Available online at www.sciencedirect.

com

ScienceDirect
Advances in Space Research 73 (2024) 4903–4921
www.elsevier.com/locate/asr

Improving MODIS-IR precipitable water vapor based on the

FIDWFT model
Xiangrong Yan a,b,c, Weifang Yang a,b,c,⇑, Nan Ding d, Fenglin Gao d, Yibo Peng a,b,c
a
Faculty of Geomatics, Lanzhou Jiaotong University, Lanzhou, China
b
Nation-Local Joint Engineering Research Center of Technologies and Applications for National Geographic State Monitoring, Lanzhou, China
c
Gansu Provincial Engineering Laboratory for National Geographic State Monitoring, Lanzhou, China
d
School of Geography, Geomatics and Planning, Jiangsu Normal University, Xuzhou, China

Received 29 November 2023; received in revised form 18 February 2024; accepted 20 February 2024
Available online 4 March 2024

Abstract

The monitoring of precipitable water vapor (PWV) is of significant importance for meteorological research and rainstorm prediction.
The moderate resolution imaging spectroradiometer (MODIS) is one of the most widely used remote sensing PWV instruments aboard
the Terra and Aqua satellites. The main objective of this study is to develop a fused inverse distance weighting and the Fourier transform
model (FIDWFT) to calibrate MODIS Infrared (IR) PWV. In this research, one year of ground observation data from 17 GPS stations
were utilized to evaluate the accuracy of MODIS IR PWV in Hong Kong. Initially, the PWV obtained by one radiosonde was used to
verify the GPS PWV. Comparison between GPS and radiosonde showed good agreement with an R-squared of 0.97. After the estimation
of GPS PWV in the study area was evaluated, MODIS IR products in the study area were extracted for evaluation. The comparison
between MODIS and GPS showed that the R-squared is 0.81 and the mean bias (MB) is
1.60 mm. Considering the spatial interpolation
and calibration problems of MODIS products, new model used an inverse distance weighting method to make MODIS IR PWV and
GPS PWV completely coincide in space, and fifteen MODIS pixel points were selected for interpolation. Then the MODIS IR PWV
was calibrated by GPS PWV using Fourier transform model. Results showed that compared with the MODIS IR PWV before calibra-
tion, the R-squared increases from 0.81 to 0.94, the root mean square error decreases from 8.72 mm to 4.98 mm, and the MB decreases
from
1.60 mm to
0.83 mm. Therefore, the method is feasible to estimate MODIS IR water vapor, and it achieves better results.
Ó 2024 COSPAR. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/
by-nc-nd/4.0/).

Keywords: GPS; MODIS IR; Precipitable water vapor; Inverse distance weighting; Fourier transform model

1. Introduction vapor (Solomon et al., 2010). At the same time, water

vapor plays a significant role in severe disaster weather,
Water vapor is water in the gaseous state in the atmo- rainstorm prediction, and global climate change (Zhang
sphere, which accounts for a minor part of the overall et al., 2015). It is significant to national defense, economic
amount of the atmosphere but has complex temporal and construction, ecological environment, and people’s lives
spatial distribution features. Water vapor is also a signifi- and health. Therefore, in order to predict weather systems
cant greenhouse gas, and a strong correlation exists on a small- to medium-scale, it is crucial to have access to
between climate change and the rise of tropospheric water fast, precise, high spatiotemporal resolution, and all-
weather water vapor observation data.
⇑ Corresponding author at: Faculty of Geomatics, Lanzhou Jiaotong Over the past few decades, various methods such as
University, Lanzhou, China. satellite remote sensing observations (Gao et al., 2003),
E-mail address: [email protected] (W. Yang).

https://doi.org/10.1016/j.asr.2024.02.036
0273-1177/Ó 2024 COSPAR. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

radiosondes (Niell et al., 2001), microwave radiometers pressure and sea level pressure, and the error of the modi-
(Sánchez et al.,2013), ground-based sun photometry fied MODIS NIR PWV could be reduced to 1.41 mm. Xu
(Alexandrov et al.,2009) and Global Positioning System et al., (2022) established an empirical regression relation-
(GPS) receivers (Champollion et al., 2005; Elgered et al., ship between differential PWV data and MODIS NIR
1997) have been used to measure the water vapor content PWV products using the linear regression method, and
in the air. Among the above-mentioned methods, radio- the results showed that the average deviation of the model
sonde is the most traditional way to measure PWV. PWV in China is reduced to
0.38–––2.01 mm. Under the con-
may be obtained at each station with an accuracy of a straint of the Crustal Movement Observation Network of
few millimeters thanks to the use of radiosondes (Niell China (CMONOC), Liu et al., (2019) constructed the
et al., 2001). Therefore, PWV radiosonde observations MODIS NIR PWV calibration model for each season in
are frequently used as a reference by meteorologists. mainland China, and the calibration accuracy could reach
Unfortunately, limitations such as the high cost of a millimeter level. Liu et al., (2023) presented a calibration
radiosondes and low spatiotemporal resolution (twice a model for MODIS NIR PWV that utilizes the fast Fourier
day) hinder the availability of this technology for monitor- transform (FFT) and least squares approach (LS). The
ing large-scale weather changes. Another reliable and com- root mean square error (RMSE) of MODIS NIR PWV
monly used technology for PWV acquisition is GPS decreased by 35.7 % overall following this model calibra-
measurement. The benefits of GPS water vapor detection tion, while most of them were still larger than 6 mm, par-
technology are low operating costs, a wide spatiotemporal ticularly during the summer. The mean bias (MB)
detection range, strong application value, and development decreased to approximately 2 mm, a remarkable improve-
potential (Ding, 2018). In addition, the accuracy of GPS ment. This indicated that the model produced better cali-
PWV can reach 1  2 mm (Ware et al.,1993; Yao et al., bration effects and showed a higher degree of fitting,
2017; Wang et al., 2007), which can meet the requirements hence improving the MODIS NIR PWV’s reliability.
of meteorological applications such as meteorological sci- Zhang et al., (2021) constructed a linear regression equa-
entific research, rainstorm prediction, and environmental tion for MODIS NIR PWV based on GPS PWV and fused
monitoring. Though GPS PWV has discrete geographical it into the tomographic observation equations, improving
distribution characteristics and GPS stations are few, the the reconstruction quality of the three-dimensional water
technology can be combined with other water vapor detec- vapor field. Bai et al., (2021) used the GPS PWV of each
tion methods to compensate for the spatial discontinuity CMONOC station to calibrate the linear model of MODIS
issue. NIR PWV of each epoch in China, and found that RMSE
In recent years, satellite remote sensing technology has of PWV decreased significantly except for the plateau cli-
developed rapidly, and it can continuously observe water mate region, with an average relative improvement of
vapor distribution on a global scale. The Moderate Resolu- 7.75 %. Sam et al., (2019) proposed a height-function linear
tion Imaging Spectroradiometer (MODIS) has excellent model, and the MB and the RMSE of MODIS NIR PWV
spatial resolution and complete spectrum observation were reduced by 2 mm and 1 mm. However, most of these
capabilities, and it can utilize the differential approach to studies are to establish the functional relationship between
retrieve PWV based on water vapor absorption in the GPS PWV and MODIS NIR PWV, and the calibration
near-infrared and thermal-infrared channels (Kaufman models are mainly based on linear regression models. At
et al., 1992; Gao et al., 2003). And MODIS can obtain finer the same time, they usually use the nearest neighbor
surface feature information to calculate PWV. It also does method to match the MODIS pixel points with the GPS
spatial interpolation in order to produce a continuous station at the spatial point position.
PWV distribution. Unfortunately, the accuracy of MODIS Several studies have also used ground-based PWV
PWV is affected by cloud cover and ground spectral reflec- observation data to evaluate MODIS infrared (IR) PWV
tion errors, resulting in low accuracy and a relative error of observation products over areas of the world. Chang
more than 10 % (Gubuz et al., 2017; Oliveira et al., 2017). et al., (2014) proposed a differential linear calibration
Therefore, the usability of MODIS for monitoring weather model (DLCM) to calibrate the MODIS IR PWV product
changes (such as thunderstorms and short-term weather during the nighttime. Comparisons of the DLAM cali-
forecasting) is hampered by restrictions including low accu- brated MODIS IR PWV against GPS PWV showed that
racy impacted by cloud pollution and limited time resolu- the standard deviation (STD) of the mean difference
tion. Additionally, and it is necessary to calibrate decreases from 7.48 to 5.69 mm, together with the RMSE
MODIS PWV. reduction from 12.52 to 5.58 mm. Results showed that
In order to improve the accuracy of MODIS water the DLCM can effectively enhance the accuracy of MODIS
vapor products, many scholars use GPS PWV observation IR retrievals at nighttime. Merrikhpour et al., (2017) was
data to calibrate MODIS near-infrared (NIR) PWV prod- to investigate and improve the presented algorithm in to
ucts (Liu et al., 2006, 2013, 2015; McAtee and Maier, 2006; extract PWV values from MODIS IR images to optimize
Prasad and Singh, 2009; Wang et al., 2009). Liu et al., PWV estimation in Iran. The results showed that in highest
(2006) established the exponential regression equation of accuracy, the coefficient of determination R-squared and
MODIS NIR PWV and GPS PWV considering the surface RMSE was 0.878 and 2.702 mm. Maghrabi (2019)
4904
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

analyzed the power spectrum for the period 1985–2016 polation method used in this study and the model estab-
using Fourier Transform (FT) technique to study the inter- lished in this study are introduced.
annual, inter-monthly, and seasonal variations and trends
in PWV using radiosonde observations from seven stations 2.1. Datasets
in Saudi Arabia from 1985
2016. It was found that the
PWV cycle in the Arabian Peninsula may be related to 2.1.1. Radiosonde
the influence of solar activity as well as land-based meteo- Radiosonde is one of the most traditional tools for PWV
rological phenomena. Ghaderi et al., (2022) extracted PWV measurement, which can measure the information on
values using meta-heuristic algorithms for MODIS surface atmospheric pressure, water vapor pressure, temperature,
reflectance data. It is found that the Grasshopper Opti- and relative humidity at various levels of the atmosphere.
mization Algorithm (GOA) algorithm acquired the best Using the vertical profile of these atmospheric parameters,
result. MODIS IR PWV product evaluation with radio- PWV can be calculated with an error of 5 %8% (Pérez-
sonde data showed RMSE and R-squared of 6.53 mm Ramı́rez et al., 2014; Turner et al., 2003). However, it is still
and 0.69, respectively, and with GPS PWV showed RMSE often used by meteorologists as a reference for PWV mea-
and R-squared of 6.55 mm and 0.68, respectively. How- surements. There are only two sounding balloon launches
ever, most of these studies improve the algorithm of each day (UTC 00:00 and UTC 12:00) due to the expensive
extracting MODIS IR PWV, and do not consider the prob- expense of the radiosonde.
lem of incomplete coincidence in space when comparing To verify the reliability of GPS PWV, this study selects
with GPS PWV. The RMSE of the improved MODIS IR PWV data extracted from a Hong Kong radiosonde
PWV is still greater than 5 mm and the R-squared is less (HKKP) from January to December 2022 at UTC 0:00,
than 0.9 compared to the GPS PWV. which can be obtained from the University of Wyoming
Considering that MODIS pixels and GPS stations do website https://weather.uwyo.edu/upperair/souding to
not coincide completely, it is necessary to adopt appropri- download.
ate interpolation methods to make the two spatial points
consistent. Hu et al., (2013) used the inverse distance
2.1.2. GPS data
weighting, Kriging, and radial basis function methods to
PWV is estimated on an hourly basis from ground-based
perform interpolation experiments on MODIS water
GPS measurements, and it is considered as a reference for
vapor. However, only one day’s MODIS data was used
evaluating calibration models. PWV is obtained based on
in this study, and the research methods needed to be
the Zenith total delay (ZTD), and PWV retrieval is
improved. Unfortunately, most of the current MODIS
assessment and calibration studies directly compare GPS described in detail in Section 2.2. The GPS product is pro-
PWV with MODIS PWV, which is not rigorous. And in vided by the Hong Kong Satellite Positioning Reference
MODIS water vapor calibration test, the calibration mod- Station Network (https://www.geodetic.gov.hk/sc). This
els are mostly linear regression models, and the research study uses GPS observation data from 17 stations in Hong
methods are insufficient. In addition, most of the studies Kong’s continuously operating reference system (CORS)
are focused on the calibration and applications of MODIS for the experiment, the time resolution is 1 h. The study
NIR PWV measurements, whereas only a few studies have period is from January to December 2022. The location
contributed to the analyses of MODIS IR retrievals. There- of the stations is shown in Fig. 1.
fore, we combine the advantages of GPS and MODIS, and
use GPS technology to calibrate MODIS IR PWV. We 2.1.3. MODIS products
propose a fusion inverse distance weighting and Fourier MODIS is a sensor developed by the United States and
transform combined (FIDWFT) model. Compared with carried on board the Terra and Aqua satellites to acquire
previous studies, we consider the MODIS interpolation high-resolution remote sensing data of the Earth’s surface
and calibration problems together. Eight interpolation and atmosphere. MODIS has 36 medium-resolution spec-
methods are used to register MODIS pixel points and tral bands, covering the spectral region of 0.4  15 lm,
GPS station, so that the two are consistent in spatial point and observes the Earth’s surface once every 1 to 2 days,
position. Then, a linear model and five nonlinear models which can provide high-time resolution remote sensing
are constructed to calibrate MODIS PWV using GPS tech- observation data on a global scale.
nology to further improve the accuracy of MODIS PWV. In MODIS data, channels 17 (0.905 lm), 18 (0.936 lm),
Finally, the optimal interpolation parameters, interpola- and 19 (0.940 lm) have high sensitivity for NIR water
tion method, and calibration model are combined. vapor inversion. The MODIS IR water vapor retrievals
are derived from band 24 to band 36 (between 4.47 and
2. Data and methodology 14.24 mm), excluding band 26. These bands are greatly
affected by the absorption and scattering of atmospheric
This section describes the radiosonde, GPS, and water vapor, so they can be used to acquire atmospheric
MODIS datasets used in this study and their processes water vapor content. The PWV of MODIS inversion is
for generating PWV products. At the same time, the inter- referred to as MODIS PWV in this study. MODIS water
4905
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

have a lot of room for improvement. Enhancing the preci-

sion of IR PWV products can facilitate MODIS PWV
retrieval in both day and night conditions (Chang et al.,
2019). This has also helped with the analysis of IR retrie-
vals during the night, when the transportation of water
vapor from the atmosphere to the ground surface plays a
significant role in the water budget, especially in regions
that are semiarid and arid (Chen et al., 2008; Li et al.,
2003; Prasad et al., 2009). In this study, IR water vapor
products with a resolution of 5 km  5 km are used for
experiments. Good-quality 202 IR images of MOD05
and MYD05 in Hong Kong 2022 are chosen for the exper-
iment. The entire MODIS IR water vapor distribution
throughout southeast China is depicted in Fig. 2 (a), and
the GPS station distribution position on the MODIS IR
images of Hong Kong is shown in Fig. 2 (b).

2.2. Method
Fig. 1. GPS stations’ location in Hong Kong. The solid green circle
indicates the location of the 17 GPS stations in Hong Kong, and the 2.2.1. PWV inversion with radiosonde
yellow five-pointed star indicates the radiosonde Station (HKKP) location PWV can be calculated from the meteorological data
in Hong Kong.
measured by radiosonde (Yang et al., 2002):
Z p2
vapor products can be provided by NASA (https://lads- PWV ¼ qg1
qdp ð1Þ
p1
web.modaps.eosdis.nasa.gov), which include MOD05_L2
products from the Terra platform and MYD05_L2 prod- where g is the acceleration of gravity, and the unit is m/s2; q
ucts from the Aqua platform. The two products are water is the density of liquid water, expressed in kg/m3. p1 and p2
vapor observation data of different observation periods in are the upper and lower limits of pressure. q is the specific
the same area, including infrared (IR) and near-infrared humidity, and the unit is g/kg. The value of q can be
(NIR) water vapor products. Moreover, the resolution of obtained from the following equation (Wang et al., 2015):
NIR water vapor products is 1 km  1 km, and the resolu-
0:622  es
tion of IR water vapor products is 5 km  5 km. The R- q¼ ð2Þ
squared of IR PWV products is about 83 %, and the R- p
0:378  es
squared of NIR PWV products can reach more than In the equation, es represents the water vapor pressure,
90 %. IR PWV products are not as reliable as NIR PWV the unit is hPa, and its calculation equation is as follows
products (Prasad and Singh, 2009), but IR PWV products (Bolton, 1980; Murray, 1967; Wang et al., 2016):

Fig. 2. PWV values’ distributions retrieved from MODIS data for each pixel at 06:00 UTC, August 1, 2022. Fig. 2 (a) is the MODIS water vapor map of
southeast China, and Fig. 2 (b) shows the enlarged MODIS water vapor map of Hong Kong with the red box in Fig. 2 (a). The triangles represent GPS
stations. The color bar represents the PWV values, its unit is mm. Moreover, the PWV values are produce for entire region of the pixels.
4906
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

RH  Es values are k1 = 77.604 K/hPa, k2 = 64.79 K/hPa,

es ¼ ð3Þ
100 k3 = 373900 K2/hPa. Tm is the atmospheric weighted mean
where RH stands for relative humidity, Es represents satu- temperature, and the unit is K.
rated water vapor pressure, the unit is hPa, and its calcula- In addition, the acquisition of PWV uses GAMIT 10.71
tion equation adopts the Goff-Gratch formula software to process the GPS observation data, the data
recommended by the World Meteorological Organization sampling interval is 5 min, the cut-off elevation angle is
(Goff, 1957): set to 10°, the earth tide model and sea tide model are
    introduced for correction, the tropospheric delay model is
lg Es ¼ 10:79586  1
TT d1
5:02808  lg TT d1 Saastamoinen model, and the mapping function model is
"  # set to VMF1. Three auxiliary IGS stations, BAKO,
Td

8:2969 T 1
1 CHAN, and GUAM, are introduced for a joint solution
þ1:50475  10
4  1
10
ð4Þ to weaken the correlation of zenith tropospheric delay.
"  # The PWV obtained by the solution is called GPS PWV in
T
4:76955 1
T 1 this study.
þ0:42873  10
3  10 d
þ 0:78614

where, ‘‘lg” stands for logarithmic base 10. T1 = 273.16 K, 2.2.3. Interpolation method
Td represents the dew point temperature in K. The nearest neighbor interpolation (nearest) assigns the
In addition, the PWV obtained by the radiosonde is point closest to the interpolation point as the gray value of
referred to as RS PWV in this study. that point. It is one of the simplest interpolation algo-
rithms, but only the influence of the nearest point on the
2.2.2. PWV inversion with ground-based GPS interpolation point is considered.
When the satellite passes through the troposphere region The triangulation with linear interpolation (TLI) creates
of the atmosphere during the transmission process from the multiple Delaunay triangles by connecting data points. The
satellite to the ground receiver, refraction occurs due to the three closest points are selected from the best triangles.
action of the atmospheric refractive index. This causes the Linear interpolation is then used to obtain data from the
signal propagation path to bend and the signal propaga- internally interpolated points (Yan et al., 2022).
tion speed to change, forming a delayed phenomenon. The natural neighbor interpolation (natural) relies on
Meanwhile, more than 3/4 of the water vapor in the entire the average weight of the obtained data. Nevertheless, the
atmosphere is gathered in the troposphere, and more than natural constructs a Delaunay triangle for the sample point
4/5 of the signal neutral delay occurs here, so the signal and chooses a point close to the interpolated area vertex in
neutral delay is collectively referred to as the tropospheric addition to interpolating values using all nearby points and
delay (Zhao et al., 2023, 2024). Then the tropospheric delay weighted gradually. That region is proportionate to the
is converted to the zenith direction through the mapping value weight.
function, and the total zenith delay is calculated as follows Cubic spline interpolation (cubic) is a way to approxi-
(Iwabuchi et al., 2000): mate an arbitrary function via multiple 3rd order polyno-
Z
mials connected at points (called knots) that are usually
ðnðsÞ
1Þds ð5Þ spaced at regular intervals of size. The second derivatives
of each polynomial is set to zero at each endpoint to ensure
ZTD stands for zenith total delay, ZHD stands for a continuous function.
zenith hydrostatic delay, and ZWD stands for zenith wet Biharmonic spline interpolation (BHS) can be used to
delay. PWV can be calculated by multiplying the ZWD find the minimum curvature from the distribution of irreg-
by the conversion factor P (Bevis et al., 1994): ular data points by using a green function. The interpola-
PWV ¼ P  ZWD ð6Þ tion curve is a linear combination of Green functions
centered at each data point. This method is more flexible
In the equation, P has no dimension and is directly than the spline method since both slopes and values can
determined by Tm (BEVIS et al., 1992): be used to find a surface (Yulianto et al., 2020).
106 Inverse distance weighting (IDW) is based on the
P¼   ð7Þ assumption that the interpolating data should be influ-
q  Rw k3
Tm
þ k 02 enced more by the nearby data and less by the more distant
Rd data, and the interpolating data are the weighted average
k 02 ¼ k 2
k 1  ð8Þ of the scatter data. Weights are usually inversely propor-
Rw
tional to a power of distance. Its equation is as follows:
where q is the density of water, and its value is 1 g/cm3. Rw Pn Z i
and Rd are gas constants of dry and wet water vapor, and i¼1 d k
their values are Rw = 0.462 jg-1k
1, Rd = 0.287 jg-1k
1. k1, Z ¼ Pn 1
i
ð9Þ
i¼1 d k
k2 and k3 are atmospheric refraction coefficients, and their i

4907
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

where Z is the PWV estimate of the interpolation point, Zi the FT model. The parameters of the model are shown in
is the PWV value of the MODIS pixel point, di is the dis- Table 1. It should be noted that the GF and FT models
tance between the ith pixel point and the interpolation are only applicable when the PWV is less than 90 mm.
point, n is the number of pixels used for interpolation, k
is the specified power, and two is used in this study. 2.2.5. FIDWFT model
Newton interpolation (Newton) is a commonly used The technical roadmap of the FIDWFT model pro-
numerical analysis technique that can be used to find the posed in this study is shown in Fig. 3. Firstly, the accuracy
values and derivatives of polynomial functions. When the of the PWV products obtained from GPS measurements is
number of sampling points is large, the order of its interpo- evaluated using the radiosonde. As long as the accuracy of
lation polynomial will be higher, thus increasing the inter- GPS PWV products is verified, GPS PWV will be used to
polation error. evaluate the MODIS IR PWV products. Then, MODIS
Kriging, synonymous to geostatistical interpolation, and IR PWV is calibrated based on GPS PWV. The MODIS
mining engineer D.G. Krige and statistician H.S. Sichel pixel points near each GPS station are sorted according
formulated it. Kriging is founded on a concept of random to the distance from GPS stations, and the number of suit-
functions with the surface or volume assumed one realiza- able MODIS pixel points is selected according to the dis-
tion of a random function with a known spatial covariance. tance priority to participate in interpolation. The number
Its calculation equation is as follows: of parameters with the highest interpolation accuracy is
X
n obtained. Next, nearest, TLI, natural, BHS, cubic, IDW,
Z¼ ki Zðxi Þ ð10Þ Newton, and Kriging methods are used to interpolate the
i¼1 water vapor values of GPS station locations, respectively.
Furthermore, LF, MLF, GF, PIF, EF, and FT calibration
where ki is the weight coefficient of MODIS pixel points
models are constructed to calibrate MODIS IR PWV.
participating in interpolation, and the sum of its values is
Replace the MODIS IR PWV for the GPS station location
1. n indicates the number of GPS stations. xi is the position
with a GPS PWV, and then put it into the calibration
of the ith MODIS pixel.
model to calibrate the MODIS IR PWV for the entire
region. Finally, the optimal interpolation parameter num-
2.2.4. Calibration method ber, interpolation method, and calibration model are com-
In this study, MODIS PWV is calibrated using six mod- bined to compare and analyze the original MODIS IR
els, including linear fitting (LF), multiple linear fitting PWV.
(MLF), power index fitting (PIF), exponential fitting
(EF), Gaussian fitting (GF), and Fourier transform (FT)
models. The equations for these models are given below: 3. Results and discussion
PWV MODIS calibration ¼ p0 þ p1  PWV MODIS initial ð11Þ 3.1. Comparison of PWV from radiosonde, GPS and
PWV MODIS calibration ¼ p0 þ p1  PWV MODIS initial þ p2  PWV 2 MODIS
MODIS initial

ð12Þ
In this section, we firstly use radiosonde data to assess
PWV MODIS calibration ¼ p1  PWV d
MODIS initial ð13Þ the quality of GPS PWV. After it is considered reliable,
the MODIS IR PWV is evaluated using the PWV of 17
PWV MODIS calibration ¼ p1  expðp 2  PWV MODIS initial Þ ð14Þ GPS stations in Hong Kong.
 
PWV MODIS initial
b
PWV MODIS calibration ¼ p1  expð
Þ ð15Þ 3.1.1. Comparison between radiosonde and GPS PWV
c
Independent measurements using radiosonde data are
PWV MODIS calibration ¼ p0 þ p1  cos ðPWV MODIS initial  wÞ necessary to confirm that GPS meteorology is a valid
PWV sensing technique in the study area (Liou et al.,
þ p2  sin ðPWV MODIS initial  wÞ ð16Þ
2001). For this purpose, the PWV of radiosonde data from
Equation (11) is the LF model, Eq. (12) is the MLF the HKKP station from January to December 2022 is com-
model, Eq. (13) is the PIF model, Eq. (14) is the EF model, pared with GPS PWV. The statistical results of GPS PWV
equation (15) is the GF model, and Eq. (16) is the FT and RS PWV during the study period are compared in
model. The PWVMODIS calibration is the calibrated MODIS Table 2. The results show that the PWV values obtained
PWV at the GPS station location. The PWVMODIS initial by GPS technology agree with those obtained by radio-
is the MODIS PWV before calibration, and it is the sonde, and the average R-squared is 0.97. In addition,
MODIS PWV after interpolating GPS stations to measure the difference between RS PWV and GPS PWV is slight,
the station location. p0 is the constant term; p1  p2 are the with mean bias (MB), root mean square error (RMSE),
model coefficients; b is the peak positions of the GF model; standard deviation (STD), and mean absolute error
c is the half-width information of the GF model; d is the (MAE) of 0.69 mm, 2.32 mm, 4.43 mm, and 1.56 mm,
power of the PIF model; w is the angular frequency of respectively.
4908
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

Table 1
Parameters of the MODIS PWV calibration model.
calibration Model p0 p1 p2 b c d w
LF 0.7897 6.47 \ \ \ \ \
MLF
7.16 1.55
0.0092 \ \ \ \
PIF \
2.00 \ \ \ 0.8016 \
EF \ 17.65 0.0183 \ \ \ \
GF \ 54.33 \ 65.84 46.37 \ \
FT 32.87
21.41
0.1107 \ \ \ 0.0477

Fig. 3. Technology roadmap.

Table 2
Comparison of PWV of 17 GPS stations and radiosonde in Hong Kong.
GPS Station MB/mm RMSE/mm MRE/% MAE/mm R-squared
HKCL 0.10 2.03 4.04 1.65 0.97
HKKS 0.67 2.14 4.31 1.44 0.97
HKKT 0.39 2.05 4.38 1.40 0.97
HKLM 0.44 2.10 4.43 1.48 0.97
HKLT 1.05 2.44 5.32 1.73 0.97
HKMW 1.04 2.34 4.84 1.55 0.97
HKNP 1.47 3.54 4.47 2.08 0.96
HKOH 1.00 2.74 4.65 1.94 0.96
HKPC 0.08 1.91 4.59 1.48 0.97
HKQT 0.35 2.05 4.15 1.38 0.97
HKSC 0.29 2.02 4.02 1.31 0.97
HKSL 1.01 2.24 4.60 1.50 0.97
HKSS 0.31 2.09 4.56 1.50 0.97
HKST 0.68 3.24 4.53 1.59 0.96
HKTK 1.11 2.36 3.83 1.86 0.97
HKWS 0.68 2.14 4.38 1.43 0.97
T430 1.00 2.08 4.13 1.23 0.97

4909
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

Therefore, the comparison between GPS PWV and RS also different. Table 3 shows the specific error statistics.
PWV in this study is consistent with the results of previous During this period, the overall MB of MODIS IR PWV
studies (e.g., Khaniani et al., 2019; Kwon et al., 2007; is
1.66 mm, and PWV products under all interpolation
Sharifi et al., 2015). methods show negative deviation, indicating that the
MODIS IR PWV is slightly underestimated. The Nash-
3.1.2. Comparison between MODIS and GPS PWV Sutcliffe efficiency (NSE) for all eight interpolation meth-
In order to compare the difference between GPS PWV ods are greater than 0, which indicates that the interpola-
and MODIS IR PWV, we calculate the correlation between tion results are plausible. And the NSE of IDW method
the two and use a scatter plot to present the relationship is 0.68, which is greater than the other methods and has
between the two, as shown in Fig. 4. According to the the highest credibility. As shown in Table 3, the MB of
Fig. 4, there is a good agreement between GPS PWV and IDW is
0.99 mm, and the MODIS IR PWV obtained
MODIS IR PWV during the study period in this study by IDW interpolation has a good agreement with GPS
area. However, the error dispersion degree is large, which PWV, with an R-squared of 0.84. Cubic, nearest, and
indicates that the relationship between the two is not a sim- TLI have a high MB of
2.15,
1.88, and
1.88 mm,
ple linear relationship. Therefore, it is necessary to study respectively. MODIS IR PWV obtained by these three
the calibration of MODIS IR PWV to improve its interpolation methods tends to underestimate atmospheric
accuracy. water vapor seriously. In addition, the RMSE, STD, MAE,
NSE, and interquartile range (IQR) ranges of MODIS IR
3.2. Regional MODIS IR PWV product spatial interpolation PWV obtained by eight interpolation methods are 7.83  1
3.78 mm, 7.41  16.64 mm, 6.26  10.16 mm, 0.31  0.68,
In this section, this study calibrates MODIS IR PWV and 20.81  23.43 mm, respectively. That’s because that
based on GPS PWV, the appropriate number of MODIS nearest only considers the impact of the nearest point on
pixels is selected to participate in the interpolation, and the interpolation point, with the largest error and the worst
eight interpolation methods are used to interpolate the correlation. The TLI data are uniformly distributed in the
water vapor values of GPS stations. study area without considering the water vapor difference
of stations in different regions. The weights of natural inter-
3.2.1. Comparison of interpolation methods polation points are affected by the average weights of adja-
Since GPS PWV and MODIS IR PWV are at different cent points. Cubic uses piecewise interpolation, which
points in space, it is necessary to interpolate MODIS IR cannot guarantee the smoothness of interpolation points
PWV and perform spatial registration for MODIS IR throughout the study area. BHS provides a highly smooth
PWV and GPS PWV. In this study, eight different interpo- fitted curve but can cause problems with excessive smooth-
lation methods are used to compare and analyze MODIS ing, resulting in extreme values. IDW takes into account
IR PWV. The comparison between MODIS IR PWV the variation in the distribution of water vapor and
obtained by different interpolation methods and GPS employs the distance as the weight between the sample
PWV is shown in Fig. 5, which contains all interpolation point and the interpolation point. Newton changes the
parameters from 1 to 22 selected. Parameter refers to the shape of the fitting polynomial by adding more sampling
number of selected MODIS pixel points participating in points, but the resulting fitting polynomial could oscillate
the interpolation. As can be seen from Fig. 5, each interpo- and be prone to extreme values. Kriging considers the dis-
lation method can choose different parameters to partici- tance and spatial relationship between the interpolation
pate in the interpolation, so the scatter plot density is point and the sample point, but this method requires much
calculation and can quickly generate extreme values. In
summary, the interpolation effect is ranked from best to
worst: IDW, Kriging, natural, TLI, cubic, nearest, BHS,
and Newton. The R-squared and NSE between MODIS
IR PWV and GPS PWV are relatively low, and the RMSE,
STD, MAE, and IQR are relatively high, mainly because
MODIS IR products are greatly affected by cloudy
conditions.

3.2.2. Select different parameters to participate in the

interpolation
The parameter in this study refers to the number of
selected MODIS pixel points participating in the interpola-
tion. In the process of interpolation, the accuracy and reli-
ability of interpolation can be improved using different
Fig. 4. Scatter density map for PWV from GPS and MODIS. The color parameters. If only the same parameter is used for interpo-
scale represents the relative density of points. lation, the influence of other factors on the data is often
4910
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

Fig. 5. Correlation between GPS PWV and MODIS IR PWV under different interpolation methods. Figures (a) to (h) are GPS PWV and MODIS IR
PWV scatter plots based on nearest, TLI, natural, cubic, BHS, IDW, Newton, and Kriging interpolation methods. The color scale represents the relative
density of points.

ignored, which leads to a significant deviation in the inter- MODIS pixel point and GPS station. Different interpola-
polation results. Using only one parameter in the interpo- tion methods can be used for each parameter, and the error
lation process may not accurately reflect the scenario statistics are also different. As can be seen from Table 4, the
since the data distribution and variation trend of the many RMSE and MB of MODIS IR PWV obtained by interpo-
parameters chosen to take part in the interpolation differ. lation of different parameters show a trend of first decreas-
Therefore, the minimum 1  22 parameters of the distance ing, then rising, and finally slowly decreasing. The STD
total are chosen after sorting the distance between each and MAE show decreasing and gradually stabilizing

4911
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

Table 3
Specific error statistics between MODIS IR PWV and GPS PWV.
Interpolation method R-squared NSE MB (mm) RMSE (mm) STD (mm) MRE (%) MAE (mm) IQR (mm)
nearest 0.79 0.53
1.88 9.50 14.78 24.90 7.70 21.00
TLI 0.82 0.61
1.88 8.65 14.22 23.44 6.92 21.19
natural 0.82 0.62
1.82 8.51 14.13 23.16 6.81 21.10
cubic 0.81 0.58
2.15 8.95 14.46 24.25 7.17 21.37
BHS 0.63 0.51
1.31 9.62 14.82 24.60 7.61 21.75
IDW 0.84 0.68
0.99 7.83 13.59 20.94 6.26 20.81
Newton 0.61 0.31
1.78 13.78 16.64 34.75 10.16 23.43
kriging 0.82 0.62
1.48 8.54 8.41 22.51 6.83 20.94

Table 4
Global error statistics of MODIS IR PWV are obtained by selecting all interpolation methods with different parameters.
Interpolation parameter R-squared NSE RMSE (mm) STD (mm) MB (mm) MRE (%) MAE (mm)
1 0.79 0.5262 9.50 14.78
1.88 24.90 7.70
2 0.79 0.5386 9.41 14.72
1.87 24.70 7.60
3 0.77 0.4991 9.78 14.80
0.87 24.98 7.82
4 0.77 0.4969 9.79 14.61
0.92 25.23 7.57
5 0.78 0.5266 9.50 14.50
1.39 24.68 7.36
6 0.78 0.5280 9.49 14.50
1.47 24.52 7.32
7 0.78 0.5269 9.51 14.52
1.64 24.58 7.32
8 0.79 0.5353 9.42 14.50
1.65 24.45 7.27
9 0.79 0.5367 9.39 14.50
1.67 24.33 7.25
10 0.79 0.5368 9.39 14.49
1.69 24.38 7.24
11 0.79 0.5324 9.44 14.52
1.71 24.41 7.26
12 0.79 0.5365 9.40 14.49
1.70 24.35 7.24
13 0.79 0.5354 9.41 14.48
1.70 24.40 7.24
14 0.79 0.5434 9.32 14.46
1.73 24.32 7.21
15 0.79 0.5563 9.39 14.48
1.71 24.40 7.23
16 0.79 0.5418 9.34 14.45
1.71 24.31 7.20
17 0.79 0.5428 9.33 14.46
1.73 24.28 7.20
18 0.79 0.5531 9.29 14.44
1.72 24.18 7.18
19 0.79 0.5459 9.31 14.45
1.71 24.22 7.20
20 0.79 0.5478 9.29 14.44
1.71 24.19 7.19
21 0.79 0.5456 9.31 14.45
1.69 24.19 7.19
22 0.79 0.5502 9.27 14.44
1.69 24.17 7.17

trends. The two errors are stable around 14.45 and 6 parameters. The RMSE is increasing, while the MB is
7.20 mm, respectively. The mean relative error (MRE) decreasing. When 14 parameters are selected, the error of
shows a slow decreasing trend. R-squared presents a slow BHS begins to stabilize. IDW can be interpolated with
growth trend. The NSE shows an overall increasing trend more than 3 parameters, and the RMSE and MB show a
with increasing parameters and is maximum at the fifteenth slow decreasing trend. Newton can use more than 2 param-
parameter. In summary, choosing 5  15 parameters to eters for interpolation. The RMSE and MB show an
participate in the interpolation is better, of which fifteen increasing trend. Moreover, the two errors’ variation trend
have the highest precision. is rough, sometimes increasing and sometimes decreasing.
As shown in Fig. 6, the MODIS IR PWV obtained by Kriging can use more than 3 parameters for interpolation.
nearest had a more significant RMSE. TLI can be interpo- The RMSE and MB show a decreasing trend, while the
lated with more than 4 parameters, and the RMSE of MB is unstable with positive and negative. When Kriging
MODIS IR PWV increases briefly and then becomes selects 15 parameters, the error begins to stabilize. There-
stable, the MB decreases slowly as the number of parame- fore, the interpolation accuracy of the eight interpolation
ters increases. The error of TLI begins to stabilize when 15 methods is sorted from best to worst as follows: IDW, krig-
parameters are selected. Natural can be interpolated with ing, natural, TLI, cubic, nearest, BHS, Newton.
more than 3 parameters, and the RMSE decreases with To sum up, as the number of parameters increases, the
the increase of the number of parameters. The MB complexity of interpolation also increases. Some interpola-
increases briefly and then slowly decreases. Parameter tion methods appear to have overfitting phenomena, such
selection, RMSE and MB error trends for cubic are the as BHS and Newton, and some interpolation methods have
same as for TLI. BHS can be interpolated with more than errors that gradually become stable as the number of

4912
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

Fig. 6. Error statistics of different parameters of MODIS IR PWV and GPS PWV. Figure (a) to (h) shows the error statistics of 1 to 22 parameters selected
by eight interpolation methods. The blue line represents the root mean square error (RMSE), and the orange line represents the mean bias (MB).

parameters increases. According to the experimental results and the R-squared of the FT model is the highest. From
of this study, the best effect occurs when 15 interpolation the RMSE perspective, the FT model has the best calibra-
parameters are selected, the precision of IDW is the high- tion effect, and the RMSE decreases by 13.6 % to 8.14 mm,
est. The MODIS IR PWV obtained by selecting different while the LF calibration effect is the worst, and the RMSE
parameters and interpolation methods is different, and even increases. From the perspective of STD, the EF model
there are some errors with GPS PWV, so it is necessary had the best calibration effect, with the STD reduced by
to choose a suitable calibration model to calibrate MODIS 28.2 % to 10.44 mm, and the STD change before and after
IR PWV. the MLF model revision is the least, with a decrease of
19.4 %. The EF model has the most minor standard devi-
3.3. Regional MODIS PWV product calibration ation, which means that the degree of difference between
the data points and the average is the least, which means
In this section, linear fitting and five nonlinear fitting that the degree of dispersion of the model is also the least.
calibration models are constructed. From the perspective of MB, the FT model has the best
calibration effect, with the MB decreasing by 64.3 % to
3.3.1. MODIS PWV calibration model 0.019 mm, while the LF model has the worst calibration
GPS PWV and MODIS IR PWV are not a simple linear effect, with the MB even increasing. From the perspective
relationship. By calibrating MODIS IR PWV using GPS of MRE, the FT model has the best calibration effect,
PWV, atmospheric water vapor may be estimated more and the MRE decreases by 8.4 %, while the EF model’s
accurately and more accurately reflects variations in water MRE increases. From the perspective of MAE, the FT
vapor across time and space. Fig. 7 shows the calibration model has the best calibration effect, and the MAE
results of LF, MLF, PIF, EF, GF, and FT models. The decreases by 12.8 % to 6.37 mm, while the MAE of the
MB of MODIS IR PWV of the six models except the LT LF model increases. In summary, the FT model has a good
model is significantly reduced, and the error range is
0. calibration effect in this region, which can further weaken
08  0.80 mm. In addition, the LT, MLF, GF, and FT the influence of cloud cover, precipitation, and other fac-
models are positive deviations, indicating that the MODIS tors on the reliability of MODIS IR PWV product data.
IR PWV product is slightly overestimated. The correlation The above error indicators can be used to see the general
coefficient between GF and FT models is higher than that error distribution of the results of the six models and the
of the other four models. The R-squared is above 0.8, the degree of concentration of errors at the baseline. The quan-
RMSE is below 8.2 mm, and the MB is below 0.025 mm. tity of outliers is a tangible indicator of the stability of the
model solution, but they are not sensitive to the outliers in
3.3.2. Model reliability analysis the mistakes and cannot be determined intuitively. There-
This study takes GPS PWV as the evaluation criterion, fore, we use the box chart to quantify the outliers. The
the comparison of error histograms before and after cali- box graph’s main benefit is that, although it can identify
bration is shown in Fig. 8. The R-squared of the LF model outliers in the data, it remains unaffected by outliers and
is lower than that of the original model, the R-squared of can provide a rather consistent description of the discrete
other models is higher than that of the original model, distribution of data. The box chart contains five statistical
4913
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

Fig. 7. Comparison between GPS PWV and MODIS IR PWV after calibration of six fitted models. The blue scatters represent the scatters of GPS PWV
and MODIS IR PWV, and the red curves represent the fit curves obtained by each fitting method. The color scale represents the relative density of points.

elements: lower boundary, lower quartile (Q1), median is the smallest, followed by the FT model. The IQR is a
(Q2), upper quartile (Q3), and upper boundary. The crite- quartile distance, and the smaller the value, the tighter
rion for determining outliers in the box chart is based on the data distribution. The FT model has a smaller IQR
the interquartile range (IQR), IQR = Q3–Q1. The position value, and the EF model has a more immense IQR value.
1.5  IQR higher than Q1 is the upper boundary, and the The box’s width reveals the box diagram’s flat degree,
position 1.5  IQR lower than Q3 is the lower boundary. and its relative narrowness indicates the data distribution
The data outside the upper and lower boundaries is the is more concentrated and the data change range is tiny.
outlier. The box diagram is narrow, the data are slightly concen-
As shown in Fig. 9, the box graphs corresponding to the trated, and the absolute values of the upper and lower
errors of the six models have their characteristics. The out- bounds of the FT model are minimal. The LF and EF
liers of the PIF and EF models are more extensive and models have wider boxes and more discrete data. The out-
more discrete, while the outliers of the GF and FT models liers reflect the robustness of the calibration model’s
are smaller and more concentrated. The box chart can results. The number of outliers in the FT model is the least,
roughly show the distribution of model result error, and and the box of the FT model is narrow, which indicates
the specific values of the statistical elements corresponding that the FT model is relatively stable and the errors are rel-
to the six box charts in the chart are shown in Table 5. Q2 atively concentrated. In addition, it can also be found from
reflects the data deviation to some extent. The deviation of the figure that the PIF and EF models have many outliers.
the LF model is more significant than that of other This phenomenon is because the two models have the prob-
schemes, and the absolute value of Q2 of the PIF model lem of underfitting.

4914
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

Fig. 8. Overall error histogram statistics of the six calibration models.

From the above analysis, it can be seen that the box

chart of the FT model is the most narrow, the outliers
are the least, and the absolute value of IQR and Q2 is
the smallest, which can significantly reduce the calibration
model error in order to obtain high-precision MODIS IR
PWV products.

3.4. Accuracy verification of calibration model

3.4.1. Accuracy verification of the combination of different

interpolation methods and calibration models
In order to further compare the calibration effects
of different models combined with different interpolation
methods, six calibration models are applied to MODIS
IR PWV products obtained by different interpolation
methods in 2022. The statistical error results of different
calibration models and interpolation methods are shown
in Table 6.
Fig. 9. Box diagram of six calibration models.

Table 5
Box graph statistics for six calibration models.
statistics
model Q1 Q2 Q3 IQR lower boundary upper boundary
LF
0.75 4.89 10.35 11.10
17.40 27.00
MLF
5.24 0.25 5.33 10.57
21.10 21.19
PIF
5.65
0.05 5.29 10.94 –22.06 21.70
EF
6.41
0.52 5.80 12.21
24.73 24.12
GF
5.25 0.24 5.35 10.60
21.15 21.25
FT
5.23 0.23 5.27 10.50
20.98 21.02

4915
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

Table 6
Specific error statistics of different combinations of calibration models and interpolation methods.
correction model interpolation method R-squared NSE RMSE (mm) STD (mm) MB (mm) MAE (mm)
Original model nearest 0.79 0.5262 9.50 14.78
1.88 7.70
TLI 0.82 0.6063 8.65 14.22
1.88 6.92
natural 0.82 0.6184 8.51 14.13
1.82 6.81
cubic 0.81 0.5781 8.95 14.46
2.15 7.17
BHS 0.63 0.5144 9.62 14.82
1.31 7.61
IDW 0.84 0.6823 7.83 13.59
0.99 6.26
Newton 0.61 0.3730 13.78 16.64
1.78 10.16
Kriging 0.82 0.6190 8.54 14.10
1.48 6.83
LF model nearest 0.79 0.7509 9.75 11.67 4.77 7.73
TLI 0.82 0.7566 9.18 11.23 4.66 7.40
natural 0.82 0.7566 9.14 11.16 4.71 7.37
cubic 0.81 0.7576 9.19 11.42 4.44 7.39
BHS 0.63 0.7422 10.11 11.70 5.19 8.05
IDW 0.84 0.7649 9.34 10.73 5.44 7.56
Newton 0.61 0.7395 12.73 13.14 4.55 9.57
Kriging 0.82 0.7518 9.43 11.13 5.08 7.59
MLF model nearest 0.80 0.9583 8.25 11.66
0.05 6.55
TLI 0.83 0.9618 7.64 11.41
0.28 6.11
natural 0.84 0.9618 7.57 11.38
0.26 6.05
cubic 0.83 0.9618 7.80 11.51
0.46 6.23
BHS 0.79 0.9490 8.44 11.86 0.41 6.66
IDW 0.85 0.9688 7.27 11.15 0.37 5.85
Newton 0.61 0.9353 12.10 13.52 0.35 8.80
Kriging 0.83 0.9583 7.64 11.33 0.10 6.12
PIF model nearest 0.79 0.9488 8.40 11.45
0.20 6.64
TLI 0.82 0.9510 7.83 11.05
0.33 6.27
natural 0.83 0.9510 7.75 10.98
0.29 6.21
cubic 0.82 0.9512 7.97 11.22
0.55 6.37
BHS 0.79 0.9492 8.57 11.52 0.23 6.77
IDW 0.85 0.9593 7.52 10.61 0.41 6.04
Newton 0.61 0.9339 11.89 13.06
0.27 8.77
Kriging 0.82 0.9488 7.86 10.95 0.07 6.28
EF model nearest 0.76 0.8609 9.04 10.80
0.38 7.18
TLI 0.79 0.8603 8.56 10.20
0.43 6.89
natural 0.79 0.8603 8.49 10.10
0.37 6.83
cubic 0.78 0.8605 8.68 10.46
0.68 6.98
BHS 0.75 0.8526 9.19 10.70
0.04 7.32
IDW 0.81 0.8699 8.35 9.52 0.41 6.74
Newton 0.58 0.8493 11.97 12.15
1.16 9.00
Kriging 0.78 0.8630 8.62 10.10 0.02 6.92
GF model nearest 0.80 0.9708 8.23 11.58 0.00 6.58
TLI 0.84 0.9737 7.56 11.47
0.27 6.06
natural 0.84 0.9737 7.48 11.45
0.26 6.00
cubic 0.83 0.9736 7.73 11.52
0.43 6.20
BHS 0.80 0.9703 8.37 11.77 0.42 6.62
IDW 0.86 0.9751 7.16 11.32 0.33 5.77
Newton 0.63 0.9663 11.40 12.63 0.23 8.44
Kriging 0.84 0.9708 7.57 11.41 0.10 6.06
FT model nearest 0.80 0.9693 8.23 11.59 0.00 6.57
TLI 0.84 0.9721 7.55 11.48
0.27 6.06
natural 0.84 0.9721 7.47 11.46
0.26 6.00
cubic 0.83 0.9720 7.72 11.53
0.43 6.19
BHS 0.80 0.9698 8.36 11.76 0.42 6.61
IDW 0.86 0.9756 7.16 11.34 0.33 5.76
Newton 0.64 0.9545 11.29 12.50 0.16 8.38
Kriging 0.84 0.9693 7.56 11.42 0.10 6.05

4916
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

It can be seen from Table 6 that the changes in the RMSE, STD, and MAE of IDW are the smallest,
R-squared and MAE after combining all interpolation decreasing by 8.49 %, 16.69 %, 7.91 % to 7.16, 11.32, and
methods with the LF model are small, and the NSE is 5.77 mm, respectively. The nearest has an MB of almost
much smaller than the other models, which indicates that 0 mm, and the correlation of IDW can reach 0.86. The
the combination effect of all interpolation methods and FT model has the largest increase in correlations and the
the LF model is deplorable. In the MLF model, the largest reduction in all errors. In the FT model, the RMSE,
R-squared and NSE are enhanced, and all errors are STD, and MAE of IDW are the smallest, decreasing by
reduced, indicating that the MLF model has a better effect. 8.60 %, 16.59 %, and 8.01 % to 7.16, 11.34, and
At the same time, the RMSE, STD, and MAE of IDW are 5.76 mm, respectively. The nearest MB is almost 0 mm,
the smallest, decreasing by 7.16 %, 18 %, and 6.59 % to the correlation of IDW reaches 0.86 and the NSE can reach
7.27, 11.15, and 5.85 mm, respectively. The MB of nearest 0.9756. In summary, the FT model has the highest accu-
is the smallest, decreasing by 97.26 % to
0.05 mm. The racy, and the accuracy and correlation of each interpola-
correlation of IDW is also the highest, the R-squared is tion method are improved, among which IDW has the
0.85 and the NSE is 0.9688. However, the correlation of highest accuracy and Newton has the lowest.
the PIF model does not increase much, NSE is even a little
and the reduction of the errors is relatively small, indicat- 3.4.2. Accuracy verification of optimal calibration model
ing that the effect of the PIF model does not have a better In this study, fifteen interpolation parameters are
effect than the MLF model. In the PIF model, the RMSE, selected as the optimal parameters, IDW is used as the
STD, and MAE of IDW are the smallest, decreasing by optimal interpolation method to obtain MODIS IR
3.92 %, 21.97 %, 6.59 % to 7.52, 10.61, and 5.85 mm, PWV, and the optimal calibration model, FT model, is
respectively, while the MB of Kriging is the smallest, combined to calibrate MODIS IR PWV. This calibration
decreasing by 94.99 % to 0.07 mm. The EF model is not method is called FIDWFT. We compare the MODIS IR
very stable, the R-squared and the NSE decrease, and the PWV before and after calibration with the corresponding
changes of RMSE and MAE are bumpy and unstable, GPS PWV, and the MB generated by both is shown in
which indicates that the EF model has a poor effect, but Fig. 10. It can be seen from Fig. 10 that the FIDWFT cal-
the STD decreases considerably, and the degree of data dis- ibrated model has a higher agreement with the GPS PWV
persion is weakened. In the GF model, the correlation of model, and the MB error is smaller. While the MODIS IR
the GF model is enhanced, and all the error reduction is PWV model before calibration has a significant error and
greater than that of the MLF model. At the same time, the agreement with the GPS model is low. In other words,
the accuracy of the FIDWFT model in this study is greatly
improved compared with that of the original MODIS PWV
model, and it is in good agreement with the GPS PWV
model.
Table 7 collects error data for MODIS IR PWV and
corresponding GPS PWV before and after calibration in
2022. The R-squared of the FIDWFT model improved sig-
nificantly, from 0.86 to 0.94, indicating that the PWV of
the new model is in good agreement with the GPS PWV.
The NSE improves from 0.72 to 0.87, the confidence of
the model increases and the fit to the observations
improves. The MB decreases from
1.89 to
0.83 mm, a
relative increase of 55.9 %. Comparing the RMSE of the
two models, it can be found that the RMSE decreases sig-
nificantly, from 7.32 to 4.98 mm, with a relative increase of
31.9 %. The STD, MAE, and MRE increase by 10.4 %,
36.6 %, and 29.7 %, respectively. These improvements
Fig. 10. Mean bias (MB) comparison of MODIS IR PWV and GPS PWV
show that the calibrated model is promising.
models before and after calibration. The before calibration PWV model is As MODIS products are affected by cloud cover, cloud
the MODIS IR PWV obtained using only the nearest interpolation. The cover can lead to poorer quality of MODIS IR PWV data,
after calibration PWV model refers to the FIDWFT model. with RMSE larger than 10 mm (Jia and Liu, 2019). The

Table 7
Error statistics between the FIDWFT model and the original model.
R-squared NSE MB (mm) RMSE (mm) STD (mm) MAE (mm) MRE (%)
Before calibration PWV 0.86 0.72
1.89 7.32 13.58 5.83 20.24
After calibration PWV 0.94 0.87
0.83 4.98 12.17 3.70 14.23

4917
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

Fig. 11. MODIS IR PWV distribution map before and after calibration in Hong Kong at 6:00 UTC, August 1, 2022. The color scale represents the
frequency of PWV values (Blue represents the highest frequency and red represents the lowest frequency).

accuracy of MODIS IR PWV products under clear-sky signals in the time domain into signals in the frequency
conditions is higher than that under cloudy conditions, domain, which facilitates the analysis and processing of
and the RMSE under clear-sky conditions is generally the frequency characteristics of the signals. Meanwhile,
4  6 mm (Bai et al., 2021; Sam et al., 2019). By using PWV has inter-annual or inter-decadal cycles and short-
our FIDWFT model, the MODIS IR PWV product can term cyclic variations. Maghrabi (2019) analyzed the vari-
achieve an average RMSE of 4.98 mm, obtaining higher ations of PWV in the Arabian Peninsula under 1-year, 6-
correlation, smaller RMSE, and lower MB relative to the month, 4-month, and 3-month cycles for the years
reference PWV data. Due to the effect of clouds on the 1986–––2016 by using the Fast Fourier Transform (FFT).
MODIS IR channel, the MODIS IR PWV product does Therefore, these two methods are plausible in the case
not provide good quality water vapor measurements in where the experimental area is a uniform distribution of
cloudy conditions, which can result in RMSEs in excess stations and the calibrated PWV period is one year. In this
of 4.98 mm in cloudy conditions, but the FIDWFT model experimental region, we believe that the model is equally
is still expected to be an effective tool for improving the applicable even if the experimental time period changes.
accuracy of the MODIS IR PWV product.
4. Conclusions
3.4.3. MODIS water vapor calibration for Hong Kong based
on GPS Water vapor in the atmosphere has a thermal feedback
MODIS IR PWV has the feature of planar continuous effect and is a greenhouse gas. Weather change is primarily
spatial distribution. The accuracy of the PWV distribution caused by water vapor as well. Currently, cloud cover,
map can be improved by using GPS PWV calibration of atmospheric conditions, and satellite sensor performance
point data. In this study, MODIS IR PWV in Hong Kong all have an impact on MODIS IR PWV accuracy, which
is calibrated based on the FIDWFT model, and Fig. 11 is not particularly high. However, MODIS can provide
shows the MODIS IR PWV distribution map before and water vapor content information globally with a particular
after the calibration in Hong Kong at 6:00 UTC, August spatial resolution. Therefore, the main requirements are the
1, 2022. evaluation and calibration of MODIS water vapor. This
As can be seen from Fig. 11, the PWV distribution map study proposes a MODIS water vapor calibration model
that has been calibrated has a truly smooth surface, which fusing inverse distance weighting and the Fourier trans-
can lessen the influence of local data and more accurately form (FIDWFT) to improve the accuracy of MODIS IR
reflect the distribution of water vapor in the area. The PWV. Our model shows two advantages: (1) improving
PWV distribution map that has not been calibrated creates the accuracy of MODIS IR PWV; (2) GPS PWV can be
a ‘‘bull’s eye” phenomenon near GPS stations, which acquired in real-time, while MODIS IR PWV has a specific
overly depends on local data. The low accuracy of MODIS time delay. By combining them, PWV data with real-time,
IR PWV can be successfully improved, the geographical high precision, and large range can be obtained.
resolution of the calibration region can be improved, and This study used one year of PWV data from GPS sta-
the dependability of PWV data can be increased by cali- tions in Hong Kong to evaluate the accuracy of MODIS
brating MODIS IR PWV using the IDWFT model. IR water vapor data. First, the PWV obtained by the
IDW interpolation tends to work better with a relatively radiosonde (HKKP) is used to verify the GPS PWV. The
uniform distribution of stations. The FT model can convert R-squared between the two is 0.94, showing a good
4918
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

agreement. In addition, GPS PWV is used as the reference Declaration of Competing Interest
water vapor value to evaluate MODIS IR PWV. The R-
squared between the two is 0.81, the MB is
1.60 mm, The authors declare that they have no known competing
and the RMSE is 8.72 mm. Then, we built the FIDWFT financial interests or personal relationships that could have
model. The new model first considers the interpolation appeared to influence the work reported in this paper.
problem, including the number of interpolation points
and methods. Using different interpolation methods to
Acknowledgment
interpolate MODIS IR PWV, it is found that IDW has
the highest accuracy, the RMSE is 7.83 mm, the MB is
This research is supported by the National Natural

0.99 mm, the NSE is 0.68, and the R-squared is 0.84.
Science Foundation of China (42061076), and LZJTU
Then, a different number of MODIS pixel points are
EP201806. The authors thank the Survey and Mapping
selected to participate in the interpolation, and it is found
Office of the Hong Kong Lands Department for providing
that the interpolation effect is best when 15 parameters
GPS data from the Hong Kong Satellite Positioning Refer-
are selected. Then, the new model considers the calibration
ence Station Network (SatRef). We also thank the Mas-
problem. The FT model has the best calibration effect, the
sachusetts Institute of Technology (MIT) for GAMIT/
RMSE is reduced by 13.6 % to 8.14 mm, and the MB is
GLOBK software, the University of Wyoming for radio-
reduced by 64.3 % to
0.019 mm.
sonde data, and NASA for MODIS MOD_05 and
The new model can improve the accuracy of MODIS IR
MYD_05 products. The author is very grateful to the edi-
PWV. The MODIS IR PWV before calibration has an R-
torial and review team for their valuable comments, which
squared of 0.86 with an RMSE of 7.32 mm and an MB
significantly improved the quality of the paper. In addition,
of
1.89 mm relative to GPS PWV. After calibrating
this study uses the RINEX format of the GPS data, which
PWV with the new model, the R-squared of MODIS IR
can be downloaded from the website (https://www.geo-
PWV is increased to 0.94, the RMSE is reduced by
detic.gov.hk/sc/rinex/DOWNV.ASPX). MODIS water
31.9 % to 4.98 mm, and the MB is reduced by 56.08 % to
vapor products can be downloaded from the website

0.83 mm. It indicates that the MODIS IR PWV of the
(https://ladsweb.modaps.eosdis.nasa.gov).
new model exhibits better precision and accuracy. In addi-
tion, the new model’s PWV distribution map is smoother
and more accurate, which effectively addresses the low References
accuracy of MODIS IR PWV, enhances the calibration
region’s spatial resolution, and more properly depicts the Alexandrov, M.D., Schmid, B., Turner, D.D., Cairns, B., Oinas, V., Lacis,
A.A., Gutman, S.I., Westwater, E.R., Smirnov, A., Eilers, J., 2009.
distribution of water vapor in the area.
Columnar water vapor retrievals from multifilter rotating shadowband
In order for the MODIS IR PWV to completely coincide radiometer data. J. Geophys. Res. 114.
with the GPS PWV at the spatial point location, we first Bai, J., Lou, Y., Zhang, W., Zhou, Y., Zhang, Z., Shi, C., 2021.
address the spatial interpolation problem in this study. Assessment and calibration of MODIS precipitable water vapor
We employ eight different interpolation techniques, arrang- products based on GPS network over China. Atmos. Res. 254, 105504.
Bevis, M., Businger, S., Herring, T.A., Rocken, C., Anthes, R.A., Ware,
ing them according to the distance priority. Then the water
R.H., 1992. GPS meteorology: Remote sensing of atmospheric water
vapor calibration problem is considered and six models are vapor using the global positioning system. J. Geophys. Res. 97 (D14),
constructed to calibrate the MODIS IR PWV. Finally, it is 15787. https://doi.org/10.1029/92jd01517.
discovered that the MODIS water vapor calibration model Bevis, M., Businger, S., Chiswell, S., Herring, T.A., Anthes, R.A.,
with IDW and FT has the best accuracy. However, the FT Rocken, C., Ware, R.H., 1994. GPS meteorology: mapping zenith wet
delays onto precipitable water. J. Appl. Meteorol. 33, 379–386.
model does not include an error correction term and is only
Bolton, D., 1980. The computation of equivalent potential temperature.
applicable in the MODIS water vapor calibration process Mon. Weather Rev. 108, 1046–1053.
when the PWV is less than 90 mm. As a result, the model’s Champollion, C., Masson, F., Bouin, M.N., Walpersdorf, A., Doerflinger,
applicability is limited, and additional research on the E., Bock, O., Van Baelen, J., 2005. GPS water vapour tomography:
model is required in future studies. Preliminary results from the ESCOMPTE field experiment. Atmos.
Res. 74, 253–274.
Chang, L., Gao, G., Jin, S., He, X., Xiao, R., Guo, L., 2014. Calibration
CRediT authorship contribution statement and evaluation of precipitable water vapor from MODIS infrared
observations at night. IEEE Trans. Geosci. Remote Sens. 53.
Xiangrong Yan: Conceptualization, Data curation, Chang, L., Xiao, R., Prasad, A., Gao, G., Zhang, Y., 2019. Cloud mask-
Writing – original draft, Writing – review & editing, Visu- related differential linear adjustment model for MODIS infrared water
vapor product. Remote Sens. Environ. 221, 650–664.
alization, Investigation, Validation, Formal analysis,
Chen, S.H., Zhao, Z., Haase, J., Chen, A.D., Vandenberghe, F., 2008. A
Methodology, Supervision, Resources, Project administra- study of the characteristics and assimilation of retrieved MODIS Total
tion and Software. Weifang Yang: Funding acquisition, precipitable water data in severe weather simulations. Mon. Weather
Writing – review & editing, Resources, Project administra- Rev. 136, 3608–3628.
tion and Software. Nan Ding: Data curation, Writing – Ding, N., 2018. Research on key technologies of GNSS water vapor
tomography for foundation. China University of Mining and
review & editing. Fenglin Gao: Investigation. Yibo Peng:
Technology.
Investigation.
4919
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

Elgered, G., Johansson, J.M., Rönnäng, B.O., Davis, J.L., 1997. solar activity: Arabian Peninsula. J. Atmos. Sol. Terr. Phys. 185.
Measuring regional atmospheric water vapor using the swedish https://doi.org/10.1016/j.jastp.2019.01.011.
permanent GPS network. Geophysical Research Letters - GEOPHYS McAtee, B.K., Maier, S.W., 2006. A comparison of algorithms for near-
RES LETT. 24, 2663–2666. real time water vapour retrieval from MODIS. Int. J. Remote Sens. 27,
Gao, B.C., Kaufman, Y.J., 2003. Water vapor retrievals using moderate 5145–5163. https://doi.org/10.1080/01431160600707383.
resolution imaging spectroradiometer (MODIS) near-infrared chan- Merrikhpour, M.H., Rahimzadegan, M., 2017. Improving the algorithm
nels. J. Geophys. Res. Atmos. 108 (D13). of extracting regional Total precipitable water vapor over land from
Gao, B.C., Yang, P., Guo, G., Park, S.K., Wiscombe, W.J., Chen, B., MODIS images. IEEE Trans. Geosci. Remote Sens., 1–10
2003. Measurements of water vapor and high clouds over the tibetan Murray, F., 1967. On the computation of saturation vapor pressure. J.
plateau with the Terra MODIS instrument. IEEE Trans. Geosci. Appl. Meteorol. 6, 203–204.
Remote Sens. 41, 895–900. Niell, A.E., Coster, A.J., Solheim, F.S., Mendes, V.B., Toor, P.C.,
Ghaderi, M., Rahimzadegan, M., 2022. Improving precipitable water Langley, R.B., Upham, C.A., 2001. Comparison of measurements of
vapor estimations of the moderate resolution imaging spectroradiome- atmospheric wet delay by radiosonde, water vapor radiometer, GPS,
ter (MODIS) using metaheuristic algorithms. Adv. Space Res. 69. and VLBI. Journal of Atmospheric and Oceanic Technology - J
Goff, J.A., 1957. Saturation pressure of water on the new kelvin ATMOS OCEAN TECHNOL. 18.
temperature scale. Transactions of the American Society of Heating Oliveira, G., Brunsell, N., El, M., Shimabukuro, Y., Bertani, G., dos, S.T.,
and Ventilating Engineers. 63, 347–354. Aragão, L., 2017. Evaluation of MODIS-based estimates of water-use
Gurbuz, G., Jin, S., 2017. Long-time variations of precipitable water efficiency in Amazonia. Int. J. Remote Sens. 38, 5291–5309. https://
vapour estimated from GPS, MODIS and radiosonde observations in doi.org/10.1080/01431161.2017.1339924.
Turkey. Int. J. Climatol. 37 (15), 5170–5180. Pérez-Ramı́rez, D., Whiteman, D.N., Smirnov, A., Lyamani, H., Holben,
Hu, B., Wang, H., Hou, J., Jiang, L., Shen, Q., Liu, L., 2013. Correction B.N., Pinker, R., Andrade, M., Alados-Arboledas, L., 2014. Evalua-
of InSAR atmosphere effect based on MODIS water vapor data. tion of AERONET precipitable water vapor versus microwave
Journal of Geodesy and Geodynamics 33 (02), 82–86. radiometry, GPS, and radiosondes at ARM sites. J. Geophys. Res.
Iwabuchi, T., Naito, I., Mannoji, N., 2000. A comparison of global Atmos. 119. https://doi.org/10.1002/2014JD021730.
positioning system retrieved precipitable water vapor with the numerical Prasad, A.K., Singh, R.P., 2009. Validation of MODIS Terra, AIRS,
weather prediction analysis data over the Japanese Islands. J. Geophys. NCEP/DOE AMIP-II reanalysis-2, and AERONET sun photometer
Res. 105, 4573–4585. https://doi.org/10.1029/1999JD901007. derived integrated precipitable water vapor using ground-based GPS
Jia, H., Liu, Z., 2019. Comparison of satellite-derived precipitable water receivers over India. J. Geophys. Res.-Atmos. 114. https://doi.org/
vapor through near-infrared remote sensing channels. IEEE Trans. 10.1029/2008JD011230.
Geosci. Remote Sens., 1–11 https://doi.org/10.1109/TGRS.2019. Sam, A., Nikraftar, Z., Zakeri, S., 2019. Evaluation of MODIS near-IR
2932847. water vapor product over Iran using ground-based GPS measure-
Kaufman, Y.J., Gao, B.C., 1992. Remote sensing of water vapor in the ments. Atmos. Res. 231, 104657.
near IR from EOS/MODIS. IEEE Trans. Geosci. Remote Sens. 30 (5), Sánchez, J.L., Posada, R., Garcı́a-Ortega, E., López, L., Marcos, J.L.,
871–884. 2013. A method to improve the accuracy of continuous measuring of
Khaniani, A.S., Nikraftar, Z., Zakeri, S., 2019. Evaluation of MODIS vertical profiles of temperature and water vapor density by means of a
near-IR water vapor product over Iran using ground-based GPS ground-based microwave radiometer. Atmos. Res. 122, 43–54.
measurements. Atmos. Res. 104657. Sharifi, M.A., Sam Khaniani, A., Joghataei, M., 2015. Comparison of
Kwon, H.T., Iwabuchi, T., Lim, G.H., 2007. Comparison of precipitable GPS precipitable water vapor and meteorological parameters during
water derived from ground-based GPS measurements with radiosonde rainfalls in Tehran. Meteorol. Atmos. Phys. 127. https://doi.org/
observations over the Korean peninsula. Journal of the Meteorological 10.1007/s00703-015-0383-3.
Society of Japan - J METEOROL SOC JPN. 85, 733–746. https://doi. Solomon, S., Rosenlof, K.H., Portmann, R.W., Daniel, J.S., Davis, S.M.,
org/10.2151/jmsj.85.733. Sanford, T.J., Plattner, G.K., 2010. Contributions of stratospheric
Li, J., Wetzel, S., Menzel, W., Gumley, L., Seemann, S., 2003. Operational water vapor to decadal changes in the rate of global warming. Science
retrieval of atmospheric temperature, moisture, and ozone from 327, 1219–1223.
MODIS infrared radiances. J. Appl. Meteorol. 42 (8), 1072–1091. Turner, D.D., Lesht, B.M., Clough, S.A., Liljegren, J.C., Revercomb, H.E.,
Liou, Y.A., Teng, Y., VanHove, T., Liljegren, J., 2001. Comparison of Tobin, D.C., 2003. Dry bias and variability in vaisala RS80-H
precipitable water observations in the near tropics by GPS, microwave radiosondes: the arm experience. J. Atmos. Ocean. Technol. 20, 117–132.
radiometer, and radiosondes. J. Appl. Meteorol. Climatol. 40, 5–15. Wang, M., Fang, X., Hu, S., Hu, H., Li, T., Dou, X., 2015. Variation
https://doi.org/10.1175/1520-0450(2001)0402.0.CO. characteristics of water vapor distribution during 2000–2008 over
Liu, J., Liang, H., Sun, Z., Zhou, X., 2006. Validation of the moderate- Hefei (31.9°N, 117.2°E) observed by L625 Lidar. Atmos. Res. 164.
resolution imaging spectroradiometer precipitable water vapor pro- https://doi.org/10.1016/j.atmosres.2015.04.003.
duct using measurements from GPS on the tibetan plateau. J. Wang, Y., Liu, Y., Liu, L.i., Guo, Z., Ge, X., Xu, H., 2009. Retrieval of
Geophys. Res. 111. https://doi.org/10.1029/2005JD007028. the change of precipitable water vapor with zenith tropospheric delay
Liu, H., Tang, S., Zhang, S., Hu, J., 2015. Evaluation of MODIS water in the Chinese mainland. Advances in Space Research - ADV SPACE
vapour products over China using radiosonde data. Int. J. Remote RES. 43, 82–88. https://doi.org/10.1016/j.asr.2007.07.050.
Sens. 36 (2), 680–690. Wang, Y., Yang, K., Pan, Z.Y., Qin, J., Chen, D.L., Lin, C.G., Chen, Y.
Liu, B., Wang, Y., Lou, Z., Lou, Z., Zhan, W., 2019. The MODIS PWV Y., La, Z., Tang, W.J., Han, M.L., Lu, N., Wu, H., 2017. Evaluation
correction based on CMONOC in Chinese mainland. Acta Geodaetica of precipitable water vapor from four satellite products and four
Et Cartographica Sinica 48 (10), 1207–1215. reanalysis datasets against GPS measurements on the Southern
Liu, X., Wang, Y., Zhan, W., Yu, T.L., 2023. Improving MODIS Tibetan Plateau. J. Clim. 30, 5699–5713. https://doi.org/10.1175/
precipitable water vapour in mainland China based on the LSF model. JCLI-D-16-0630.1.
Adv. Space Res. 72. https://doi.org/10.1016/j.asr.2023.06.041. Wang, X., Zhang, K., Wu, S., Fan, S., Cheng, Y., 2016. Water-vapor-
Liu, Z., Wong, M.S., Nichol, J., Chan, P.W., 2013. A multi-sensor study weighted mean temperature and its impact on the determination of
of water vapour from radiosonde, MODIS and AERONET: a case precipitable water vapor and its linear trend. J. Geophys. Res. Atmos.
study of Hong Kong. Int. J. Climatol. 33, 109–120. https://doi.org/ 121. https://doi.org/10.1002/2015JD024181.
10.1002/joc.3412. Ware, R., Rocken, C., Solheim, F., Van Hove, T., Alber, C., Johnson, J.,
Maghrabi, A., 2019. Multi- decadal variations and periodicities of the 1993. Pointed water vapor radiometer corrections for accurate global
precipitable water vapour (PWV) and their possible association with positioning system surveying. Geophys. Res. Lett. 20 (23), 2635–2638.

4920
X. Yan et al. Advances in Space Research 73 (2024) 4903–4921

Xu, J., Liu, Z., 2022. A linear regression of differential PWV calibration Yulianto, E., Abdullah, F., Novrizal, N., Riyanto, A., 2020. Pseudo 3D
model to improve the accuracy of MODIS NIR all-weather PWV seismic using biharmonic spline interpolation. IOP Conference Series:
products based on ground-based GPS PWV data. IEEE J. Sel. Top. Earth and Environmental Science. 481, 012054.
Appl. Earth Obs. Remote Sens., 1–24 Zhang, B., Yao, Y., Xu, C., 2015. Global empirical model for estimating
Yan, X., Wang, X., Liang, W., Mu, Y., Song, Y., Ding, N., Zhang, W., water vapor seale height. Acta Geodaeti-Ca Et Cartographica Sinica
2022. An improved IDW interpolation method based on GNSS water 44 (10), 1085–1091.
vapor distribution and interstation distance. Journal of Nanjing Zhang, W., Zhang, S., Zheng, N., Ding, N., Xin, L., 2021. A new
University of Information Science & Technology (Natural Science integrated method of GNSS and MODIS measurements for tropo-
Edition) 14 (03), 287–293. spheric water vapor tomography. GPS Solutions 25 (2).
Yang, J., Qiu, J., 2002. A method for estimating precipitable water and Zhao, Q., Su, J., Xu, C., Yao, Y., Zhang, X., Wu, J., 2023. High-precision
effective water vapor content from ground humidity parameters. Chin. ZTD model of altitude-related correction. IEEE J. Sel. Top. Appl.
J. Atmos. Sci. 01, 9–22. Earth Obs. Remote Sens. 16, 609–621. https://doi.org/10.1109/
Yao, Y., Shan, L., Zhao, Q., 2017. Establishing a method of short-term JSTARS.2022.3228917.
rainfall forecasting based on GNSS-derived PWV and its application. Zhao, Q., Wang, W., Li, Z., Du, Z., Yang, P., Yao, W., Yao, Y., 2024. A
Sci. Rep. 7 (1), 12465. high-precision ZTD interpolation method considering large area and
height differences. GPS Solutions 28, 4.

4921