The Raman lidar method can assess the water vapor mixing ratio profiles through inelastic backscattering signals from nitrogen at 387 nm and from water vapor at 407 nm (Whiteman et al., 1992; Mattis et al., 2002; Adam et al., 2010). At the lowest height, the intersection of the laser beam with the receiver's field of view in the bistatic system is incomplete. Nevertheless, the overlap of both Raman channels is presumed to be equivalent; thus, the overlap effect could be minimal concerning water vapor measurements. The collected water vapor measurements, along with concurrent temperature profiles from a collocated MWR, allow us to obtain RH profiles. The vertical and temporal resolution of Raman lidar and other instruments are listed in Table 1.

The microwave radiometer (MWR) serves as a passive instrument designed to measure atmospheric emissions across two frequency bands within the microwave spectrum (Cimini et al., 2006; Crewell and Löhnert, 2007). There are seven channels set along the 22.235 GHz H2O absorption line. Humidity information can be extracted from these observations. The seven channels from 51–58 GHz within the O2 absorption complex encompass the vertical temperature profile data. Consequently, the fully automatic MWR enables the derivation of temperature and humidity profiles with a temporal resolution of up to 5 min. The method for inverting temperature and humidity profiles is the neural network method in this study. It uses statistical methods to optimize the long-term average radiosonde data and relies on previous radiosonde data (Yang et al., 2023).

We use radiosonde data from the China Meteorological Administration (CMA) station for reference analysis. The radiosonde is located in the same place as the Raman lidar and provides on-site measurements of atmospheric pressure, temperature, and RH. During the observing campaign, radiosondes were launched twice a day (08:00 and 20:00 Local Standard Time, LST). The height can be determined by the ascent time of the radiosonde balloon. The vertical resolution of the raw data is 3 mlayer-1. To match other data, the vertical resolution of the raw data is interpolated to 30 m (0–3000 m) and 250 m (3000–10 000 m).

In 2016 and 2021, China successfully deployed two second-generation geostationary meteorological satellites, Fengyun-4A (FY4A) and Fengyun-4B (FY4B). They are both equipped with the Geostationary Interferometric Infrared Sounder (GIIRS). The GIIRS therefore became the first geostationary orbiting meteorological satellite (Yang et al., 2023). This approach could achieve the detection of weather systems across China and its neighboring regions with high temporal and spatial resolution. Therefore, it enables a more comprehensive understanding of the atmospheric vertical structure, including the retrieval of atmospheric temperature profiles for 1000 m layers and moisture profiles for 2000 m layers (Yang et al., 2017). In comparison to FY4A, the GIIRS on FY4B exhibits a broader spectral range, improved spectral resolution in the long-wave IR band, and superior detection sensitivity. Specifically, the temporal resolution of GIIRS is enhanced from 2.5 h for FY4A to 2 h for FY4B. In the meantime, the spatial resolution has progressed from 16 000–12 000 m at nadir. The atmospheric humidity profiles utilized in this study, derived from GIIRS, are generated through the neural network algorithm created by the National Satellite Meteorological Centre (NSMC) (Bai et al., 2023). The data are available online: http://www.nsmc.org.cn/nsmc/cn/theme/FY4B.html (last access: 11 July 2025).

This study aims to obtain a continuous time series of RH profiles by integrating ground-based remote sensing techniques, including Raman lidar, MWR, and satellite data, in a straightforward manner to facilitate a wide range of applications. The retrieval process involves a systematic four-step algorithm that integrates the Raman lidar water mixing ratio profile and MWR brightness temperatures along with satellite data. The retrieval framework is shown in Fig. 1, and the retrieval process is detailed in the following paragraphs.

Step 1: data quality control. Data with quality control codes of 0 and 1 for FY4B and 0 for ground-based remote sensing data are selected. The lidar only retains data with a signal-to-noise ratio (SNR) greater than 3. The threshold value of the SNR is set as 3 based on our extensive comparisons with radiosonde data from CMA's long-term observations. The results indicate that selecting lidar signals with SNR>3 can significantly improve the consistency between retrieved RH profiles and radiosonde measurements. Therefore, in the data selection period, the Raman signal starts with the first SNR greater than 3 and ends with five consecutive SNRs less than 3. The real-time observing data are designated as Rradio, Rlidar, RMWR, and Rsatellite in Fig. 2.

Step 2: data spatial–temporal matching. This process aims to match the above quality-controlled data with the radiosonde data at a height of 0–10 000 m in time and space. For the time matching, temperature from MWR and water vapor data from Raman lidar are selected corresponding to the radiosonde data time (08:00 LST and 20:00 LST). In terms of spatial matching, the FY4B data are selected from the nearest grid point to the ground observing station for the horizontal scale. The data at vertical heights are interpolated to the resolution of 30 m (0–3000 m) and 250 m (3000–10 000 m).

Step 3: correction coefficient determination. The deviation between the temperature and humidity data of satellites and ground-based remote sensing data at each height is quantitatively calculated and analyzed to prepare for the optimal stitching process in the next step. Here, the deviations of lidar, MWR, and FY4B are designated as Dlidar, DMWR, and Dsatellite, respectively. 1Dlidar=Rlidar-Rradio2DMWR=RMWR-Rradio3Dsatellite=Rsatellite-Rradio The correction coefficients Clidar, CMWR, and Csatellite are calculated as follows. 4Clidar=|Dsatellite|+|DMWR|2⋅|Dsatellite|+|DMWR|+|Dlidar|5CMWR=|Dsatellite|+|Dlidar|2⋅|Dsatellite|+|DMWR|+|Dlidar|6Csatellite=|DMWR|+|Dlidar|2⋅|Dsatellite|+|DMWR|+|Dlidar| Step 4: synergetic algorithm iteration and evaluation. Based on the above spatial–temporal data matching and correction coefficient calculation at different heights, a dynamic optimal stitching algorithm (Fig. 2) is conducted. To ensure the independence between the tested sample and the true value, the temperature and humidity profiles of the current time are fused using the correction coefficient of the previous time and then compared with the radiosonde data at the same time for evaluation. The correlation coefficient (R), the root-mean-square error (RMSE), and the mean bias (MB) are used as inspection indexes. Finally, the retrieved RH information SRH could be obtained through the following equation. 7SRH=Rsatellite⋅Csatellite+RMWR⋅CMWR+Rlidar⋅Clidar From this process, we can see that, compared to these existing techniques, our new method not only incorporates satellite data but also dynamically determines optimal fusion coefficients. Because the fusion coefficients are dynamically determined by comparing the deviations from other measurements with the reference of radiosonde, they highlight that this new algorithm is real-time calibrated and that it can guarantee the device model independence and geographical adaptability. Thus, it eliminates constraints imposed by equipment specifications or observation locations, ensuring broad applicability across diverse scenarios.

To evaluate the performance of the synergetic algorithm for RH profiles, a comparative analysis was conducted between retrieved values and actual radiosonde measurements. Let N represent the total number of samples. The measured value is designated as Oi, with i representing the sample label. The value obtained through the new synergetic algorithm is designated as Gi. The evaluation indicators consist of MB, mean absolute bias (MAB), and RMSE, defined by the following equations. 8MB=∑i=1N(Gi-Oi)N9MAB=∑i=1N|Gi-Oi|N10RMSE=∑i=1N(Gi-Oi)2N

A 5-month data set has been chosen for a statistical analysis of RH profiles. The observation period spans 1 July–30 November 2024. The observation elements are RH data from 47 stations in China (yellow circles in Fig. 3) at the height of 0–10 000 m. To investigate RH retrieval accuracy, we provide the comparison results of four methods (lidar, MWR, satellite, and synergetic algorithm) utilizing the radiosonde data as the reference at 47 sites in Table 2. Then, Huhehaote (HHHT; northern China), Yibin (YB; middle China), and Qingyuan (QY; southern China) are selected as 3 representative sites (red stars in Fig. 3) for more detailed analysis, as shown in Fig. 4 and Table 3. The data samples for HHHT, YB, and QY are 3773, 7452, and 8110, respectively, after quality control and elimination of the precipitation data.

Generally, the synergetic algorithm at 47 sites presents the maximum correlation coefficient R value of 0.98 with the minimum RMSE of 5.27 % in Table 2. For three representative sites, the regression line from the synergetic algorithm at all heights similarly provides the best-fitting results, with the largest correlation coefficients R of 0.94, 0.92, and 0.93 in HHHT, YB, and QY, respectively (Table 3). The correlation coefficient R for lidar measurement follows with marginally higher values of 0.83 in HHHT, 0.86 in YB, and 0.86 in QY. It indicates its greater applicability compared to other single instruments. MWR presents the lowest R of 0.74 and 0.80 in HHHT and YB, performing better (R=0.75) than satellite (R=0.66) in QY. In terms of RMSE, the lidar-, MWR-, and satellite-derived RH all show values larger than 18 % at three sites. The synergistic use of a multi-source algorithm decreases the RMSE to the lowest value of 10 % in HHHT.

The regression line for lidar and MWR in HHHT, as illustrated in Fig. 4, exhibits a slope that is less than the one-to-one line. This implies that greater variations arise with increased RH in HHHT. Though the synergetic algorithm also presents similar trends, its RMSE decreased to 10 % in HHHT. The regression lines of lidar and MWR in YB and QY are larger than the one-to-one line, indicating the larger bias for less humidity.

As RH vertical profiles are height-dependent, Fig. 5 presents the MB profiles observed at different heights in terms of four methods. Generally, the MB in the RH of lidar in the lower troposphere (below 3000 m) outperforms the other two single methods (MWR and satellite) at three sites. No significant biases between radiosonde and lidar are noticeable. Specifically, the lowest MB values (4.93 % in HHHT, 2.63 % in YB, and 2.40 % in QY) in the comprehensive region of the tropospheric region are achieved when lidar data are incorporated into the synergetic algorithm. This is because lidar is an active remote sensing technology with more accuracy compared to MWR and satellite. The lidar data's efficacy is enhanced at heights below 3000 m when integrated with data from other sources within the boundary layer.

However, the MB from lidar increased drastically above this height, up to the highest value of 28.67 % in HHHT, 29.91 % in YB, and 20.09 % in QY. It is reasonable that the atmosphere changes so fast that radiosondes do not assess exactly the same air mass as lidar. In the meantime, lidar is increasingly constrained at elevated heights because of a decreased SNR; hence lidar is more trustworthy in the lower layer, i.e., below 3000 m.

In contrast, the MB from satellite (FY4B) over 3000 m varied steadily within the range of approximately 15 % at three sites. Therefore, the satellite data in the far height range would be more reliable and could be employed in the synergetic algorithm at higher layers. Compared to lidar and satellite, the MB from MWR gives the largest uncertainty in HHHT at all heights. This may result from the discrepancy between the temperature recorded by the radiosonde and that obtained from the MWR in HHHT. However, it yields relatively less variation than lidar and satellite in YB and QY. In any case, the synergetic method gives the best result over three observing sites at almost all heights, and accurate measurements of RH vertical profiles provided here are highly beneficial for analyzing the hygroscopic growth of local aerosols.

The sources of the discrepancy can stem from several aspects. Firstly, although all instruments are collocated in the ground, radiosondes deviate at higher heights. Signals can be disrupted if clouds are present. Secondly, satellites provide gridded data, requiring the selection of ground observation points closest to the grid's latitude and longitude, which introduces uncertainty. Finally, both MWR and satellite are passive remote sensing technologies, which are inherently less precise than active remote sensing technologies. Besides the inherent hardware difference, the errors during the retrieval process (e.g., neural networks for MWR) are also unavoidable.

RH mean monthly vertical profiles have been derived from the synergistic method illustrated in Fig. 6 because RH profiles are retrieved from water ratio profiles and temperature profiles. For this property, the RH seasonal behavior may be more complicated. For example, no obvious seasonal behavior of RH profiles is found in HHHT or YB. However, QY still presents the most likely seasonal characteristic at most of the heights, with the highest mean values in summer at 1000–2000 m (80.65 % in July) and the lowest values at 7000–10 000 m in late autumn (20.50 % in November) in Fig. 6e and f. The elevated RH observed in QY's summer may be related to the sufficient water vapor and large transport volume, as QY is located in coastal areas. So the characteristic of QY would be more dependent on water vapor.

For comparison, HHHT and YB are relatively random. At over 3000 m in HHHT (Fig. 6a and b), RH in August shows predominantly high values, with the highest value of 65.37 % at 5000–7000 m. Unlike in HHHT and QY, the RH profiles of YB in November interestingly show the highest values (83.95 %) in the lower atmosphere (0–1000 m) in Fig. 6c and d. This suggests that the reduced temperatures observed in autumn of YB promote proximity to saturation conditions, resulting in elevated RH values in November. It is also worth noting that RH above 3000 m in November of YB decreases dramatically as height increases, with the minimum RH of 13.91 % at 7000–10 000 m. This could be explained by more rapid fluctuations in the water vapor density and temperature in YB in the higher layer under the control of the subtropical monsoon climate zone. In any case, this plot illustrates a clear decrease in the RH values with heights at three sites.

Though there is no obvious RH uncertainty caused by regional differences, we found that QY exhibits the predominant seasonal feature throughout most heights. In contrast, no discernible seasonal characteristics in RH profiles are observed in HHHT or YB. Thus, we believe diverse atmospheric circulation patterns and geographical environments could result in regional variations in RH values.

We selected two different severe convective events in YB (one hailfall from 20:00–23:30 LST on 15 April and one heavy precipitation from 14:00 LST on 25 May to 08:00 LST on 26 May) for comparison in Fig. 7. At 23:00 LST on 15 April, a thunderstorm with strong winds and hail occurred. The RH profile retrieved by the synergetic algorithm showed that, before 22:00 LST, the RH was high (around 90 %) at 3000 m height, low (20 %–50 %) between 3500 m and 8000 m, and above 80 % between 8000 m and 9000 m (Fig. 7a). This indicates that, before the severe convection, the upper and lower layers were relatively moist, while the middle layer (3500–8000 m) was dry (red arrow in Fig. 7a). Such conditions favor the evaporation and cooling of ice particles descending from the upper atmosphere, leading to refreezing and hail formation.

In contrast, the RH profile from 25–26 May showed that the entire troposphere (0–10 000 m) presented high RH values (>70%) starting at 19:00 LST, which was conducive to heavy precipitation (Fig. 7b). According to ground station observations, YB recorded an hourly rainfall of 52 mm at 21:00 LST, along with gale-force winds of 23 ms-1 (9th grade). Most areas in YB experienced precipitation, with localized heavy thunderstorms. From the above two cases, we can see that the RH in the middle troposphere can be used to distinguish between hail and heavy precipitation during severe convective events.