Exploring Precipitable Water Vapor (PWV) Variability and Subregional Declines in Eastern China

Abstract

In recent years, China has experienced growing impacts from extreme weather events, emphasizing the importance of understanding regional atmospheric moisture dynamics, particularly Precipitable Water Vapor (PWV), to support sustainable environmental and urban planning. This study utilizes ten years (2013–2022) of Global Navigation Satellite System (GNSS) observations in typical cities in eastern China and proposes a comprehensive multiscale frequency-domain analysis framework that integrates the Fourier transform, Bayesian spectral estimation, and wavelet decomposition to extract the dominant PWV periodicities. Time-series analysis reveals an overall increasing trend in PWV across most regions, with notably declining trends in Beijing, Wuhan, and southern Taiwan, primarily attributed to groundwater depletion, rapid urban expansion, and ENSO-related anomalies, respectively. Frequency-domain results indicate distinct latitudinal and coastal–inland differences in the PWV periodicities. Inland stations (Beijing, Changchun, and Wuhan) display annual signals alongside weaker semi-annual components, while coastal stations (Shanghai, Kinmen County, Hong Kong, and Taiwan) mainly exhibit annual cycles. High-latitude stations show stronger seasonal and monthly fluctuations, mid-latitude stations present moderate-scale changes, and low-latitude regions display more diverse medium- and short-term fluctuations. In the short-term frequency domain, GNSS stations in most regions demonstrate significant PWV periodic variations over 0.5 days, 1 day, or both timescales, except for Changchun, where weak diurnal patterns are attributed to local topography and reduced solar radiation. Furthermore, ERA5-derived vertical temperature profiles are incorporated to reveal the thermodynamic mechanisms driving these variations, underscoring region-specific controls on surface evaporation and atmospheric moisture capacity. These findings offer novel insights into how human-induced environmental changes modulate the behavior of atmospheric water vapor.

1. Introduction

Water vapor plays a vital role in processes from local precipitation to global climate regulation [1,2]. Specifically, variations in the lower troposphere influence the frequency and severity of extreme weather events and are closely related to precipitation [3,4]. Precipitable Water Vapor (PWV) represents the total amount of atmospheric water vapor in the atmosphere [5] and effectively summarizes the distribution of water vapor throughout the entire atmospheric column. Therefore, PWV is fundamental to understanding and monitoring atmospheric moisture, offering valuable insights into sustainable climate management, water resource planning, and the assessment of weather- and climate-related risks.

PWV is usually retrieved using radiosondes and water vapor radiometers [6,7]. However, these strategies are constrained by their inadequacies. For instance, sounding stations, usually located 200–300 km apart, launch sounding balloons twice daily [8]. Consequently, radiosondes offer water vapor distribution data with limited horizontal and temporal resolutions. PWV can also be obtained from reanalysis products, which provide variables of global atmospheric circulation with excellent spatial integrity and continuity of data, using sophisticated data assimilation techniques [9,10]. However, the accuracy and quality of the PWV measurements are limited [11,12,13,14]. Since the seminal work of Bevis et al. [15], which introduced the concept of PWV retrieval using the data of Global Navigation Satellite Systems (GNSSs), the field of GNSS meteorology has undergone rapid advancement. The utilization of GNSS PWV data offers distinct advantages, including cost-effectiveness, enhanced precision, high temporal resolutions, and capacity to operate under diverse meteorological conditions [16,17].

In recent years, numerous researchers have studied precipitable water using time-series data. In the time domain, some researchers have utilized multiple data sources to retrieve PWV and found that most regions of the globe exhibit an increasing trend, while a few areas remain unchanged or show a decreasing trend [18,19,20,21]. Studies on the frequency-domain characteristics of atmospheric water vapor have highlighted distinct patterns across different climates and regions. Typical bimodal distributions of water vapor frequency have been found in monsoon regions, and lognormal distributions have been found in subtropical and temperate zones [22]. Some studies have extracted multiyear periods in the Greenland region using the Fast Fourier Transform (FFT) method based on GPS and ERA5 reanalysis PWV data [23]. Khutorova et al. [24] investigated the long-term oscillations in the time series of the surface partial pressure of water vapor for Russia’s European region and found that variations in timescales ranging from two to six years accounted for between 35 and 60 percent of the interannual variance. Jadala et al. [25] used an FFT and harmonic analysis to extract the amplitude and phase of diurnal, semi-diurnal, and 8 h oscillations at the tropical Hyderabad site.

Similarly, some researchers analyzed GNSS water vapor data from stations of the Crustal Movement Observation Network of China (CMONOC), and the results indicated the existence of annual and semi-annual periods and revealed notable regional disparities and seasonal fluctuations in water vapor frequency characteristics, further emphasizing the complexity of atmospheric moisture dynamics [26,27]. These findings further underscore the complexity of atmospheric moisture dynamics. However, most studies only briefly discuss the PWV trends and periodic behaviors, without a thorough investigation into the underlying causes. Moreover, the feature extraction methods employed for PWV analysis in these studies remain relatively limited, lacking integration for a comprehensive assessment.

Hu Huanyong’s foundational research historically revealed that eastern China holds an absolute advantage over the western regions in terms of both population and economics [28]. Moreover, eastern China encompasses a broader range of the country’s major climate zones, many of which have become focal areas for scientific investigation. In recent decades, this regional disparity has intensified, and the global increase in extreme precipitation events has posed greater risks to eastern China [29]. These factors underscore the urgent need for a comprehensive investigation of the PWV characteristics in this region.

In this study, the PWV time series over key regions of eastern China were derived from GNSS observations at stations of the International GNSS Service (IGS). Through frequency-domain analysis, we explored the multiscale temporal characteristics and dominant factors influencing PWV variability. These findings contribute to a deeper understanding of regional atmospheric water dynamics, supporting sustainable climate monitoring, environmental planning, and early warning systems for extreme weather events.

Furthermore, PWV contributes to achieving the 2030 Agenda goals. The long-term trend of PWV can be used as a proxy variable for changes in atmospheric humidity, complementing traditional temperature-based indicators and aiding in the prediction of urban heat island effects. Its cyclical characteristics directly reflect the frequency and intensity of extreme precipitation events, enabling early warning capabilities. These insights offer a new perspective for assessing urban hydrometeorological risks and can inform the optimization of urban drainage infrastructure and disaster-prevention strategies.

2. Research Data and Methods

2.1. Research Data

2.1.1. GNSS Data

The GNSS observation files used in this study were collected from 15 in situ stations of the IGS (Figure 1), an international organization founded in 1993 by the International Association of Geodesy (IAG), to provide services for GNSS applications. The data are stored in RINEX format, and all 15 stations are equipped with receivers that continuously track GNSS signals. The observational data can be downloaded from the IGS Data Center in Wuhan (http://www.igs.gnsswhu.cn/index.php, last access on 10 December 2023), and the study period extends from 1 January 2013 to 31 December 2022. The data used in this study have been quality-checked according to WMO standards, with non-compliant data excluded.

Table 1. Statistics of GNSS station areas, temporal coverage, and number of valid data points before and after exclusion.

Stations	Areas	Temporal Coverage	Data Points Before Exclusion	Data Points After Exclusion
BJFS	Beijing	2013-01-01–2022-12-31	83,926	83,829
BJNM	Beijing	2013-01-01–2020-07-08	53,865	53,772
CHAN	Changchun	2013-01-01–2022-12-31	83,544	82,914
CKSV	Southern Taiwan	2016-10-16–2022-12-31	50,900	50,900
HKSL	Hong Kong	2013-01-01–2022-12-31	85,189	85,189
HKWS	Hong Kong	2013-01-01–2022-12-31	85,280	85,280
JFNG	Wuhan	2014-03-26–2022-12-31	75,026	75,024
KMNM	Kinmen County	2016-10-16–2022-12-31	50,681	50,681
NCKU	Southern Taiwan	2016-04-05–2022-05-19	47,560	47,560
SHAO	Shanghai	2013-01-01–2021-07-08	44,489	44,479
TCMS	Northern Taiwan	2013-01-01–2021-04-24	48,595	48,595
TNML	Northern Taiwan	2013-01-01–2022-12-31	45,966	45,966
TWTF	Northern Taiwan	2013-01-01–2022-12-31	81,810	81,810
WUH2	Wuhan	2016-08-31–2022-12-31	50,693	50,690
WUHN	Wuhan	2013-01-04–2022-12-31	72,874	72,867

shows the locations of the 15 stations, temporal coverage, and number of valid points before and after exclusion in this study. The JFNG station has provided observations since 2014, whereas the CKSV, KMNM, NCKU, and WUH2 stations have provided observations since 2016.

2.1.2. Radiosonde PWV

Radiosondes are meteorological sounding instruments carried by balloons to measure atmospheric parameters, such as temperature, pressure, and humidity, at different altitudes. These measurements enable the accurate calculation of PWV using standard formulas [30]. Radiosonde PWV is commonly used as a reference value for the validation of satellite PWV products owing to its excellent accuracy [31,32]. In this study, radiosonde data from six sites, obtained from the University of Wyoming (http://weather.uwyo.edu/upperair/sounding.html, last access on 5 March 2024), were downloaded for the period 2013–2022 to evaluate the accuracy of the GNSS-derived PWV. These radiosonde sites are located within a 50 km radius of the GNSS stations. Figure 2 depicts the distribution of the six GNSS stations and neighboring radiosonde sites.

2.1.3. Temperature

GNSS stations lack meteorological measurements, while radiosonde stations provide such data but are sparse. Ground-based weather stations offer only surface-level observations. ERA5 is the fifth generation of the European Center for Medium-Range Weather Forecasts (ECMWF) atmospheric reanalysis dataset of global climate from January 1950 to the present, produced by the Copernicus Climate Change Service (C3S) [33]. In contrast, the ERA5 reanalysis data provide an adequate vertical temperature dataset across the study area. Therefore, the temperature dataset from ERA5 can be used to analyze vertical atmospheric temperature profiles. The temperatures used in this study are labeled as T_2m (thermodynamic temperature at a height of 1.5–2 m) and T_h (atmospheric temperature at 14 pressure levels from 1000 hPa to 600 hPa). The grid for latitude and longitude is 0.25° × 0.25°, and the time resolution is one hour. We then matched the latitude and longitude of the GNSS stations on a 0.25° grid to obtain atmospheric data from the GNSS stations.

2.2. GNSS PWV Retrieval

2.2.1. GNSS Data Solution Software

As is well known, there are numerous GNSS data processing software tools, among which GAMIT and Bernese are representative and widely used globally. GAMIT is designed for scientific use, supporting satellite orbit calculation, atmospheric delay correction, earth rotation parameter estimation, and other purposes. It provides millimeter-level accuracy and allows code customization. Bernese is another widely adopted tool that provides comprehensive functions for both precise point positioning (PPP) and network solutions [34]. Its automated processing capabilities and relatively user-friendly interface make it suitable for both research and engineering applications. However, its high cost may limit access to individual users.

In recent years, China has developed several high-precision GNSS processing tools. PRIDE and GREAT are two notable platforms that offer competitive performance in high-precision GNSS applications. PRIDE is an open-source software package with a user-friendly GUI for high-frequency GNSS data processing [35,36,37]. It offers centimeter-level accuracy and has strong dynamic tracking capabilities, making it particularly suitable for crustal movement monitoring, time–frequency transfer, and aerial photogrammetry. GREAT is a powerful tool for precise positioning, multi-source fusion navigation, and satellite precise orbit determination [38,39,40]. Its modular and open-source structure is suited to complex navigation scenarios and integrated Earth science research.

Considering the high precision, scientific applicability, and flexible customization of GAMIT 10.7, it was selected in this study for GNSS data processing. Its robust tropospheric delay modelling capabilities and millimeter-level accuracy make it well-suited for extracting reliable PWV time-series data in climate-related research. In this study, double-difference relative positioning strategies were adopted to process the GNSS observations within the research area. The processing model and detailed settings are presented in

Table 2. Data processing strategies in GAMIT.

	Processing Key Points	Specific Models and Parameters
Double-difference relative positioning	Coordinate parameters	Double-difference network adjustment
	Receiver clock offset	Differential technique
Observations	Type of observation	GPS observation data
	Observation periods	2013-01-01–2022-12-31
	Sampling interval	30 s
Correction models	Cutoff height angle	10°
	Phase entanglement	Correction
	Tidal load	Correction of Earth tides, polar tides, and ocean tides
	Relativistic effect	Correction
	ZTD prior model	VMF1+Saastamoinen
	Projection function	$\mathrm{VMF}1 (2°\times 2.5$°)
Parameter estimation strategies	Satellite orbit	IGS Final Orbital Precision Ephemeris
	Ambiguity solution	Least-Squares Ambiguity Decorrelation Adjustment
	Ambiguity parameter	Floating point solution

. The sampling rate of GNSS observations was 30 s. The final ephemeris of the IGS was used to estimate the satellite orbits. A cutoff angle of 10° was adopted to remove low-quality observations. The RELAX mode of GAMIT was employed, with data corrections applied using the VMF1 mapping function and tidal model [41]. Zenith Tropospheric Delay (ZTD) corrections were estimated hourly using the GAMIT default horizontal gradient.

2.2.2. GNSS PWV Retrieval Method

ZTD includes Zenith Hydrostatic Delay (ZHD) and Zenith Wet Delay (ZWD). ZHD can be accurately calculated using the Saastamoinen model [42]:

$$\mathrm{Z}\mathrm{H}\mathrm{D}=\left(2.2768\pm 0.0024\right){\mathrm{P}}_{\mathrm{s}}/(1-0.00266\times \cos 2\mathsf{\theta }+0.00028\times \mathrm{H})$$

(1)

where ${\mathrm{P}}_{\mathrm{s}}$ is the atmospheric pressure obtained from a ground-based weather station co-located at the GNSS station, $\mathsf{\theta }$ is the latitude of the GNSS station, and $\mathrm{H}$ is the GNSS station height.

ZWD is obtained by subtracting ZHD from ZTD, and PWV can be derived using the following formula [43]:

$$\mathrm{P}\mathrm{W}\mathrm{V}=\prod \times \mathrm{Z}\mathrm{W}\mathrm{D}$$

(2)

where the conversion coefficient, Π, can be calculated using the Bevis formula:

$$\prod ={10}^6/\left[\right({\mathrm{k}}^3\times {\mathrm{T}}_{\mathrm{m}}^{-1}+{\mathrm{k}}_2^{\prime })\times {\mathrm{R}}_{\mathrm{V}}]$$

(3)

where ${\mathrm{k}}_2^{\prime }$ (22.1 ± 2.2 K/hPa) and ${\mathrm{k}}^3$ (3.735 × 105 ± 0.012 × 105 K²/hPa) are the atmospheric refractivity constants; ${\mathrm{R}}_{\mathrm{V}}$ (4.613 × 106 erg/g.°) is the gas constant of water vapor; and ${\mathrm{T}}_{\mathrm{m}}$ is the weighted mean temperature, derived from the global barometric pressure and temperature (GPT) model of Boehm and Schuhti [44,45].

By plotting the GNSS PWV time series, we filtered and removed the PWV with values less than or equal to zero.

2.3. Bayesian Statistics

Bayesian statistics uses prior knowledge to probabilistically deal with model parameters and structures and provide a unified framework for resolving various forms of uncertainty [46]. It uses Bayesian model averaging (BMA) [47], which applies Bayesian integrated modelling techniques to combine numerous competing models and uncover complex nonlinear dynamics from time-series data. Although the model was originally developed for remote sensing [48,49], it can be applied to any time-series data that satisfy its assumptions [50]. In Bayesian statistics, time-series data are equal to the sum of trends, seasonality, and residuals:

$${\mathrm{Y}}_{\mathrm{t}}={\mathrm{T}}_{\mathrm{t}}+{\mathrm{S}}_{\mathrm{t}}+{\mathrm{e}}_{\mathrm{t}}, \mathrm{t}=1, \dots ,\mathrm{n}$$

(4)

where ${\mathrm{Y}}_{\mathrm{t}}$ ($\mathrm{t}$ denotes time) is the time series, ${\mathrm{T}}_{\mathrm{t}}$ is the trend, ${\mathrm{S}}_{\mathrm{t}}$ is the seasonality, ${\mathrm{e}}_{\mathrm{t}}$ is the random variation of the variable, and $\mathrm{n}$ is the number of observations in the time series.

The seasonal component, ${\mathrm{S}}_{\mathrm{t}}$, mentioned above is modelled as a piecewise harmonic model, whose specific form depends on the seasonal change points in the time series. For each time interval [${\mathsf{\xi }}_{\mathrm{k}}$, ${\mathsf{\xi }}_{\mathrm{k}+1}$], the seasonal component can be expressed as in Formula (5). The trend component, ${\mathrm{T}}_{\mathrm{t}}$, is modeled as a piecewise linear function, whose specific form depends on the trend change points in the time series. For each time interval [${\mathsf{\tau }}_{\mathrm{j}}$, ${\mathsf{\tau }}_{\mathrm{j}+1}$], the trend component can be expressed as in Formula (6):

$$\mathrm{S}\left(\mathrm{t}\right)=\sum _{\mathrm{l}=1}^{{\mathrm{L}}_{\mathrm{k}}}\left[{\mathrm{a}}_{\mathrm{k},\mathrm{l}} \ast \sin \left({\displaystyle \frac{2\mathsf{\pi }\mathrm{l}\mathrm{t}}{\mathrm{P}}}\right)+{\mathrm{b}}_{\mathrm{k},\mathrm{l}} \ast \cos \left({\displaystyle \frac{2\mathsf{\pi }\mathrm{l}\mathrm{t}}{\mathrm{P}}}\right)\right]; {\mathsf{\xi }}_{\mathrm{k}}\le \mathrm{t}\le {\mathsf{\xi }}_{\mathrm{k}+1}$$

(5)

$$\mathrm{T}\left(\mathrm{t}\right)={\mathrm{a}}_{\mathrm{j}}+{\mathrm{b}}_{\mathrm{j}}\mathrm{t}; {\mathsf{\tau }}_{\mathrm{j}}\le \mathrm{t}\le {\mathsf{\tau }}_{\mathrm{j}+1}$$

(6)

where $\mathrm{P}$ denotes the seasonal cycle (e.g., one year), ${\mathrm{L}}_{\mathrm{k}}$ is the harmonic order of the $\mathrm{k}$th time period, ${\mathrm{a}}_{\mathrm{k},\mathrm{l}}$ and ${\mathrm{b}}_{\mathrm{k},\mathrm{l}}$ are the coefficients of the $\mathrm{l}$th harmonic of the $\mathrm{k}$th time period, ${\mathsf{\xi }}_{\mathrm{k}}$ is the position of the seasonal change point, ${\mathrm{a}}_{\mathrm{j}}$ and ${\mathrm{b}}_{\mathrm{j}}$ denote the linear coefficients of the $\mathrm{j}$th time period, and ${\mathsf{\tau }}_{\mathrm{j}}$ is the position of the trend change point.

The trend (trend probability) is rigorously evaluated using a Bayesian approach with Markov Chain Monte Carlo sampling and model averaging, ensuring robust uncertainty quantification. This method provides a probabilistic understanding of trend changes and indirectly assesses trend probabilities through the statistical analysis of posterior distribution samples. The general form of the posterior distribution is as follows:

$$\mathrm{p}\left({\mathsf{\beta }}_{\mathrm{M}}, {\mathsf{\sigma }}^2, \mathrm{M}|\mathrm{D}\right)\propto \mathrm{p}\left(\mathrm{D}\right|{\mathsf{\beta }}_{\mathrm{M}}, {\mathsf{\sigma }}^2, \mathrm{M}\left)\mathsf{\pi }\right({\mathsf{\beta }}_{\mathrm{M}}, {\mathsf{\sigma }}^2, \mathrm{M})$$

(7)

where $\mathrm{p}$ denotes the probability density function; $\mathrm{D}$ is the observed data; ${\mathsf{\beta }}_{\mathrm{M}}$ and ${\mathsf{\sigma }}^2$ are the model parameters and noise variance, respectively; and $\mathrm{M}$ is the model structure (including the number and location of trend and seasonal change points).

In this study, we used Bayesian statistics to obtain the trend/seasonality components and trend probabilities of the PWV time series to analyze the time-domain characteristics of PWV.

2.4. Wavelet Transform

Time series are often nonlinear and complex, and different results can be obtained by analyzing data at different timescales. The wavelet transform can be used for time–frequency analysis because of its multi-resolution property and ability to characterize the local features of a signal in the time and frequency domains [51,52]. The wavelet transform method is a process of shifting a basic wavelet function by $\mathsf{\tau }$ and subsequently performing the inner product operation with the signal to be analyzed ($\mathrm{t}$) at different scales $\mathsf{\mu }$. It can be expressed as follows:

$${\mathrm{W}}_{\mathrm{f}}\left(\mathrm{a},\mathsf{\tau }\right)=<\mathrm{f}\left(\mathrm{t}\right), {\mathsf{\Psi }}_{\mathrm{a},\mathsf{\tau }}\left(\mathrm{t}\right)>={\displaystyle \frac{1}{\sqrt{\mathrm{a}}}}{\int }_{\mathrm{R}}{\mathsf{\Psi }}^{*}\left({\displaystyle \frac{\mathrm{t}-\mathsf{\tau }}{\mathrm{a}}}\right)\mathrm{d}\mathrm{t}$$

(8)

where $\mathrm{f}\left(\mathrm{t}\right)$ denotes the original signal, ${\mathrm{W}}_{\mathrm{f}}\left(\mathrm{a},\mathsf{\tau }\right)$ describes the degree of similarity between a signal and a wavelet function ${\mathsf{\Psi }}_{\mathrm{a},\mathsf{\tau }}\left(\mathrm{t}\right)$ at a specific scale $\mathrm{a}$ and translation $\mathsf{\tau }$ [53], and R is the set of real numbers. The duration of the wavelet at different scales broadens with increasing scale and decreases inversely with amplitude, $\sqrt{\mathrm{a}}$ [54], while the shape of the wave remains constant.

This study employed the complex Morlet (Cmor) wavelet [55] to identify the main modes of variability and their temporal variations, as well as to examine the amplitude and phase of the oscillations. Moreover, the amplitude of the wavelet spectrum reflects the temporal variations in the relative contribution of the periodic components at different scales to the original signals. At any given moment, it allows for the estimation of the intensity of parameter variations across all timescales. The phase wavelet spectrum, on the other hand, indicates the timing of the maximum value of each mode. Owing to these capabilities, wavelet analysis has been widely employed to identify multi-timescale variations and reveal hidden features in complex time series, particularly in the geosciences [24,56]. Cmor is expressed as follows:

$$\mathsf{\Psi }\left(\mathrm{t}\right)={\mathrm{e}}^{\mathrm{i}{\mathrm{w}}_0\mathrm{t}}{\mathrm{e}}^{-{\displaystyle \frac{{\mathrm{t}}^2}{{2\mathsf{\sigma }}^2}}}$$

(9)

where $\mathsf{\Psi }\left(\mathrm{t}\right)$ denotes the value of the complex Morlet wavelet function at time t; $\mathrm{i}$ denotes the imaginary number; ${\mathrm{w}}_0$ denotes the center frequency, which determines the principal frequency component of the wavelet function; and $\mathsf{\sigma }$ denotes the scale parameter, which controls the width of the wavelet function.

2.5. Fast Fourier Transform

The discrete Fourier transform (DFT) establishes a correspondence between the time-domain characteristics of the signal and the frequency-domain characteristics and realizes the mutual conversion of the signal in the time and frequency domains. The FFT, which was proposed by James W. Cooley in 1965, speeds up the computation process compared with the computationally intensive DFT [57]. The employment of the unit complex root as a rotational element not only adeptly captures energy fluctuations in multidimensional datasets, but also substantially diminishes the computational demands of the discrete Fourier transform. Furthermore, it elucidates the amplitude and phase attributes of continuous signals that remain imperceptible in the time domain in the frequency domain. The FFT is widely used in the frequency-domain analysis of various types of time series, and its core formula is as follows:

$$\mathrm{x}\left(\mathrm{k}\right)={\displaystyle \frac{1}{\mathrm{N}}}\sum _{\mathrm{j}=1}^{\mathrm{N}}\mathrm{X}\left(\mathrm{j}\right){\mathrm{W}}_{\mathrm{N}}^{(\mathrm{j}-1)(\mathrm{k}-1)},\mathrm{k}=1,2,\dots ,\mathrm{N}; \mathrm{j}=1,2,\dots ,\mathrm{N}$$

(10)

$${\mathrm{W}}_{\mathrm{N}}={\mathrm{e}}^{-\mathrm{j}{\displaystyle \frac{2\mathsf{\pi }}{\mathrm{N}}}}$$

(11)

where $\mathrm{x}\left(\mathrm{k}\right)$ denotes the distribution of the signal in the frequency domain, $\mathrm{X}\left(\mathrm{j}\right)$ denotes the distribution of the signal in the time domain, and ${\mathrm{W}}_{\mathrm{N}}$ is a rotation factor with periodicity and symmetry.

In this study, FFT was used to analyze the GNSS PWV time series in eastern China, and its oscillation characteristics were analyzed from a frequency-domain perspective.

2.6. Evaluation Indicators

Three evaluation metrics, namely, the correlation coefficient (R), Root Mean Square Error (RMSE), and mean bias (MB), were used to evaluate the precision of the GNSS PWV. The absolute values of the correlation coefficients are typically interpreted as very strong (0.8–1.0), strong (0.6–0.8), moderate (0.4–0.6), weak (0.2–0.4), or very weak or negligible (0.0–0.2) correlations. The RMSE measures the gap between the mean and true values of the test values. MB determines whether the test values are underestimated or overestimated. The formulas for the three evaluation metrics are as follows:

$$\mathrm{R}={\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{N}}(\mathrm{x}-\overline{\mathrm{x}})(\mathrm{y}-\overline{\mathrm{y}})}{\sqrt{\sum _{\mathrm{i}=1}^{\mathrm{N}}{(\mathrm{x}-\overline{\mathrm{x}})}^2}\sqrt{\sum _{\mathrm{i}=1}^{\mathrm{N}}{(\mathrm{y}-\overline{\mathrm{y}})}^2}}}$$

(12)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{\mathrm{N}}}\sum _{\mathrm{i}=1}^{\mathrm{N}}{(\mathrm{x}-\mathrm{y})}^2}$$

(13)

$$\mathrm{M}\mathrm{B}={\displaystyle \frac{1}{\mathrm{N}}}\sum _{\mathrm{i}=1}^{\mathrm{N}}(\mathrm{x}-\mathrm{y})$$

(14)

where $\mathrm{x}$ is the test value, $\mathrm{y}$ is the validation value, $\overline{\mathrm{x}}$ is the mean of the test value, $\overline{\mathrm{y}}$ is the mean of the validation value, and $\mathrm{N}$ is the sample size.

3. Results and Discussion

3.1. Accuracy Assessment of GNSS PWV

Before exploring GNSS PWV, we used observations from radiosonde stations near the GNSS stations to evaluate the accuracy of GNSS PWV. Observations at radiosonde sites were performed twice daily. Therefore, only the GNSS PWV values at the corresponding times were adopted to align with the radiosonde PWV values.

Comparisons of the radiosonde and GNSS PWV time series at the six stations are shown in Figure 3. As can be seen from the subfigures, the regression line for each station nearly coincides with the one–one line. The correlation coefficients at the six stations range from 0.95 to 0.99, the RMSE values vary between 2.6 and 4.2 mm, and the MB values range from −1.2 to 0.5 mm. The results indicate that the GNSS PWV retrieval in this study has high accuracy, which is consistent with previous studies and meets the application requirements [58,59,60,61].

3.2. Spatial and Temporal Variations in GNSS PWV

3.2.1. Overall Characteristics of PWV in the Time Domain

Ten years of PWV time series were obtained by solving GNSS observations of each station from 1 January 2013 to 31 December 2022. Due to a significant number of missing observations, PWV datasets from two neighboring stations, CKSV and NCKU, were merged into a single time series, as were those from TCMS and TNML. The datasets from the BJNM, SHAO, and WUH2 stations were not considered in this section because of their short time spans. By utilizing Bayesian statistics, we extracted the time series of seasonality and trend components, as well as the trend probability of PWV, from the monthly average PWV data. The extracted components and trend probabilities of the GNSS PWV for each station are shown in Figure 4.

As can be seen from the first subplots of all stations, the monthly average GNSS PWV values reach their peaks during the summer months and drop to their valley values in winter. This pattern is primarily driven by the monsoon climate of eastern China, where winters are cold and dry and summers are warm and humid. As shown in Figure 4a,b, the BJFS and CHAN stations exhibit the highest peak values of approximately 40 mm, with a short peak duration. In contrast, as illustrated in Figure 4c–j, the stations in Wuhan, Hong Kong, Taiwan, and Kinmen County display higher peak values of approximately 60 mm, accompanied by a longer peak duration. Moreover, the valley values also exhibit obvious differences, as they are approximately 0–5 mm in Beijing and Changchun, 10–20 mm in Wuhan, and 20–30 mm in the other areas.

The inter-station disparities exhibit a statistically significant latitudinal dependence. This latitudinal gradient aligns with thermodynamic constraints: reduced solar irradiance reduces surface evaporation efficiency, and colder winter temperatures at higher latitudes suppress atmospheric moisture retention. Shanghai and Wuhan are located at lower latitudes than Beijing and Changchun, resulting in more sunlight. Meanwhile, Kinmen County, Hong Kong, and Taiwan, located closer to the equator, experience higher annual temperatures, which enhance the surface evaporation efficiency and increase the atmospheric moisture retention capacity, ultimately contributing to the observed PWV patterns.

3.2.2. Trend Variations in PWV Time Series

From the second subplots of Figure 4b,e,g,h,j, it is evident that the red envelopes are more pronounced, indicating that the PWV in Changchun, Hong Kong, northern Taiwan, and Kinmen County is likely to follow an increasing trend. This phenomenon has also been reported in other studies [27,62,63], indicating that the increasing trend may be associated with the long-term rise in temperature and human activities in recent years [62,64]. In contrast, Figure 4a,c,d,i show that the blue envelopes over Beijing, Wuhan, and southern Taiwan are prominent, suggesting that the PWV in these regions is likely to decrease.

To further explore the factors contributing to the declining trend in specific regions, we conducted an additional analysis from the perspective of natural conditions and atmospheric environments. For Beijing and Wuhan, two inland cities, we explored this from the perspective of variations in the origin of the water vapor. The China Land Cover Dataset (CLCD) for 2013 and 2022 was used to indicate land cover changes in Beijing and Wuhan. The CLCD categorizes land cover into nine classes, which we reclassified into two broad categories: impervious and non-impervious surfaces. Subsequently, a raster calculator was used to obtain the increase in impervious surfaces by 2022. Figure 5 presents the extent of impervious surfaces in 2013 and highlights the expansion of impervious areas in Wuhan and Beijing by 2022.

In Wuhan, which has experienced rapid urbanization over the past decade, the area of impervious surfaces in the city has increased by 26.83%. Rapid urbanization impacts the hydrological balance of a city [65], which can lead to urban dry and wet island effects [66]. This may reduce local evapotranspiration, decrease water vapor sources, and reduce atmospheric PWV. A previous study demonstrated that atmospheric humidity decreased with the rapid urbanization of Wuhan [67]. Simultaneously, Beijing has increased its impervious surface by 7.08% over the past ten years, resulting in lower humidity in the atmosphere [68,69]. Moreover, Beijing is located in a semi-arid region, and groundwater has been experiencing persistent drought over the past 10 years [70,71], leading to reduced groundwater evaporation through soil pores [72]. Therefore, increased urbanization and persistent groundwater droughts are likely to be the key drivers of the decreasing PWV trend in Beijing.

For tropical southern Taiwan, previous research has suggested that the decreasing trend may be attributed to changes in ENSO phases [73]. The trend of water vapor in the tropics during 1988–2003 may be related to interdecadal Pacific oscillation interdecadal variations from a warm period (1977–1998) to a cool period (1999–2003) [74]. Thus, we explored it from the ENSO perspective, utilizing the two most commonly used ENSO discriminant indices to characterize ENSO, namely, MEI and ONI (https://psl.noaa.gov/enso/, last access on 10 April 2024).

Figure 6 illustrates the MEI and ONI from 2013 to 2022 at a temporal resolution of one month. ENSO underwent several phase changes during the study period. A stronger El Niño event occurred during 2015–2016. However, in the following years, La Niña events occurred more frequently and lasted longer, especially during 2020–2022, when La Niña persisted for nearly two years. The La Niña phenomenon usually enhances easterly winds and upwelling, leading to a decrease in sea surface temperature [75], which reduces water vapor evaporation. Therefore, the PWV in southern Taiwan shows a downward trend.

3.3. Frequency-Domain Characterization of GNSS PWV

3.3.1. Annual/Semi-Annual Period of GNSS PWV

The time–frequency features of GNSS PWV from different areas were investigated using the Cmor wavelet. We downsampled the temporal resolution of the GNSS PWV time-series data to a daily scale and interpolated the missing values to obtain continuous datasets. To mitigate the boundary effects at both ends of the series, we extended the data on both sides and removed the extended wavelet transform coefficients, thereby generating wavelet coefficient plots of water vapor for each regional station.

Figure 7 shows the wavelet coefficient plots for BJFS, BJNM, CHAN, JFNG, WUH2, and WUHN, while those for SHAO, HKWL, HKWS, and KMNM are presented in Figure 8 and those for TWTF, TCMS, TNML, CKSV, and NCKU are shown in Figure 9.

The first column in Figure 7, Figure 8 and Figure 9 shows the contour maps of the modulus for the stations, where the horizontal and vertical coordinates indicate the observation time and timescale, respectively, which can determine whether there is a periodic change. The second column shows the wavelet variance of the wave energy distribution of the PWV time series at different scales for each station. The main period (timescale) was determined by calculating the wavelet variances. The third column shows the fluctuation in the change in the main periods.

According to column 1, there is a significant fluctuation at all stations in the vertical coordinate timescale of approximately 500–700 d. Moreover, for the BJFS, BJNM, CHAN, JFNG, WUH2, and WUHN stations, there is another fluctuation with a small amplitude and inconsistent fluctuation time in the timescale of approximately 200–300 d, whereas the other nine stations do not exhibit such a fluctuation. The exact first and second main periods for each station were identified by locating the amplitude peaks based on the variance trends shown in column 2. The main periods for all the stations are listed in

Table 3. Main periods and fluctuation periods in the main periods of GNSS stations.

Stations	First Main Period/d	First Fluctuation Period/d	Second Main Period/d	Second Fluctuation Period/d
BJFS	552	363	278	182
BJNM	562	370	286	184
CHAN	552	361	277	183
JFNG	549	365	277	182
WUH2	542	368	280	185
WUHN	542	364	281	180
SHAO	611	386	/	/
HKSL	553	364	/	/
HKWS	553	364	/	/
KMNM	544	366	/	/
TWTF	551	362	/	/
TCMS	558	366	/	/
TNML	569	365	/	/
CKSV	542	364	/	/
NCKU	539	366	/	/

. Except for the SHAO station, the first main periods at the other stations are concentrated around 550 d. Additionally, the second main periods at BJFS, BJNM, CHAN, JFNG, WUH2, and WUHN are nearly identical, at approximately 277–286 d.

The extracted first and second main periods of each station are further used to illustrate the wavelet coefficient fluctuations, as shown by the blue and red wavy lines in column 3, respectively. The first main period of each station exhibits fluctuations with a cycle of approximately one year. In the second main period at BJFS, BJNM, CHAN, JFNG, WUH2, and WUHN, a semi-annual (~6 month) cycle can be observed. However, its amplitude is relatively small, unstable, and occasionally variable compared to the annual cycle.

Therefore, it can be concluded that all stations across the seven regions exhibit a pronounced annual cycle with a significant amplitude. Additionally, inland areas, such as Beijing, Changchun, and Wuhan, also display semi-annual cycles of smaller amplitudes, which are absent in the coastal regions.

Based on the results of the annual/semi-annual period, it can be verified that the stations in coastal cities have only an annual period, whereas the stations in inland cities have both annual and semi-annual periods with low amplitudes. For coastal cities, water vapor is mainly dependent on the ocean, and the stabilizing and regulating effect of the ocean creates a continuous and smooth supply of water vapor, resulting in insignificant semi-annual cycle fluctuations. The ocean warms up more slowly than land in summer, creating low-pressure areas that may attract more water vapor inland; the opposite is true in winter. Moreover, periodic variations in monsoon activities (such as the East Asian monsoon) may also lead to phased increases or decreases in water vapor in inland regions during specific seasons (e.g., late spring to early summer and late autumn to early winter), resulting in semi-annual cycle fluctuations [76]. Therefore, sea–land thermal differences and monsoon systems may contribute to the phenomena where the semi-annual cycle is significant in inland cities but not in coastal cities.

3.3.2. Seasonal/Monthly Fluctuation Period of GNSS PWV

To identify the period on a smaller timescale, the time series of PWV in 2018 was selected for wavelet transformation to obtain wavelet coefficient maps of each regional station. Considering the continuity of the datasets, the BJNM, TCMS, and TNML stations were excluded from this part of the experiment. The wavelet coefficient plots of the stations in each region are shown in Figure 10 and Figure 11.

As shown in column 1 of Figure 10, there is a prominent wave peak and a secondary peak of smaller amplitude at BJFS and CHAN, corresponding to their first and second main periods, respectively. JFNG, WUH2, WUHN, and SHAO display three peaks of comparable amplitudes, suggesting more balanced periodic components. For HKSL, HKWS, KMNM, TWTF, CKSV, and NCKU, shown in column 1 of Figure 11, there are multiple large-amplitude and small peaks, indicating that the main periods are more complex in these areas. The main periods extracted for all the stations are listed in

Table 4. Main periods and fluctuation periods of GNSS stations in 2018.

Stations	Main Period/d	Fluctuation Period/d
BJFS	160/43	105/31
CHAN	176/44	113/26
JFNG	94/59/39	62/38/25
WUH2	97/59/39	64/40/26
WUHN	94/60/39	64/40/26
SHAO	97/60/40	56/40/25
HKSL	184/136/86/41/18	121/90/56/27/11
HKWS	184/136/86/42/18	122/92/57/27/12
KMNM	186/133/91/41/18	125/93/58/27/12
TWTF	132/90/41/18	93/58/27/12
CKSV	194/88/43/18	121/56/27/12
NCKU	184/88/43/18	121/56/28/12

.

The wavelet coefficient fluctuations are plotted based on these identified dominant periods, as illustrated in column 3 of Figure 10 and Figure 11. At BJFS and CHAN, the first main period shows fluctuations with a seasonal cycle, whereas the second main period exhibits a shorter fluctuation of approximately one month. At JFNG, WUH2, WUHN, and SHAO, fluctuations with approximately two-month, forty-day, and one-month cycles can be identified. Several medium- and short-period fluctuations can be observed at stations with multiple complex dominant periods, such as HKSL, HKWS, KMNM, TWTF, CKSV, and NCKU.

In summary, cycles of seasonality and one month exist in Beijing and Changchun. Shanghai and Wuhan display two-month, forty-day and one-month cycles, while multiple medium and short periods are exhibited in Kinmen County, Hong Kong, and Taiwan. According to the findings for the seasonal/monthly period, the medium-term periodic characteristics of PWV demonstrate a clear latitude-dependent pattern.

At higher latitudes such as Beijing and Changchun, PWV cycles tend to be longer and more stable. This is largely due to the dominance of large-scale, stable atmospheric circulation systems such as the polar vortex and westerlies, which introduce regular seasonal variations [77,78]. In these regions, the relatively flat terrain also provides minimal interference with atmospheric motion, facilitating the formation of regular oscillations.

In contrast, mid- and low-latitude regions experience complex and irregular PWV variations. Mid-latitude areas, such as Shanghai and Wuhan, lie in transitional zones influenced by a mixture of weather systems, including the subtropical high, the Meiyu front, cold fronts, and typhoons. These interacting systems generate multiscale fluctuations, such as two-month plum rain cycles and forty-day typhoon intervals. In low-latitude regions such as Kinmen, Hong Kong, and Taiwan, atmospheric dynamics are dominated by tropical convection and short-lived systems, such as thunderstorms and tropical cyclones. Furthermore, water vapor transport is more stable and zonally directed at higher latitudes, while at lower latitudes it is dominated by disorganized local convection, further weakening identifiable cycles [79].

3.4. Short-Time Frequency-Domain Characterizations of GNSS PWV

To further investigate the short-term frequency-domain characteristics of PWV, time-series data from 2018 were selected and processed using the FFT method to extract the frequency-domain variations in PWV at each station. As shown in Figure 12, the vertical coordinates represent the amplitude and the horizontal coordinates represent the time spans from 0 to 2 d (0–48 h), which corresponds to a time frequency of 0.5~∞ cpd.

The stations located in the Shanghai and northern Taiwan areas have only a significant 1 d period (diurnal) in the frequency domain. In contrast, stations in Wuhan, Hong Kong, and Kinmen County display prominent period changes of 0.5 d (semi-diurnal). Stations in Beijing and southern Taiwan demonstrate both 0.5 d and 1 d periodic components, with the amplitude of the diurnal period being greater than that of the semi-diurnal period. Compared with these regions, the frequency-domain characteristics of PWV in the Changchun area exhibit minimal variation and lack any prominently amplified signals, reflecting more stable short-term fluctuations.

In summary, except for Changchun, GNSS stations in six typical cities display pronounced periodic variations in PWV at semi-diurnal, diurnal, or both frequencies. In contrast, Changchun exhibits insignificant variations in the frequency domain.

This may be because diurnal water vapor cycles are primarily driven by solar radiation. Nonetheless, other factors, such as temperature, local atmospheric circulation, vertical air motion, evapotranspiration, condensation, and precipitation, also modulate short-term variability [80]. Since solar radiation is generally stronger at lower latitudes, water vapor responds more sharply, with multiple fluctuation periods occurring within a single day.

Changchun, located at mid-to-high latitudes, has a temperate continental monsoon climate and receives less solar radiation. This limits evaporation and reduces atmospheric moisture. Although eastern and southern Changchun are not far from the ocean, moist summer winds are weakened by the Changbai Mountain. In winter, the open terrain of the Songliao Plain facilitates the dominance of cold, dry Siberian air masses. Together, these geographic and climatic factors limit the occurrence of short-term PWV periodicities in the Changchun region.

3.5. The Effect of Atmospheric Temperature on PWV

To further explore the main factors driving the periodic characteristics of PWV, experiments focusing on atmospheric temperature were conducted. Similarly, we selected the 10-year and 2018 time series to apply wavelet transforms to the temperature (T_2m and T_h) data of the representative stations. For T_h, the stratified air temperature level with the highest correlation with the PWV at each station was selected based on the correlation analysis presented in

Table 5. Correlation statistics of vertical atmospheric temperature (T_2m and T_h) and PWV at each station.

Height	BJFS	CHAN	JFNG	SHAO	HKSL	TWTF	CKSV	KMNM
2 m	0.71	0.76	0.71	0.76	0.74	0.73	0.71	0.78
1000 hPa	0.70	0.77	0.70	0.76	0.73	0.75	0.74	0.77
975 hPa	0.69	0.77	0.69	0.75	0.73	0.74	0.74	0.75
950 hPa	0.69	0.77	0.70	0.75	0.75	0.74	0.74	0.75
925 hPa	0.69	0.77	0.72	0.76	0.78	0.74	0.75	0.76
900 hPa	0.70	0.78	0.73	0.77	0.8	0.74	0.76	0.77
875 hPa	0.71	0.78	0.75	0.77	0.81	0.74	0.76	0.78
850 hPa	0.72	0.79	0.76	0.78	0.81	0.74	0.76	0.78
825 hPa	0.74	0.79	0.77	0.78	0.8	0.75	0.75	0.78
800 hPa	0.75	0.80	0.78	0.78	0.79	0.76	0.74	0.77
775 hPa	0.76	0.80	0.79	0.79	0.77	0.76	0.73	0.77
750 hPa	0.77	0.81	0.80	0.79	0.75	0.75	0.72	0.76
700 hPa	0.79	0.81	0.81	0.8	0.68	0.74	0.68	0.73
650 hPa	0.8	0.81	0.82	0.81	0.59	0.70	0.60	0.67
600 hPa	0.81	0.82	0.82	0.81	0.48	0.65	0.50	0.60

.

As a result, the temperature at 600 hPa was selected for BJFS, CHAN, JFNG, and SHAO, whereas the temperatures at 850 hPa were chosen for HKSL, TWTF, CKSV, and KMNM in the wavelet transform analysis. The wavelet coefficient plots for the 10-year temperatures (T_2m and T_h) are presented in Figure 13 and Figure 14, respectively. As shown, all stations exhibit one main period of approximately 550 d, corresponding to an annual variation cycle. Notably, no significant semi-annual periodic components can be observed in the temperature data for any of the stations.

The wavelet coefficient plots of temperature (T_2m and T_h) in 2018 for the BJFS, CHAN, JFNG, and SHAO stations are presented in Figure 15, whereas those for the HKSL, TWTF, CKSV, and KMNM stations are shown in Figure 16. By comparing Figure 15 and Figure 16 with Figure 8 and Figure 9, it can be observed that the PWV characteristics of BJFS and CHAN align with the wavelet coefficient patterns of T_2m, both exhibiting variability periods of approximately four months and one month. For the CKSV station, PWV characteristics on the timescale of 1–120 d is also more consistent with the wavelet coefficient patterns of T_2m at 1–120 d. For the other stations, their PWV features on the timescale of 1–120 d conform to the wavelet coefficient patterns of T_h at 1–120 d.

Temperature affects the atmospheric PWV in two ways. On the one hand, when the temperature near the surface rises, surface water evaporation accelerates, leading to an increase in PWV. On the other hand, when the relative humidity in the mid–lower troposphere (especially over oceanic regions) remains constant, the amount of water vapor increases with increasing air temperature. According to the Clausius–Clapeyron equation, the water vapor content can increase by 6–7% per 1 K increase in temperature [73,81]. As the temperature rises, water molecules move more rapidly, allowing more molecules to overcome the surface tension of the liquid and enter the atmosphere, causing the saturated water vapor pressure to rise. This also directly drives the actual water vapor pressure to increase in the same proportion to maintain the dynamic equilibrium between water vapor and liquid water.

The periodic characteristics of PWV in Beijing and Changchun are consistent with those of T_2m, suggesting that PWV in these two regions is predominantly influenced by temperature-driven surface water evaporation. The characteristics of southern Taiwan within the timescale of 1–120 d are consistent with T_2m, illustrating that the medium- and short-term periods of this region are also predominantly influenced by temperature-driven surface water evaporation. Conversely, in Wuhan, Shanghai, Hong Kong, northern Taiwan, and Kinmen County, the PWV variation conforms to the Th patterns of the strongest correlation coefficient at the pressure level in the timescale of 1–120 d. This indicates that, in these regions, the PWV is more affected by the second mechanism within this timescale.

The annual PWV period aligns with the temperature variation patterns, although the affecting mechanism remains unclear. For Beijing and Changchun, the seasonal and monthly periods are affected by temperature-accelerated surface water evaporation, and medium and short periods in southern Taiwan are also affected by this mechanism. However, for the other regions, except for the seasonal period, the medium and short periods are affected by temperatures which enhance the moisture retention capacity of the atmosphere, suggesting that seasonal periods may be affected by more complex weather systems. In summary, these findings indicate that temperature plays a crucial role in regulating annual, medium-, and short-term PWV variation periods across the major regions of eastern China.

4. Limitations and Prospects

In this study, we focused on major cities in eastern China to investigate PWV variability. While this focus was partly driven by the region’s economic development and high population density, it also imposed notable limitations. China is a vast country with huge differences in climate conditions, topography, and atmospheric circulation in different regions [82], all of which shape distinct PWV distribution characteristics, changing patterns, and influencing mechanisms. Therefore, the findings of this study may not be fully applicable to other regions not included in the study.

Moreover, PWV, a key parameter in the field of meteorology and hydrology, is affected by many complex factors, including atmospheric physical processes, geographic environmental features, and human activities. However, our study primarily considered the effect of temperature on PWV and only at a qualitative level. The absence of a quantitative assessment limits a deeper understanding of the temperature–PWV relationship in this region.

Future research should broaden the scope of this study to encompass all Chinese cities in the research field, comprehensively consider the mechanisms of multiple influencing factors on PWV, and deepen studies on the quantitative relationship between these factors and PWV. Notably, abrupt changes in PWV have been identified as precursors to rainfall events [83]. Previous studies have shown that PWV typically rises sharply within 0–12 h before precipitation onset, maintains high-level fluctuations during the event, and rapidly returns to baseline levels afterwards [84]. A deeper understanding of these dynamics will support more accurate characterization of PWV variability and provide a more reliable scientific basis for the fields of meteorological forecasting, water resource management, and climate change research.

5. Conclusions

In this study, GNSS observations from IGS stations between 2013 and 2022 were utilized to retrieve PWV data across representative regions in eastern China. By integrating long-term GNSS-derived PWV time series with multiple frequency-domain analysis methods (Bayesian statistics, wavelet transform, and FFT), the multiscale time–frequency characteristics of PWV were described in Changchun, Beijing, Shanghai, Wuhan, Taiwan, Hong Kong, and Kinmen County. Additionally, the vertical atmospheric temperature was uniquely incorporated to explore the driving mechanisms behind the regional PWV variability. The following findings can be drawn:

(1) Subregional declines in PWV can be identified in Wuhan, Beijing, and southern Taiwan—contrasting with the increasing trends in other regions. In Wuhan, rapid urbanization has driven the growth of impervious surfaces, resulting in the suppression of local atmospheric moisture levels. In Beijing, the decline is primarily driven by a combination of groundwater depletion and urbanization, both of which contribute to reduced local evapotranspiration. In southern Taiwan, this decline is primarily driven by ENSO-related climate anomalies, which disrupt regional moisture transport patterns.

(2) The PWV periodicities exhibit clear latitudinal and coastal–inland gradients. Inland stations (Beijing, Changchun, and Wuhan) show weaker annual and semi-annual signals, whereas coastal stations (Shanghai, Kinmen County, Hong Kong, and Taiwan) primarily exhibit annual cycles. Higher-latitude stations display stronger seasonal and monthly oscillations, mid-latitude stations present intermediate-scale variations, and low-latitude regions exhibit richer medium- and short-term fluctuations.

(3) Pronounced diurnal and semi-diurnal PWV oscillations can be observed in six typical cities, while Changchun exhibits minimal short-term fluctuations—likely due to its northern latitude and complex topographic influences.

(4) Temperature plays a key role in modulating PWV periodicities. The annual PWV cycle is closely aligned with temperature fluctuations. In Beijing, Changchun, and southern Taiwan, near-surface temperature drives groundwater evaporation, amplifying seasonal signals. In contrast, in other regions, the medium- and short-term PWV variations are influenced by temperature-driven increases in atmospheric moisture capacity, indicating more complex thermodynamic interactions.

Overall, this study offers solid data support and an analytical framework for discovering the geographic and climatic driving mechanisms underlying water vapor variations. These findings provide valuable insights for enhancing weather and climate modelling, sustainable urban planning, and regional drought and flood early warning.