the west side (
Figure 8b), RS-voxels on the south side (
Figure 8c), and RS-voxels on the north side (
Figure 8d).
Table 2 shows the accurate range of the tomographic area and the number of voxels in these four schemes.
In addition, to verify the accuracy and reliability of the HFM, two sub-groups of solutions were generated for the four schemes; one only used the rays travelling from the top boundary, while the others combined both rays passing from the top and side boundaries. It should be noted that if the effective observations from the HKSC station are included, the data is not consistent with the signal distribution of the general tomographic model. In addition, due to the RS-voxels always being crossed by rays in the four schemes, the benefit of the side signal for improving the accuracy of the water vapor density of the un-punctured voxels cannot be evaluated accurately. Therefore, all observations from HKSC station were excluded in the four schemes.
4.2. Contribution Analysis of the Side Signals
The isotropic and anisotropic SWV values of inside signals were first estimated by the HFM, as listed in
Figure 9. A slight consistency between the total SWV and the isotropic one can be noticed from the four sub-figures. In contrast, the anisotropic parts were inconsistent with the totals and only accounted for an average of 2.31%, 2.56%, 2.61%, and 2.25% of the corresponding total component in the four schemes, which reveals that the isotropic portion has a more marked significance than the anisotropic portion in estimating the SWV value of the inside signals. However, the latter, with a value of approximately 20 mm for certain rays, is also important.
Furthermore, we developed a detailed analysis of the number of rays crossing from the RS-voxel columns.
Figure 10 demonstrates the benefit of absorbing inside signals into the traditional tomography model. Evidently, in the East, West, and North schemes, no signals passed from the voxel from layer 1 to layer 8 in the general tomography model, while only the first four layers of voxels were not penetrated by rays in the improved model. As far as the South scheme is concerned, it is noticeable that the RS-voxel columns were all passed by GNSS rays in the optimized model, except for the voxel on the first layer. The number of penetrated signals tripled for voxels from 2 to 11 layers when adding the inside rays.
The average number of effective rays and the mean signal utilization rate for the traditional model and optimized model during the tomographic period are also compared in
Figure 11, where the histogram shows the number of signals and the line chart represents the signal utilization rate. The former refers to the 31-day average number of GNSS rays selected in the tomography model during the two epochs. Similarly, the latter represents the 31-day mean utilization rate of the number of rays used in the equations out of the total number of GNSS signals in each tomography epoch. Because the GNSS rays with an elevation angle below 15° were eliminated, the mean utilization rate of integral signals and inside ones was about 80% in the four schemes. With side signals absorbed into the tomography system, the mean effective rays increased by 32.33%, whereas the average utilization rate of GNSS signals was enhanced by 33.42%, from 47.12% to 80.54%. In particular, the number of effective signals combining inside ones remained an approximative constant for any tomographic scheme, which enhanced the stability of the improved tomography model.
The three-dimensional WVD was reconstructed using the traditional method and the proposed approach, respectively. To assess the retrieval capabilities of the optimized tomographic model, the SWV differences between the observed SWV and reconstructed SWV from the three-dimensional tomographic water vapor field were calculated, and the SWV differences between the traditional method and the proposed approach were obtained, respectively.
Figure 12 illustrates a comparison of the SWV differences between the two methods in the four schemes.
There is a noticeably similar tendency, in that the SWV differences of GNSS signals with an elevation angle larger than 45° are in the range of −5 to 5 mm, while those of rays with an elevation angle lower than 45° sharply increase with a decreasing elevation angle in all four schemes. As far as the comparison of SWV differences between the general and optimal tomography model is concerned, a remarkable improvement can be identified in the East scheme, West scheme, and South scheme, whereas the proposed method is slightly better than the traditional one in the North scheme. The same conclusion can also be deduced from
Table 3, where the statistics, including the RMSE, standard deviation (STD), and bias, for the SWV differences from the four schemes are compared. The improvements of the combined observations with respect to only the integral one are 3.11/3.34, 2.55/2.19, 2.60/2.62, and 1.27/1.35 mm for the RMSE/STD in the four schemes, respectively. In the bias comparison, there is a visible advancement of the tomography solutions when inside signals are included.
In addition, the tomographic results derived from the general model and improved model were compared with radiosonde data at 00:00 UTC and 12:00 UTC, when the radiosonde data provided accurate water vapor profiles at different altitudes.
Table 4 lists the mean RMSE and STD of the tomography results retrieved from the integral signals and combined signals for August 2017. It can be observed that the tomography results are closer to the radiosonde data when the inside rays are absorbed. The improved tomography model displays a smaller RMSE than the conventional one in the four test schemes, with decreases of 33.12%, 30.86%, 19.27%, and 30.67% in the mean RMSE, and the approximate consistent decrease rate of the STD for the four schemes can also be computed from
Table 4.
4.3. Comparison with the Existing Correction Model
In this section, the HFM in the current study is compared with the existing correction model in Yao and Zhao [
14], which is named the Yao model in this work for convenience. In the Yao model, a similar exponential relationship between the scale factor and the height is established, as follows [
14]:
where
is the scale factor, and
and
represent the coefficients of the scale factor, respectively.
denotes the height of the intersection point between the effective side signals and the side face of the tomography area.
To compare the performance of the HFM and Yao model in retrieving the three-dimensional water vapor field, three further schemes were designed for tomography modeling based on all GNSS station observations from the Hong Kong SatRef, in addition to the HKSC station. The tomographic region for the three schemes covers from 113.82°E to 114.36°E in longitude and 22.16°N to 22.56°N in latitude. Other tomographic model configurations, such as a horizontal and vertical resolution, are consistent with the four special experimental schemes. The three schemes in this section are as follows.
Scheme I: The traditional tomography method that only considers the GNSS signals crossing the top boundary is adopted to construct the observation equations;
Scheme II: The Yao model is used to estimate the SWV of side rays and build the equations with both rays passing from the top and side boundary;
Scheme III: The HFM is employed to calculate the and structure the tomography system with top and side observations.
Based on GNSS data at two tomography epochs of 00:00-00:30 UTC and 12:00-12:30 UTC during the period of DOY 213-243, 2017, the scale factor in the Yao model was calculated and is illustrated in
Figure 13, where a similar exponential relationship can be observed. It is noticeable that owing to the neglection of GNSS signals with an elevation angle of less than 15°, all of the heights of the intersection points are higher than approximately 1 km. Accordingly, the coefficients
and
were estimated by the least-squares method, and the results are given in
Table 5 [
27].
With the scale factor in the Yao model and the HFM proposed in this paper, the SWV of the side signals could be estimated based on the height of the intersection points. Consequently,
Figure 14 presents a comparison of the SWV value of effective side rays between the Yao model and HFM at two tomography epochs during each of the test days. It is clear that the HFM can estimate slightly higher SWV values than the Yao model for most signals, and the latter provides a rather small SWV estimation, with a value of approximately 20 mm, for several observation rays.
To assess the water vapor reconstruction capabilities of different methods, the tomographic results derived from the three schemes, described in
Figure 15, were employed to illustrate the water vapor vertical profile at 00:00 UTC and 12:00 UTC on DOY 220 (sunny day) and DOY 234 (rainy day).
It can be observed that the water vapor profiles derived from the three considered schemes agree with the reference profile obtained by the radiosonde data. Furthermore, a higher coincidence of the vertical profile can be seen on majority layers when combining the inside signals. As far as the quantitative comparison of WVD is concerned, the accuracy of the tomographic solutions from Scheme III (mean RMSE are 0.69 and 0.95 g/m3 on both days) is superior to those from Scheme I (mean RMSE are 1.40 and 1.30 g/m3 on both days) and Scheme II (mean RMSE are 0.73 and 1.20 g/m3 on both days), which reveals that the proposed model has better reconstruction capabilities than the Yao model and traditional model in both weather conditions.
In addition, to further show the superiority of the proposed method in the WVD comparison, the RMSE of tomographic results from the three schemes was calculated, and the results are shown in
Figure 16, including the 31-day period from DOY 213 to DOY 243, 2017, at 00:00 and 12:00 UTC daily. It is clear that, among the three schemes, Scheme III has the smallest RMSE compared to Scheme I and Scheme II in the majority of time periods. The average improvement of Scheme III with respect to Scheme I and Scheme II was 0.51 and 0.20 g/m
3 in terms of the RMSE, respectively. As a further assessment,
Table 2 lists the statistics of the three tomography solutions, including the maximum, minimum, and mean values of RMSE, STD, and bias.
It should be noted from
Table 6 that all of the means of the three statistics in Scheme II and III are superior to those of Scheme I, which suggests that there is an improvement of the GNSS tomography solution when side observations are included in the tomography modeling. As far as Scheme II and III are concerned, the average RMSE and STD for the former are 1.28 and 1.25 g/m
3 and the latter are 1.08 and 1.07 g/m
3, respectiely, and the mean accuracy improvement rates are 15.62% and 14.40% for both statistics, respectively, which demonstrates that the proposed HFM provides more accurate side SWV observations and three-dimensional water vapor fields.
Following this, to quantitatively compare the accuracy of each layer’s tomographic results, the reference value of WVD, along the vertical columns, was interpolated using the radiosonde data, and the mean difference between the tomographic results obtained from the three schemes and the radiosonde data during the tomographic period was calculated and is reported in
Figure 17. It is evident that the absolute value of the average differences dramatically decreases with increasing height layers. Besides, it should be noted that in the earth surface layer from 0 to 2 km, the difference of most voxels is less than 0 in the three schemes, i.e., the WVD retrieved from the GNSS tomography solution is lower than that obtained from the radiosonde. However, it is noticeable that there is an impressive improvement of the WVD in these layers in Scheme II and Scheme III, although the voxels in this range are not penetrated by side signals according to
Figure 10. This can be explained by the fact that when combining both integral signals and inside ones for GNSS tomography, the number of effective observations is obviously increased and the spatial geometric defect of the three-dimensional tomographic model and the ill condition of the tomographic observation equations are gradually remedied, which is beneficial for improving the accuracy and stability of the tomography results.
Furthermore, the conclusion that the differences of Scheme III are smaller than those of Scheme II in each layer can be drawn from
Figure 17, which highlights the advantages of the proposed HFM. This could be because of the higher side SWV observations estimated from the HFM in
Figure 14, which results in the tomography solutions being closer to the radiosonde measurements.