remote sensing

Article
Multi-Source Precipitation Data Merging for Heavy Rainfall
Events Based on Cokriging and Machine Learning Methods
Junmin Zhang 1,2 , Jianhui Xu 2,3, *, Xiaoai Dai 1 , Huihua Ruan 4 , Xulong Liu 2,3 and Wenlong Jing 2,3

1 College of Earth Science, Chengdu University of Technology, Chengdu 610059, China;

[email protected] (J.Z.); [email protected] (X.D.)
2 Guangdong Province Engineering Laboratory for Geographic Spatio-Temporal Big Data, Key Laboratory of
Guangdong for Utilization of Remote Sensing and Geographical Information System, Guangdong Open
Laboratory of Geospatial Information Technology and Application, Guangzhou Institute of Geography,
Guangdong Academy of Sciences, Guangzhou 510070, China; [email protected] (X.L.);
[email protected] (W.J.)
3 Southern Marine Science and Engineering Guangdong Laboratory (Guangzhou), Guangzhou 511458, China
4 Guangdong Meteorological Observation Data Center, Guangzhou 510080, China;
[email protected]
* Correspondence: [email protected]

Abstract: Gridded precipitation data with a high spatiotemporal resolution are of great importance
for studies in hydrology, meteorology, and agronomy. Observational data from meteorological
stations cannot accurately reflect the spatiotemporal distribution and variations of precipitation over
a large area. Meanwhile, radar-derived precipitation data are restricted by low accuracy in areas of
complex terrain and satellite-based precipitation data by low spatial resolution. Therefore, hourly
precipitation models were employed to merge data from meteorological stations, Radar, and satellites;



the models used five machine learning algorithms (XGBoost, gradient boosting decision tree, random


forests (RF), LightGBM, and multiple linear regression (MLR)), as well as the CoKriging method. In
Citation: Zhang, J.; Xu, J.; Dai, X.; the north of Guangdong Province, data of four heavy rainfall events in 2018 were processed with
Ruan, H.; Liu, X.; Jing, W.
geographic data to obtain merged hourly precipitation data. The CoKriging method secured the best
Multi-Source Precipitation Data
prediction of spatial distribution of accumulated precipitation, followed by the tree-based machine
Merging for Heavy Rainfall Events
learning (ML) algorithms, and significantly, the prediction of MLR deviated from the actual pattern.
Based on Cokriging and Machine
Learning Methods. Remote Sens. 2022,
All machine learning methods showed poor performances for timepoints with little precipitation
14, 1750. https://doi.org/10.3390/ during the heavy rainfall events. The tree-based ML method showed poor performance at some
rs14071750 timepoints when precipitation was over-related to latitude, longitude, and distance from the coast.

Academic Editor: Elisa Palazzi

Keywords: heavy rainfall events; data merging; CoKriging; machine learning; multi-source precipitation
Received: 25 February 2022
Accepted: 2 April 2022
Published: 6 April 2022
1. Introduction
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in Precipitation is a key meteorological parameter that influences the global water cycle
published maps and institutional affil- and surface environmental conditions. It is a crucial component of the water cycle and
iations. an avenue of energy exchange in the climate system [1,2], making it an essential indicator
for characterizing climate change. Gridded precipitation estimates with a high spatial
resolution are crucial for scientific research in various fields (e.g., hydrology, meteorology,
climatology, and agronomy). Ground-based meteorological measurements are the most
Copyright: © 2022 by the authors. direct method of acquiring precipitation data, providing the highest single-point accuracy
Licensee MDPI, Basel, Switzerland. over relatively long periods [3,4]. However, this method is limited by the density and spatial
This article is an open access article
distribution of observation stations, making it difficult to accurately capture spatiotemporal
distributed under the terms and
distribution and variations [5–7]. Precipitation data with a high spatiotemporal resolution
conditions of the Creative Commons
can be obtained through ground-based radar observations, but data accuracy is easily
Attribution (CC BY) license (https://
affected by complex terrain [8]. Satellite-based precipitation measurement techniques
creativecommons.org/licenses/by/
have developed from infrared/visible light sensing to passive and active microwave
4.0/).

Remote Sens. 2022, 14, 1750. https://doi.org/10.3390/rs14071750 https://www.mdpi.com/journal/remotesensing

Remote Sens. 2022, 14, 1750 2 of 23

sensing, with the most recent satellite-based efforts attempting to integrate the advantages
of infrared and microwave sensing [9–11]. These merging projects include pioneering
efforts such as the CPC Merged Analysis of Precipitation (CMAP), the Global Precipitation
Climatology Project (GPCP) [12], and the Tropical Rainfall Measuring Mission (TRMM),
operating since 1997. The current standard for such projects is the Global Precipitation
Measurement (GPM) mission, which forms the basis for two widely used data products:
Integrated Multi-satellite Retrievals for GPM (IMERG) and Global Satellite Mapping of
Precipitation (GSMaP). IMERG can provide good half-hourly precipitation estimates with
a spatial resolution of 0.1◦ × 0.1◦ , but some studies of extreme weather have found this
resolution too coarse [13]. GSMaP data outperform IMERG for extreme precipitation
events [14], but they are likewise limited by low resolution. There may be a significant
systematic deviation between individual satellite-derived and radar-derived precipitation
data, and the utility of satellite and radar data can be greatly enhanced by merging them
with station data for correction and calibration [15,16].
In recent years, methods have been developed to merge precipitation data from differ-
ent sources in order to improve spatiotemporal resolution and combine the advantages of
the different sources [17]. Various methods have been introduced for merging precipitation
data from meteorological stations and satellites, including artificial neural networks [18,19],
optimal interpolation [20], the Filtersim multiple-point statistics method [21], convolutional
neural network–long short-term memory (CNNLSTM) deep fusion modeling [22], and geo-
graphically weighted ridge regression [23]. The ordinary Kriging [24,25], CoKriging [26],
and spatial–temporal local weighted linear regression Kriging (STLWLRK) [27] methods
have been developed for merging data from meteorological stations and ground-based
radar. Multilayer perceptron networks [28] and the MetNet neural weather model [29]
can be used to merge satellite and ground-based radar data. To further improve spatial
resolution without reducing accuracy, researchers have introduced the high-resolution
spatial structure analysis of radar precipitation data based on techniques of merging sta-
tion and satellite data, as well as developing methods for merging data from all three
sources (stations, satellite, and radar), including Monte Carlo-based multi-objective opti-
mization [30], Bayesian averaging [31], geographically weighted regression, and artificial
neural networks [32].
Despite the progress in developing methods for daily precipitation estimation based
on multi-source data merging, the use of daily precipitation as an indicator of precipitation
intensity remains a potential source of bias. The intensity of prolonged light precipitation
may be overestimated, while the intensity of brief heavy precipitation may be underesti-
mated, and two different intensity figures may be reported for a single precipitation event
spanning two days [33]. Using hourly data provides a more accurate indicator of the actual
precipitation intensity, reducing the sampling error while recording more details regarding
each precipitation event [34]. Compared to normal rainfall events, heavy rainfall events
are associated with higher precipitation values and more pronounced spatial differences
in precipitation, leading to lower data accuracy. In addition, there is a need to test the
applicability of multi-source hourly precipitation data merging for studying different types
of heavy rainfall events, such as monsoon rainstorms and typhoons. Therefore, in this
study, we analyzed the correlations between selected variables and the hourly precipitation
observed at 250 meteorological stations in the mountainous areas of Northern Guangdong
Province during four heavy rainfall events in 2018 (event I, 23–27 April; event II, 7–10 May;
event III, 26–30 August; and event IV, 16 and 17 September). The auxiliary data analyzed
included radar precipitation data, satellite precipitation data, elevation, distance from the
coastline, and latitude and longitude. In addition, we sought to determine the optimal
multi-source precipitation data merging method under the theoretical framework of ma-
chine learning and geostatistics. Accordingly, we analyzed the heavy rainfall events by
data merging using five machine learning algorithms (XGBoost, GBDT, RF, LightGBM, and
MLR) and the CoKriging precipitation merging model, then compared the results.
 Remote Sens. 2022, 14, x FOR PEER REVIEW 3 of 24

GBDT, RF, LightGBM, and MLR) and the CoKriging precipitation merging model, then
Remote Sens. 2022, 14, 1750 compared the results. 3 of 23

2. Materials and Methods

2.
2.1.Materials
Study Areaand
andMethods
Data Sources
2.1. Study of
Overview Area
theand Data
Study Sources
Area
Overview of the Study Area
The study area is located in the mountainous area of Northern Guangdong Province,
with The
a latitude
study range
area isoflocated
23°33′16″ to 24°33′16″
in the mountainous N and a longitude
area of Northern range of 112°18′2″
Guangdong to
Province,
114°18′38″
with E. It includes
a latitude range ofmost23◦ 33 0 1600ofto
areas 24◦ 330 16City
Qingyuan 00 N and
andparts of Zhaoqing,
a longitude of 112◦ 180 200 to
rangeGuangzhou,
114 ◦ 180 3800
Huizhou, andE. Shaoguan,
It includesas shown
most areasin Figure 1. The eastern
of Qingyuan City andandparts
western parts of theGuangzhou,
of Zhaoqing, study
area are mainly composed of mountain ranges, so the overall
Huizhou, and Shaoguan, as shown in Figure 1. The eastern and western parts of theterrain is high in the eaststudy
and west and low in the middle, with a maximum elevation difference
area are mainly composed of mountain ranges, so the overall terrain is high in the east of 1421 m. The area
of lowest
and west elevation
and low in is the
the Beijiang
middle, River
with aValley
maximumin Yingde, Qingxin,
elevation and Qingcheng
difference in The
of 1421 m. the area
southeast of Qingyuan, mostly below 20 m. The study area is in a subtropical
of lowest elevation is the Beijiang River Valley in Yingde, Qingxin, and Qingcheng in the monsoon
climate zone,
southeast with an average
of Qingyuan, mostlyannual
below temperature
20 m. Thebetween
study area18.9 is
°Cinand 22 °C. Rainfall
a subtropical is
monsoon
abundant, with an average annual precipitation of 1631.4–2149.3 mm
climate zone, with an average annual temperature between 18.9 ◦ C and 22 ◦ C. Rainfall and an annual aver-
age of 160–173 days with precipitation (daily precipitation ≥ 0.1 mm/d). The area is located
is abundant, with an average annual precipitation of 1631.4–2149.3 mm and an annual
in one of the three belts of heavy rainfall in Guangdong Province and is typical of the parts
average of 160–173 days with precipitation (daily precipitation ≥ 0.1 mm/d). The area is
of Guangdong known as “rain nests” [35]. The heaviest rainfall in the study area is con-
located in one of the three belts of heavy rainfall in Guangdong Province and is typical of
centrated in Southeastern Qingyuan, Northeastern Guangzhou, Southern Shaoguan, and
the parts of Guangdong known as “rain nests” [35]. The heaviest rainfall in the study area
Northern Huizhou.
is concentrated in Southeastern Qingyuan, Northeastern Guangzhou, Southern Shaoguan,
and Northern Huizhou.

Figure 1. Study area and the distribution of the meteorological stations.

Figure 1. Study area and the distribution of the meteorological stations.
2.2. Research Data
2.2. Research Data
Surface topography has a significant effect on precipitation [36]. Therefore, the research
data Surface topography
also included has ageographic
auxiliary significant effect on precipitation
parameters. [36]. data
Precipitation Therefore, the re-
were observed at
search data also included auxiliary geographic parameters. Precipitation data were
meteorological stations by radar and by satellite; the auxiliary geographic parameters ob-
served at elevation,
included meteorological stations
distance by coastline,
to the radar andand
by satellite;
latitude the
andauxiliary geographic pa-
longitude.
rameters included elevation, distance to the coastline, and latitude and longitude.
2.2.1. Precipitation Data
2.2.1. Precipitation Data
The station-observed precipitation data used in this study were obtained from the
The station-observed precipitation data used in this study were obtained from the
Guangdong Meteorological Bureau (http://data.cma.cn/wa accessed on 1 April 2020). We
Guangdong Meteorological Bureau (http://data.cma.cn/wa accessed on 1 April 2020). We
selected quality-controlled hourly precipitation data, deleting implausible values (data is
selected quality-controlled hourly precipitation data, deleting implausible values (data is
null or out of reasonable range), obtained from 250 meteorological stations in the study
null or out of reasonable range), obtained from 250 meteorological stations in the study
area (Figure 1) for four heavy rainfall events in 2018: event I (monsoon rainstorm, 23
area (Figure 1) for four heavy rainfall events in 2018: event I (monsoon rainstorm, 23
April), event II (monsoon rainstorm, 7–10 May), event III (extreme monsoon rainstorm,
April), event II (monsoon rainstorm, 7–10 May), event III (extreme monsoon rainstorm,
26–30 August),and
26–30 August), andevent
eventIVIV(Typhoon
(Typhoon Mangkhut,
Mangkhut, 16 and
16 and 17 September).
17 September). Hourly
Hourly radar-radar-
derived precipitationdata
derived precipitation dataforfor
thethe four
four heavy
heavy rainfall
rainfall events
events werewere
also also
used used
in thisinstudy.
this study.
We
We used the radar-based quantitative precipitation estimation (RQPE) product provided by
used the radar-based quantitative precipitation estimation (RQPE) product provided
the Guangdong Meteorological Bureau, with a spatial resolution of 1 km and a temporal
resolution of 6 minutes. First, the reflectivity (Z) during 6-minute volume scans and the
corresponding hourly precipitation intensity (I) were used to construct a model of the
Z–I relationship, and the 6-minute radar precipitation estimates were inverted; then, the
radar data were calibrated by the station data, and the cumulative sum method was used
to calculate radar-derived hourly precipitation data with a spatial resolution of 1 km.
Satellite precipitation data were obtained from IMERG [37] and GSMaP [38], which use
Remote Sens. 2022, 14, 1750 4 of 23

data from GPM mission. GPM is a new-generation global precipitation measurement

program, the successor to TRMM and jointly implemented by the National Aeronautics and
Space Administration (NASA) and Japan Aerospace Exploration Agency (JAXA) in 2014.
Both IMERG and GSMaP are among the most widely used data products currently [39].
This study used IMERG Final Run data with a temporal resolution of half an hour and a
spatial resolution of 0.1◦ . GSMaP precipitation data represent a reanalyzed version of data
from the near-real-time Global Rainfall Watch (conducted by JAXA for meteorological and
climate research) [40], with a temporal resolution of 1 h and a spatial resolution of 0.1◦ .
GSMaP and IMERG data were downscaled to a spatial resolution of 0.01◦ and a temporal
resolution of 1 h by interpolation using the area-to-point Kriging (ATPK) [41] method and
a cumulative summation of the IMERG data.

2.2.2. Auxiliary Geographic Parameters

Elevation was determined based on global digital surface model (DSM) data with a
spatial resolution of 30 m, provided by Advanced Land Observing Satellite, Panchromatic
Remote-sensing Instrument for Stereo Mapping (ALOS PRISM) (https://www.eorc.jaxa.jp/
ALOS/en/aw3d30/index.html accessed on 10 June 2019). ALOS DSM data are produced
based on the ALOS panchromatic three-line array imagery (front view, vertical view, and
rear view) with a resolution of 2.5 m and global coverage. Their horizontal and vertical
accuracy can reach 10 m, and they are widely used to monitor the progress of urban
construction and forest growth. By aggregating and averaging the DSM and cropping
based on the vector boundaries of study area, a digital elevation model with a spatial
resolution of 1 km was obtained.
Data on distance from the coastline were extracted from 2017 Worldview-3 high-
resolution remote sensing images through human–computer interactive vectorization.
Based on the spatial analysis method, the coordinates of the center point of each 1-km grid
in the study area were extracted, and the closest distance from these center points to the
coastline was calculated, giving the distance to the coastline of each 1-km grid.

2.3. Methodology
2.3.1. Multi-Source Precipitation Data Merging Methods
We constructed station–radar–satellite hourly precipitation merging models based on
machine learning algorithms and geostatistical methods, together with auxiliary geographic
parameters, including topography, latitude and longitude, and distance from the coastline.
Additionally, to facilitate the high-accuracy merging of multi-source precipitation data, we
developed a CoKriging data merging model with station-observed precipitation as the
primary variable and the radar precipitation data as a covariate. Finally, merged hourly
precipitation data with a spatial resolution of 1 km were obtained. A flowchart of the
hourly precipitation data merging methods in this study is shown in Figure 2. There
were four main steps: First, IMERG and GSMaP data with a spatial resolution of 0.1◦
were spatially downscaled using the geostatistical ATPK method, respectively. Second, a
regression prediction model was constructed using machine learning algorithms based
on the correlation of station precipitation data with radar precipitation data, satellite
precipitation data, and auxiliary geographic variables. Third, the residuals between the
model estimates and station observations were interpolated using the ordinary Kriging
interpolation algorithm. Fourth, the model prediction results were corrected using the
interpolated model residuals, producing high-accuracy hourly precipitation merging data
with a spatial resolution of 1 km.
Remote Sens. 2022, 14, x FOR PEER REVIEW 5 of 24

Remote Sens. 2022, 14, 1750

Kriging interpolation algorithm. Fourth, the model prediction results were corrected
5 of 23
us-
ing the interpolated model residuals, producing high-accuracy hourly precipitation merg-
ing data with a spatial resolution of 1 km.

Figure 2. Flow chart of multi-source precipitation data merging.

Figure 2. Flow chart of multi-source precipitation data merging.

2.3.2.
2.3.2. Machine
Machine Learning-BasedHourly
Learning-Based Hourly Precipitation
PrecipitationData
DataMerging
MergingModels
Models
In In this
this study,five
study, fivemachine
machinelearning
learning algorithms
algorithms(GBDT,
(GBDT,XGBoost,
XGBoost, LightGBM,
LightGBM, RF, RF,
andand
MLR) were used to construct regression models for station precipitation data,
MLR) were used to construct regression models for station precipitation data, radar pre-radar precip-
itation data, and auxiliary geographic parameters. The model estimates were compared
cipitation data, and auxiliary geographic parameters. The model estimates were com-
with CoKriging interpolation results, as shown in Equations (1) and (2). In Equation (1),
pared with CoKriging interpolation results, as shown in Equations (1) and (2). In Equation
(1), ˆ ML = f ML ( Radar, I MERG, GsMaP, Lon, Lat, DEM, Coastline)
Prec (1)

ˆ ML denotes
where Prec =precipitation
( , data predicted
, , a machine
by , , learning
, )
algorithm; f ML
(1)
denotes
where a regression model constructed based on machine learning algorithms;
denotes precipitation data predicted by a machine learning algorithm; Radar,
I MERG,
denotes and GsMaP
a regression denote
model radar and satellite
constructed based onprecipitation data; Lonalgorithms;
machine learning and Lat denote ,
, and denote radar and satellite precipitation data; and denote
Remote Sens. 2022, 14, 1750 6 of 23

latitude and longitude, respectively; DEM denotes elevation; and Coastline denotes dis-
tance from the coastline. The inputs of the constructed machine learning model were radar
precipitation data (spatial resolution of 1 km) and the auxiliary geographic variables, and
the output was predicted precipitation data (spatial resolution of 1 km). Then, the residuals
of the model were interpolated using the ordinary Kriging interpolation algorithm, as
shown in Equation (2):
n
ε̂ ML ( x ) = ∑ λi ε ML (Xi ) (2)
i =1

where ε̂ ML ( x ) is the ordinary Kriging estimate for the residual of the machine learning
model at spatial location x, and λi denotes the weight of the ordinary Kriging interpolation
method. Thus, to obtain high-accuracy precipitation merging data with a spatial resolution
of 1 km, it is recommended to use the interpolation results of the model residuals to correct
the 1-km precipitation prediction results.

2.3.3. GBDT
The gradient boosting decision tree (GBDT) is an iterative decision tree model, based on
a boosting algorithm, that achieves classification and regression by continuously reducing
residuals. The GBDT algorithm generates a weak learner with each iteration, and each
learner is trained with the residuals of the learner in the previous round until a strong
classifier is finally obtained. The core concept of the GBDT algorithm is to let each tree fit
the residuals generated by the previous tree and use the cumulative results for all the trees
as the final prediction output through formula calculations.

2.3.4. XGBoost
XGBoost is an integrated learning algorithm based on the method of boosting [42]. It
is an optimized ensemble tree-based algorithm, improved and extended from the GBDT
algorithm. Its main idea is to use feature splitting to grow trees continuously, with each
generated tree representing a new function used to fit the residuals of the previous tree;
finally, the calculated value of each leaf node is added to obtain the final predictive value:

K
ŷi = ∑ f k ( Xi ) , f k ∈ F (3)
K =1

where ŷi is the model-predicted value, K denotes the number of trees, F is the ensemble
space of the regression tree (also known as CART), and Xi
denotes the feature vector of the
i-th data point. F = { f ( X ) = wq(X ) } q : Rm → T, w ∈ R T , where q denotes the structure
of each tree by which the examples are mapped to the corresponding leaf indices, T is the
number of leaves on the tree, and f k corresponds to the structure q and the leaf weight w of
k-th independent tree. Each regression tree contains consecutive scores on each leaf, and
the score on the i-th leaf is denoted by wi . The objective function of the XGBoost algorithm
includes a loss function and a regularization term:

L(∅) = ∑ l (ŷi , yi ) + ∑ Ω( f k ) (4)

i k

1
Ω( f ) = γT + λ||w||2 (5)
2
where l (ŷi , yi ) denotes the training error between the predicted value ŷi and true value of
the target yi . The regularization term Ω penalizes the complexity of the model to smooth
the final learned weight and avoid overfitting, and γ and λ denote the penalty coefficients
of the model.
Remote Sens. 2022, 14, 1750 7 of 23

2.3.5. LightGBM
Light Gradient Boosting Machine (LightGBM) is a GBDT variation for big data pro-
cessing that balances efficiency and accuracy [43]. The characteristics of the LightGBM
algorithm are as follows: (1) A leaf-wise algorithm with a depth limit is adopted to replace
the level-wise strategy used by most GBDT tools, (2) data volume and accuracy are bal-
anced using a gradient-based one-sided sampling (GOSS) algorithm that can exclude most
samples with small gradients and calculate information gain using the remaining samples,
and (3) the exclusive feature bundling (EFB) method is used to reduce the data volume
by reducing the number of features. LightGBM uses a histogram algorithm to reduce the
memory occupied by the method and the complexity of the data separation. Its core idea
is to convert continuous features into discrete values and construct a histogram, and the
cumulative statistics of each discrete value in the histogram are counted by traversing the
training data. During feature selection, the optimal splitting point can be determined by
simply traversing the discrete values in the histogram. Moreover, the histogram can be
accelerated by the difference. Leaf nodes with large histograms can be obtained based on
histogram differences between the small leaf nodes, thus minimizing the computational
effort of obtaining histograms for each leaf nodes.

2.3.6. RF
Random forest (RF) is a combination of decision trees where each tree depends on
a random vector value with the same distribution as the forest [44]. RF is a product of
integrated learning, which combines the integrated Bagging (bootstrap aggregating) [44]
and classification and regression tree (CART) algorithms. The idea of RF is to randomly
select N samples from the original training sample set repeatedly with replacements to form
the sample subsets and then generate N decision trees based on the subsets. Each decision
tree is judged to obtain N classification results, and the final classification is determined by
voting. RF has the following characteristics: (1) the subsets are independent of each other,
which enables parallel computing and ensures high efficiency; (2) because of the Bagging
method, the decision tree is not too complex and does not require pruning; and (3) the
existence of out-of-bag (oob) data makes it unnecessary to select a validation set separately.

2.3.7. MLR
Based on linear relationships between the precipitation data from meteorological sta-
tions, radar precipitation data, and auxiliary geographic parameters, we constructed a multi-
source precipitation data merging model based on the multiple linear regression (MLR)
method. The parameters of the MLR model were solved by the least squares method to ful-
fill the requirement that the residual sum of squares Q = ∑m Preci − Precˆ i 2 be minimized.


i =1

2.4. CoKriging-Based Hourly Precipitation Merging Model

The CoKriging interpolation method interpolates one or more target variables based
on several variable data and their spatial and intervariable correlations [45]. In this study,
CoKriging interpolation was performed with station-observed precipitation as the primary
variable and radar precipitation data as the covariate, as shown in Equation (6):

N1 N2
Zgauge,CK ∗ ( x0 ) = ∑ λ1i Zgauge (xi ) + ∑ λ2j Zradar


xj (6)
i =1 j =1

where Zgauge,CK ∗ ( x0 ) is the estimated value at
x0 ; Zgauge ( xi ) is the value of the primary vari-
able, station-observed precipitation; Zradar x j is the value of the covariate, the radar precip-
N
itation data; λ1i and λ2j are the weights of Zgauge and Zradar , respectively; ∑i=11 λ1i = 1; and
Remote Sens. 2022, 14, 1750 8 of 23

N
∑ j=21 λ2j = 0. The CoKriging method uses the variance function and covariance function to
perform unbiased optimal estimations according to the following formulae:

N (h)
1
∑
2
γgauge (h) = Zgauge ( xi ) − Zgauge ( xi + h) (7)
2N (h) i =1

N (h)
1
∑ Zgauge ( xi ) − Zgauge ( xi + h)

γgauge,radar (h) = 2N (h)
i =1 (8)
·[ Zradar ( xi ) − Zradar ( xi + h)]
where N (h) is the number of samples used to calculate the variance function, and h is the
sample distance.

2.5. Evaluation Method

The merging results were evaluated based on precipitation observations from the test
stations, and the evaluation indices included the correlation coefficient (CC), root mean
square error (RMSE), and mean absolute error (MAE):
 
∑im=1 Prec ˆ
ˆ i − Prec Preci − Prec


CC = r  2 q (9)
m ˆ m

2
ˆ
∑i=1 Preci − Prec ∑i=1 Preci − Prec
s
m
1
2
RMSE =
m ∑ ˆ i
Preci − Prec (10)
i =1
m
1
∑  Preci − Prec
ˆ i
 
MAE = (11)
m i =1

where m is the number of stations, Preci denotes the precipitation data from the i-th ground-
ˆ i denotes the multi-source precipitation merging data,
based meteorological station, Prec
Prec denotes the average station-observed precipitation, and Prec ˆ denotes the average of
the multi-source merging precipitation.

3. Results
3.1. Evaluation of the Accuracy of Merging Results
Hourly precipitation observations were selected from 20% of the meteorological sta-
tions (50 stations), which were evenly distributed over the study area and varied at each
timepoint (Figure 1), to evaluate the accuracy of the merged hourly multi-source pre-
cipitation data for four heavy rainfall events in terms of the three indices (CC, RMSE,
and MAE).
Figures 3–6 present the merged precipitation data obtained for the four heavy rainfall
events by the six merging methods. The scatterplots show that the main errors were
underestimation for high-precipitation timepoints and overestimation for low- or no-
precipitation timepoints. CoKriging produced fewer errors in both cases than the machine
learning methods, so it was considered the most accurate, giving CC values for the four
heavy rainfall events of 0.783, 0.806, 0.727, and 0.828, respectively. Comparing the different
machine learning algorithms, GBDT and XGBoost had the best performances for events
I (CC = 0.756) and IV (CC = 0.820), respectively, whereas RF provided the highest CC for
events II (CC = 0.787) and III (CC = 0.678). MLR resulted in the lowest CC and the highest
RMSE and MAE, indicating the lowest accuracy. Moreover, MLR provided the most severe
overestimation for the low- and no-precipitation timepoints during events I, II, and III.
022, 14, x FOR PEER REVIEW 9 of 24

022, 14, x FOR PEER REVIEW 9 of 24

Remote Sens. 2022, 14, 1750 9 of 23

Figure 3. Comparison of the observed and estimated precipitation for heavy rainfall event I: (a)
XGBoost, (b) MLR, (c) RF, (d) GBDT, (e) LightGBM, (f) CoKriging (the red line represents the fitted
curve between
Figure in-situ
3. Comparison
Figure and estimated
of3.the precipitation;
observed
Comparison and the precipitation
estimated
of the observed black line represents
and estimated that
forprecipitation
heavy thefor
rainfallratio of in-situ
event
heavy I: (a) event I:
rainfall
precipitation
XGBoost, to estimate
(b) MLR, (c) RF,precipitation
(d) GBDT, is
(e) 1:1).
LightGBM, (f) CoKriging (the red line represents the fitted
(a) XGBoost, (b) MLR, (c) RF, (d) GBDT, (e) LightGBM, (f) CoKriging (the red line represents the
curve between in-situ
fittedand
curve estimated
between precipitation; the black
in-situ and estimated line represents
precipitation; thatline
the black therepresents
ratio of in-situ
that the ratio of
precipitation to estimate precipitation
in-situ precipitation to is 1:1). precipitation is 1:1).
estimate

Figure 4. Comparison
Figureof4.theComparison
observed and estimated
of the observedprecipitation
and estimated forprecipitation
heavy rainfallfor event
heavy II: (a) event II:
rainfall
XGBoost, (b) MLR, (c) RF, (d) GBDT, (e) LightGBM, (f) CoKriging (the red line represents the fitted
(a) XGBoost, (b) MLR, (c) RF, (d) GBDT, (e) LightGBM, (f) CoKriging (the red line represents the
curve between
Figure in-situ
fittedand
4. Comparison of estimated
the
curve observed
between precipitation;
and estimated
in-situ theprecipitation
and estimatedblack line represents
forthe
precipitation; heavy that
black therepresents
rainfall
line ratio of II:
event in-situ
that(a)
the ratio of
precipitation
XGBoost, to estimate
(b) MLR, (c) RF,
in-situ precipitation is
(d) GBDT,to(e)
precipitation 1:1). precipitation
LightGBM,
estimate (f) CoKriging
is 1:1). (the red line represents the fitted
curve between in-situ and estimated precipitation; the black line represents that the ratio of in-situ
precipitation to estimate precipitation
Figure 7 shows is
the1:1).
boxplots of the evaluation indices for the precipitation timepoints
during each heavy rainfall event (when the sum of the observations at the 250 stations was
greater than 10 mm). CoKriging produced the highest average and median CC for all four
events, as well as the lowest the RMSE and MAE, indicating that CoKriging had the highest
merging accuracy. MLR produced the lowest CC distribution and the highest MAE and
RMSE. Among the machine learning methods, RF produced the highest upper limit of CC
for event I, but its lower limit was still lower than that of CoKriging. For all methods, CC
was the lowest for event III compared with the other heavy rainfall events, and the mean
CC was below 0.6, except with CoKriging, indicating that the merging performance was
worst for event III. The CC values of event IV were generally high, with relatively small
variations, and the merging effect was the best, which is because event IV involved high
 Remote Sens. 2022, 14, 1750 10 of 23

22, 14, x FOR PEER REVIEW 10 of 24

022, 14, x FOR PEER REVIEW cumulative precipitation, and the rainfall was continuous without abrupt 10 of 24 (e.g., the
declines
periods from 00:00 on 24 April to 23:00 on 25 April, 19:00 on 7 May to 22:00 on 9 May, and
15:00 to 20:00 on 26 August).

Figure 5. Comparison of the observed and estimated precipitation for heavy rainfall event III: (a)
XGBoost, (b) MLR, (c) RF, (d) GBDT, (e) LightGBM, (f) CoKriging (the red line represents the fitted
Figure
curve 5. Comparison
between in-situ of 5.
and
Figure the observed
estimated and estimated
precipitation;
Comparison of the the precipitation
observed black
and line forprecipitation
heavy
represents
estimated rainfall
that thefor event
ratio of III:
heavy in-situ(a) event III:
rainfall
XGBoost, (b) MLR, (c)
precipitation to estimate RF, (d) GBDT,
precipitation
(a) XGBoost, (e) LightGBM,
is (c)
(b) MLR, 1:1). (f) CoKriging (the red line represents the fitted
RF, (d) GBDT, (e) LightGBM, (f) CoKriging (the red line represents the
curve between in-situ and estimated precipitation; the black
fitted curve between in-situ and estimated line represents
precipitation; thatline
the black therepresents
ratio of in-situ
that the ratio of
precipitation to estimate precipitation is 1:1).
in-situ precipitation to estimate precipitation is 1:1).

Figure 6. Comparison of the

Figure observed and
6. Comparison of estimated
the observedprecipitation for heavy
and estimated rainfall
precipitation forevent
heavyIV: (a) event IV:
rainfall
XGBoost, (b) MLR, (a) (c)XGBoost,
RF, (d) GBDT, (e) LightGBM, (f) CoKriging (the red line represents the fitted
(b) MLR, (c) RF, (d) GBDT, (e) LightGBM, (f) CoKriging (the red line represents the
Figure
curve 6. Comparison
between in-situ of the
and observed
estimated and estimatedthe precipitation for heavy rainfall event
of IV: (a)
fitted curve betweenprecipitation;
in-situ and estimatedblack line represents
precipitation; that
the black the ratio
line in-situ
represents that the ratio of
XGBoost, (b) MLR, (c)
precipitation to estimate RF, (d) GBDT, (e) LightGBM, (f) CoKriging (the red line represents the fitted
in-situprecipitation
precipitation toisestimate
1:1). precipitation is 1:1).
curve between in-situ and estimated precipitation; the black line represents that the ratio of in-situ
precipitation to estimate precipitation is 1:1).
Figure 7 shows the boxplots of the evaluation indices for the precipitation timepoints
Figures 8–11 demonstrate the time series of the evaluation indices of the hourly
during each heavy rainfall event
precipitation (when
merging thefor
data sumtheof theheavy
four observations at theand
rainfall events 250the
stations
sum ofwas
the observed
Figure 7 shows theprecipitation
boxplots of data
the evaluation indices forstations
the precipitation timepoints
greater than 10 mm). CoKriging produced the highest average and median CC for all fourtime. It can
hourly from all ground-based at the corresponding
during each heavy be rainfall
seen thatevent (when
CoKriging the sum
provides of theaccurate
highly observations at the 250 stations
hourly precipitation mergingwasdata for the
events, as well as the lowest the RMSE and MAE, indicating that CoKriging had the high-
greater than 10 mm). CoKriging
four heavy produced
rainfall the highest
events, with average and
the CC accounting for median
over 39%,CC for all
much fourthan that
higher
est merging accuracy. MLR produced the lowest CC In distribution and the highest ofMAE
events, as well asofthe
thelowest
five the RMSE
machine andmethods.
learning MAE, indicating that
addition, CoKriging
the RMSE and had
MAE the high-
the four heavy
and RMSE. Among the machine learning methods, RF produced the highest upper limit
est merging accuracy. MLR produced the lowest CC distribution and the highest MAE
of CC for event I, but its lower limit was still lower than that of CoKriging. For all methods,
and RMSE. Among the machine learning methods, RF produced the highest upper limit
CC was the lowest for event III compared with the other heavy rainfall events, and the
of CC for event I, but its lower limit was still lower than that of CoKriging. For all methods,
mean CC was below 0.6, except with CoKriging, indicating that the merging performance
CC was the lowest for event III compared with the other heavy rainfall events, and the
was worst for event III. The CC values of event IV were generally high, with relatively
mean CC was below 0.6, except with CoKriging, indicating that the merging performance
 Remote
Remote Sens.Sens.
2022,2022, 14, x FOR PEER REVIEW
14, 1750 11 of 24 11 of 23

(e.g., the
rainfall periods
events from
are 00:00correlated
highly on 24 Aprilwith
to 23:00
theon 25 April, 19:00
variations of theonprecipitation
7 May to 22:00values,
on 9 with
May, and 15:00 to 20:00 on 26 August).
basically consistent trends of increases and decreases.

Remote Sens. 2022, 14, x FOR PEER REVIEW 12 of 24

Figure 7. Boxplots of the evaluation indices for the four heavy rainfall events: event I, 23-27 April;
Figure 7. Boxplots of the evaluation indices for the four heavy rainfall events: event I, 23–27 April;
event II, 7-10 May; event III, 26-30 August ; and event IV, 16 and 17 September.
event II, 7–10 May; event III, 26–30 August; and event IV, 16 and 17 September.
Figures 8–11 demonstrate the time series of the evaluation indices of the hourly pre-
cipitation merging data for the four heavy rainfall events and the sum of the observed
hourly precipitation data from all ground-based stations at the corresponding time. It can
be seen that CoKriging provides highly accurate hourly precipitation merging data for the
four heavy rainfall events, with the CC accounting for over 39%, much higher than that of
the five machine learning methods. In addition, the RMSE and MAE of the four heavy
22, 14, x FOR PEER
Remote Sens.REVIEW
2022, 14, 1750 13 of 24 12 of 23

2022, 14, x FOR PEER REVIEW 14 of 24

Figure 8. Evaluation indices

Figure of heavy rainfall
8. Evaluation event
indices of I. rainfall event I.
heavy

As shown in Figure 9, the accumulated precipitation at the stations increased from

100 mm to more than 1500 mm within five hours and then dropped to less than 100 mm
in 1 h on the first day of event II. After that, the precipitation continued from 0:00 on 9
May to 10:00 on 10 May, during which it tended to be stable, basically between 200 mm
and 600 mm. It could be seen that the CC, RMSE, and MAE of CoKriging were generally
better than the other methods for event II. In particular, the accuracy of CoKriging was
significantly higher than that of the machine learning methods for the precipitation period
from 23:00 on 8 May to 13:00 on 10 May. In addition, for some periods during this heavy
rainfall event, the CC of MLR was particularly low, while the RMSE and MAE were nota-
bly high.

Figure 9. Evaluation indices

Figure of heavy
9. Evaluation rainfall
indices eventrainfall
of heavy II. event II.

As shown in Figure 10, event III contained many precipitation peaks with relatively
short precipitation intervals, which was quite different from the characteristics of the other
monsoon rainstorm events, I and II. For event III, the five machine learning methods all
 As shown in Figure 10, event III contained many precipitation peaks with relatively
short precipitation intervals, which was quite different from the characteristics of the other
monsoon rainstorm events, I and II. For event III, the five machine learning methods all
had a CC dominance ratio of about 10%, and CoKriging was still the one with the highest
overall
Remote Sens. 2022, 14, 1750accuracy. The advantage of CoKriging was especially pronounced for the period13 of 23
from 0:00 on 26 August to 0:00 on 27 August at the beginning of the heavy rainfall event.

022, 14, x FOR PEER REVIEW 15 of 24

As shown in Figure 11, during event IV, the accumulate precipitation at the stations
began to increase from 3:00 on 16 September and peaked at 13:00 on that day. After that,
the precipitation continually decreased, showing a normal distribution trend in general.
The evaluation indices CC, RMSE, and MAE were still optimal for CoKriging for event
IV, yet the ratio of CoKriging dominance decreased compared with the first three heavy
rainfall events. In addition, the CC of all the methods except for MLR mostly exceeded
0.5, indicating
Figure that theindices
10. Evaluation
Figure overall
10. ofmerging
heavy
Evaluation performance
rainfall
indices event
of heavy for
III.
rainfall event
event III. IV was better.

Figure 11. Evaluation indices

Figure of heavyindices
11. Evaluation rainfallofevent
heavyIV.
rainfall event IV.

3.2. Merging Result Demonstration

It can be seen from Figures 8–11 that the precipitation merging accuracy was not low
at timepoints with high observed precipitation during the events, and the precipitation
Remote Sens. 2022, 14, 1750 14 of 23

As shown in Figure 8, the accumulated precipitation at the stations for event I started
to increase at 12:00 on 23 April and reached the peak of the entire event of about 800 mm at
18:00 on that day. After a pause of two days, the precipitation reached two small peaks at
0:00 on the last two days, respectively; the accuracy of CoKriging for event I was generally
higher than that of the other methods. However, there were cases when CoKriging was
inferior to other methods. For example, the CC of CoKriging was much lower than that of
the other five methods for the timepoint of 23:00 on 26 April. The reasons may be that the
precipitation at that timepoint was relatively low and that the rainfall centers were far apart.
Thus, the spatial correlation between the precipitation data was relatively weak, making
the accuracy of CoKriging lower than that of the machine learning methods. Figure 8
also shows that the CC of MLR was 0 for the timepoint of 18:00 on 23 April when the
accumulated precipitation at the stations reached its peak, and the corresponding RMSE
and MAE indices were very high. The comparison between MLR merging results plotted
Remote Sens. 2022, 14, x FOR PEER REVIEW 16 of 25
in Figures 12 and 13 and the merging results of the other methods revealed that this result
was caused by a miscalculation of the precipitation centers with MLR.

Figure 12. Cont.

 Remote Sens. 2022, 14, x FOR PEER REVIEW 17 of 25

Remote Sens. 2022, 14, 1750 15 of 23

Remote Sens. 2022, 14, x FOR PEER REVIEW 18 of 25

3–6 and the comparison between the merging results and station-observed precipitation
in Figure 13 reveal that the spatial distribution of accumulated precipitation predicted by
CoKriging agrees the best with the actual pattern, followed by the results of the tree-based
machine
Figure
Figure learning
12.12. methods.
Distribution
Distribution of the
of the Furthermore,
merging
merging results
results the
based distribution
based
on onXGBoost,
the characteristics
the XGBoost,
GBDT,GBDT, of the
RF, accumu-
LightGBM,
LightGBM, andRF, and
MLR
lated precipitation
algorithms of
and the
the MLR significantly
CoKriging deviate
interpolation from
results: (a) the actual
18:00 on 23 pattern,
April showing
algorithms and the CoKriging interpolation results: (a) 18:00 on 23 April 2018, (b) 9:00 on 7(b)
MLR 2018, May9:00ob-
on 7
2018,
vious
(c) 7:00problems
May 2018,
on 30(c)
Augustinon
7:00 events
2018, II and
30 August
and III. and
(d)2018,
13:00 (d)September
on 16 13:00 on 162018.
September 2018.

4. Discussion
4.1. Spatial Distribution Characteristics of Accumulated Precipitation
Figure 13 shows the spatial distribution of accumulated precipitation obtained by the
different merging methods for the four heavy rainfall events in 2018. It can be seen that
the accumulated precipitation as determined by the four tree-based machine learning
models GBDT, XGBoost, LightGBM, and RF has similar spatial distribution characteristics
that are significantly different from those obtained by CoKriging and MLR. Among them,
the range of areas with high accumulated precipitation in the results of the RF method is
smaller than that for GBDT, XGBoost, and LGBM; the accumulated precipitation results
of the XGBoost method are not smooth, with prominent precipitation variations; and the
accumulated precipitation results of CoKriging are obviously jagged, which is character-
istic of the interpolation method. Finally, the accumulated precipitation results of MLR
have a significantly different spatial distribution from the results of the other methods,
and the accumulated precipitation values are much higher than the other methods; these
are consistent with the hourly prediction results of MLR. The evaluation indices in Figures
Figure 13. Cont.
 Remote Sens. 2022, 14, x FOR PEER REVIEW 19 of 25

Remote Sens. 2022, 14, 1750 16 of 23

Figure13.
Figure 13.Spatial
Spatialdistribution
distributionofofthe
theaccumulated
accumulatedprecipitation
precipitationmerging
mergingresults
resultsof
ofdifferent
differentmerging
merging
methodsfor
methods for the
the four
four heavy
heavy rainfall
rainfall events:
events: (a)
(a)heavy
heavyrainfall
rainfallevent
eventI, I,(b)
(b)heavy
heavy rainfall event
rainfall II, II,
event (c)
heavy rainfall event III, and (d) heavy rainfall event IV.
(c) heavy rainfall event III, and (d) heavy rainfall event IV.
Remote Sens. 2022, 14, 1750 17 of 23

As shown in Figure 9, the accumulated precipitation at the stations increased from

100 mm to more than 1500 mm within five hours and then dropped to less than 100 mm in
1 h on the first day of event II. After that, the precipitation continued from 0:00 on 9 May to
10:00 on 10 May, during which it tended to be stable, basically between 200 mm and 600 mm.
It could be seen that the CC, RMSE, and MAE of CoKriging were generally better than the
other methods for event II. In particular, the accuracy of CoKriging was significantly higher
than that of the machine learning methods for the precipitation period from 23:00 on 8 May
to 13:00 on 10 May. In addition, for some periods during this heavy rainfall event, the CC
of MLR was particularly low, while the RMSE and MAE were notably high.
As shown in Figure 10, event III contained many precipitation peaks with relatively
short precipitation intervals, which was quite different from the characteristics of the other
monsoon rainstorm events, I and II. For event III, the five machine learning methods all
had a CC dominance ratio of about 10%, and CoKriging was still the one with the highest
overall accuracy. The advantage of CoKriging was especially pronounced for the period
from 0:00 on 26 August to 0:00 on 27 August at the beginning of the heavy rainfall event.
As shown in Figure 11, during event IV, the accumulate precipitation at the stations
began to increase from 3:00 on 16 September and peaked at 13:00 on that day. After that,
the precipitation continually decreased, showing a normal distribution trend in general.
The evaluation indices CC, RMSE, and MAE were still optimal for CoKriging for event
IV, yet the ratio of CoKriging dominance decreased compared with the first three heavy
rainfall events. In addition, the CC of all the methods except for MLR mostly exceeded 0.5,
indicating that the overall merging performance for event IV was better.

3.2. Merging Result Demonstration

It can be seen from Figures 8–11 that the precipitation merging accuracy was not low
at timepoints with high observed precipitation during the events, and the precipitation
distribution of the peak timepoints was representative. Therefore, Figure 12 demonstrates
the merging results of the timepoints with the maximum precipitation in the four heavy
rainfall events. Among these, the precipitation was more concentrated at the timepoints of
18:00 on 23 April 2018 and 9:00 on 7 May 2018 than at 7:00 on 30 August 2018 and 13:00
on 16 September 2018 in the merging results. The precipitation distribution of the MLR
merging results was significantly different from the results of the other methods at three
timepoints: 18:00 on 23 April 2018, 7:00 on 30 August 2018, and 13:00 on 16 September
2018. Referring to the characteristics of the CC, RMSE, and MAE indices of the MLR
precipitation merging results at these three timepoints in Figures 8–11, we can confirm
that MLR miscalculated the precipitation distribution. The reason for this problem was
that MLR only considered the linear relationship between the features, and the error was
relatively large for some heavy rainfall timepoints. The merging performance of MLR was
not as good as the nonlinear machine learning models [46]. The stripes in Figure 12b occur
with XGBoost, LightGBM, and RF, and this feature will be discussed in Section 4.3.

4. Discussion
4.1. Spatial Distribution Characteristics of Accumulated Precipitation
Figure 13 shows the spatial distribution of accumulated precipitation obtained by
the different merging methods for the four heavy rainfall events in 2018. It can be seen
that the accumulated precipitation as determined by the four tree-based machine learning
models GBDT, XGBoost, LightGBM, and RF has similar spatial distribution characteristics
that are significantly different from those obtained by CoKriging and MLR. Among them,
the range of areas with high accumulated precipitation in the results of the RF method is
smaller than that for GBDT, XGBoost, and LGBM; the accumulated precipitation results
of the XGBoost method are not smooth, with prominent precipitation variations; and the
accumulated precipitation results of CoKriging are obviously jagged, which is characteristic
of the interpolation method. Finally, the accumulated precipitation results of MLR have
a significantly different spatial distribution from the results of the other methods, and
Remote Sens. 2022, 14, 1750 18 of 23

the accumulated precipitation values are much higher than the other methods; these are
consistent with the hourly prediction results of MLR. The evaluation indices in Figures 3–6
and the comparison between the merging results and station-observed precipitation in
Figure 13 reveal that the spatial distribution of accumulated precipitation predicted by
CoKriging agrees the best with the actual pattern, followed by the results of the tree-based
machine learning methods. Furthermore, the distribution characteristics of the accumulated
precipitation of the MLR significantly deviate from the actual pattern, showing obvious
problems in events II and III.

4.2. Accuracy Analysis

It can be seen from Figures 3–6, Figure 7, and Figures 8–11 that CoKriging provides
the highest CC between the observed and estimated precipitation during the heavy rainfall
events, as well as the highest ratio of the timepoints with the best CC performance in the
heavy rainfall events, which is sufficient to show that the accuracy of CoKriging is the
highest. We believe that there are two main reasons for the high accuracy of the CoKriging
merging results. One is that precipitation during the heavy rainfall events has a high spatial
autocorrelation, and the other is that CoKriging introduces the highly correlated radar
precipitation data as covariate. The dependence on the covariate makes CoKriging less
accurate when the correlation between the covariate and station-observed precipitation
is relatively low. Among the hourly precipitation merging results for each heavy rainfall
event, the CC of MLR is the lowest, and its RMSE and MAE are the highest. The accuracy
of MLR is the lowest among the six methods, and its merging results are significantly
different from the other methods. The main reason is that MLR only considers the linear
relationship of the observed precipitation from the meteorological stations to the radar
precipitation data, satellite precipitation data, and auxiliary geographic variables and
cannot capture the nonlinear relationship between them, resulting in the miscalculation of
the precipitation values and precipitation distribution for some timepoints. For the four
heavy rainfall events, the accuracy of each merging method is the lowest for event III,
probably because the cumulative value of precipitation from each station in this event is
lower than those in events II and IV. In addition, the peak accumulated precipitation is only
1000 mm. There were many timepoints when the accumulated precipitation was relatively
little (accumulated precipitation greater than 10 mm and less than 100 mm) during the
heavy rainfall event in August, because the little precipitation was only observed by a small
number of meteorological stations. The prediction results for such timepoints are poor.
From Figures 8–11, it can be seen that the monsoon rainstorms in the mountainous
areas of Northern Guangdong Province lasted longer than the typhoon, with more precip-
itation intervals and more complex temporal precipitation distribution patterns. For the
monsoon rainstorm events (events I, II, and III), the machine learning methods with the
highest hourly precipitation merging accuracy are GBDT, RF, and RF, respectively, and the
merging accuracy of the RF for event I is close to that of GBDT (with CC lower than that
of GBDT by 0.007); for the typhoon event (event IV), the machine learning method with
the highest merging accuracy is XGBoost, whose merging accuracy is higher than that of
LightGBM in second place by 0.06. Therefore, the RF-based hourly precipitation merging
model is suitable for analyzing monsoon rainstorm events, and the XGBoost-based hourly
precipitation merging model is suitable for analyzing typhoon events.
As can be seen from Figures 12 and 13, the CC of machine learning methods is
mostly 0 for the timepoints with little precipitation (from 9:00 on 27 April to 23:00 on 27
April and 20:00 on 7 May to 23:00 on 8 May), while the CC of CoKriging is much higher.
This is because the relatively little precipitation was only observed by a small number of
meteorological stations. Such cases when few stations observe very little precipitation
can only be considered as intermittent precipitation, no longer heavy rainfall. Secondly,
it is difficult for the model to extract a precipitation pattern when too few stations have
precipitation records. Hence, the model treats such cases as global non-precipitation, but
the interpolation results of CoKriging still present some correlation.
Remote Sens. 2022, 14, 1750 19 of 23

Chao [47] used the multiscale geographically weighted regression (MGWR) method
to perform hourly precipitation data merging for the Ziwu River Basin, and the CC of
the merging result was 0.724. Li [48] applied the space–time multiscale analysis system
(STMAS), a multigrid variational analysis technique, to merge hourly precipitation data
during the heavy precipitation period in Jiangxi Province from May to June 2019, and the
CC of the merging result was 0.76. Figures 3–6 show the hourly precipitation merging
accuracy of six methods for four heavy rainfall events, where the CC values of the merging
results of CoKriging for events I, II, and IV (0.783, 0.806, and 0.828) are higher than the CC
of the methods in the two studies mentioned above. The CC of CoKriging for event III
(0.726) is also similar to that of GWR and STMAS, and the machine learning methods RF
and XGBoost with the highest merging accuracy for events II and IV also provide higher
CC results of 0.787 and 0.82. Therefore, the merging results for the four heavy rainfall
events in this paper have application and reference values.

4.3. Defects of the Merging Results

There are horizontal stripes in the merging results of the tree-based machine learning
methods for some timepoints. Figure 14 shows the merging results of XGBoost, LightGBM,
and RF for the rainfall-concentrated timepoint 9:00 on 7 May 2018, where striped areas
appear in the regions without concentrated precipitation (the GBDT results do not have
stripes at this timepoint, but they are present at other precipitation timepoints). First, the
precipitation patterns of these striped areas are obviously inconsistent with the actual
precipitation distribution. The correlation coefficients of the observed precipitation from
the meteorological stations at 9:00 on 7 May 2018 with the radar precipitation data, GSMaP,
IMERG, longitude, latitude, DEM, and distance from the coastline are 0.57, 0.62, 0.62, 0.21,
−0.55, −0.15, and −0.62, respectively. It can be seen that the correlation between station-
observed precipitation and latitude and distance from the coastline is very high for this
timepoint, leading to the overemphasis of XGBoost, LightGBM, and RF on the influence of
latitude and distance from the coastline; the striped regions are all high-latitude areas far
from the coastline. The striped regions at high latitudes appear under the joint effects of
high weights and high eigenvalues. These striped regions do not have particularly large
regional residual values, indicating that station-observed precipitation values within the
striped regions do not differ significantly from the predicted precipitation values for these
regions, but the striped regions cannot be removed simply by residual correction.
Although CoKriging has the highest precipitation merging accuracy, its modeling
time is also longer than the machine learning methods. Considering the issues of accuracy
and time, we suggest that spatial and temporal autocorrelation parameters should be
introduced into machine learning models in further studies and that the influences of
ambient precipitation and precipitation at previous timepoints on the modeling results
should be fully considered. In addition, the accuracy of the multi-source precipitation data
merging model could be improved by incorporating more satellite precipitation data into
the model.
Remote Sens.
Remote Sens. 2022, 14, 1750
2022, 14, x FOR PEER REVIEW 2120of
of 24
23

Figure
Figure 14.
14. Residual
Residual correction for the
correction for the results
resultsof
ofthetheXGBoost,
XGBoost,LightGBM,
LightGBM,and
andRF
RFmethods
methodsatat9:00
9:00onon7
7May
May2018:
2018:(a)(a)XGBoost,
XGBoost,(b)
(b)LightGBM,
LightGBM,and and(c)(c)RF.
RF.

Although CoKriging has the highest precipitation merging accuracy, its modeling
5. Conclusions
time We
is also longerthe
selected than the machine
mountainous learning
area methods.
of Northern Considering
Guangdong the issues
Province as theofstudy
accuracy
area
and
and time, we suggest
used data thatheavy
from four spatial and temporal
rainfall autocorrelation
events during parameters
the 2018 flood seasonshould
(23–27 be in-
April,
troduced into machine learning models in further studies and that the influences
7–10 May, 26–30 August, and 16 September) to establish precipitation data merging models of am-
bient precipitation
suitable for the area and
based precipitation
on XGBoost,atGBDT,previous timepoints
LightGBM, on the
RF, MLR, andmodeling
CoKriging. results
The
should be fully
residuals considered.
of these machineInlearning
addition,models
the accuracy of the multi-source
were corrected with the precipitation
ordinary Krigingdata
merging
method to model could
obtain be improved
hourly by incorporating
precipitation merging data more
withsatellite precipitation
a spatial resolution data
of 1 into
km,
the
andmodel.
the merging results were assessed and analyzed based on the spatial precipitation
distribution and accuracy indices. The research conclusions are as follows:
5.
(1)Conclusions
The errors in these precipitation merging results mainly involve underestimations
We selected the mountainous
for high-precipitation areaand
timepoints of Northern Guangdong
overestimations Province
for low- as the study
or no-precipitation
area and used data from four heavy rainfall events during the 2018 flood season (23–27
timepoints.
April,
(2) The7–10spatial
May, 26–30 August,ofand
distribution the16 September) to
accumulated establish precipitation
precipitation predicted by data merging
CoKriging
models suitable
agrees for with
the best the area based pattern,
the actual on XGBoost, GBDT,
followed LightGBM,
by the results ofRF,
the MLR, and
tree-based
CoKriging.
machineThelearning
residuals of these whereas
methods, machine learning models of
the distribution were corrected with
accumulated the ordi-
precipitation
nary Kriging
predictedmethod
by MLR toisobtain hourly different
significantly precipitation
frommerging
the actualdata with The
pattern. a spatial resolution
merging results
of 1 km, and the merging
of CoKriging have aresults
higherwere assessed
accuracy thanand
theanalyzed
machine based on the
learning spatialbecause
methods, precip-
itation distributionduring
precipitation and accuracy indices.
heavy rainfall The research
events conclusions
has pronounced areautocorrelation,
spatial as follows: and
radar precipitation data as a covariate are highly correlated with the station-observed
precipitation.
Remote Sens. 2022, 14, 1750 21 of 23

(3) Different machine learning methods are applicable for different types of heavy rainfall
events. The RF-based hourly precipitation merging model is suitable for analyzing
monsoon rainstorm events, and the XGBoost-based hourly precipitation merging
model is suitable for analyzing typhoon events.
(4) The merging performance of the machine learning methods is relatively poor for
the timepoints, with little precipitation during the heavy rainfall event. One reason
is that the models have difficulty in extracting features when a small number of
meteorological stations observe little precipitation; another one is the models do not
capture the temporal variability of precipitation well, while constant rain is always
observed the easiest.
(5) The hourly merging results of the tree-based machine learning models contain striped
textures at some timepoints, which is caused by an excessively high correlation
between the precipitation at these timepoints and latitude and distance from the
coastline; the MLR method showed miscalculations for the precipitation values and
locations and overestimates the accumulated precipitation for heavy rainfall events II,
III, and IV.
In summary, the hourly precipitation merging models proposed in this paper can
provide high-accuracy gridded precipitation data for heavy rainfall events in the moun-
tainous areas of Northern Guangdong Province. The models combine the advantages of
different types of precipitation data by merging precipitation observed from meteorological
stations with radar and satellite precipitation data, which is essential for studying different
types of heavy rainfall events in mountainous regions. One of their shortcomings is that
CoKriging has obvious advantages for the heavy rainfall events in South China, but its
applicability in other regions needs further research; the other is that they fail to consider
the spatiotemporal autocorrelation of precipitation during heavy rainfall events and that
the results for the spatial distribution of precipitation produced by the machine learning
methods have certain defects at some timepoints. The models need to be improved in
further studies.

Author Contributions: Conceptualization, J.X.; Methodology, J.Z. and J.X.; Software, J.Z.; Validation,
J.Z. and H.R.; Formal Analysis, J.Z., J.X. and X.D.; Investigation, J.X. and J.Z.; Resources, H.R. and
W.J.; Data Curation, J.X.; Writing—Original Draft Preparation, J.Z.; Writing—Review and Editing,
J.X., X.D. and X.L.; Visualization, J.Z.; and Supervision and Project Administration, J.X. All authors
have read and agreed to the published version of the manuscript.
Funding: This research was jointly funded by the National Natural Science Foundation of China,
grant number 41901371; the Science and Technology Planning Project of Guangdong Province, grant
number 2018B020207012; the Key Special Project for Introduced Talents Team of Southern Marine
Science and Engineering Guangdong Laboratory (Guangzhou), grant number GML2019ZD0301; the
Guangdong Innovative and Entrepreneurial Research Team Program, grant number 2016ZT06D336;
the GDAS’ Project of Science and Technology Development, grant number 2019GDASYL-0301001;
and the Science and Technology Program of Guangdong, grant number 2021B1212100006.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Acknowledgments: The authors would like to thank the anonymous reviewers for their constructive
comments. We also thank the Geographical Science Data Center of the Greater Bay Area for providing
the relevant data in this study.
Conflicts of Interest: The authors declare no conflict of interest.
Remote Sens. 2022, 14, 1750 22 of 23

References
1. Sapiano, M.; Arkin, P.A. An intercomparison and validation of high-resolution satellite precipitation estimates with 3-hourly
gauge data. J. Hydrometeorol. 2009, 10, 149–166. [CrossRef]
2. Taylor, C.M.; de Jeu, R.A.; Guichard, F.; Harris, P.P.; Dorigo, W.A. Afternoon rain more likely over drier soils. Nature 2012, 489,
423–426. [CrossRef] [PubMed]
3. Jieru, Y.; András, B. Short time precipitation estimation using weather radar and surface observations: With rainfall displacement
information integrated in a stochastic manner. J. Hydrol. 2019, 574, 672–682.
4. Tapiador, F.J.; Turk, F.J.; Walt, P.; Arthur, Y.H.; Eduardo, G.; Luiz, A.T.M.; Carlos, F.A.; Paola, S.; Chris, K.; George, J.H.; et al.
Global precipitation measurement: Methods, datasets and applications. Atmos. Res. 2012, 104, 70–97. [CrossRef]
5. Kidd, C.; Becker, A.; Huffman, G.J.; Muller, C.L.; Joe, P.; Skofronick-Jackson, G.; Kirschbaum, D.B. So, how much of the Earth’s
surface is covered by rain gauges? Bull. Am. Meteorol. Soc. 2017, 98, 69–78. [CrossRef]
6. Rana, S.; McGregor, J.; Renwick, J. Precipitation seasonality over the Indian subcontinent: An evaluation of gauge, reanalyses,
and satellite retrievals. J. Hydrometeorol. 2015, 16, 631–651. [CrossRef]
7. Xie, P.; Arkin, P.A. Analyses of global monthly precipitation using gauge observations, satellite estimates, and numerical model
predictions. J. Clim. 1996, 9, 840–858. [CrossRef]
8. Yilmaz, K.K.; Adler, R.F.; Tian, Y.; Hong, Y.; Pierce, H.F. Evaluation of a satellite-based global flood monitoring system. Int. J.
Remote Sens. 2010, 31, 3763–3782. [CrossRef]
9. Arkin, P.A.; Meisner, B.N. The relationship between large-scale convective rainfall and cold cloud over the western hemisphere
during 1982-84. Mon. Weather. Rev. 1987, 115, 51–74. [CrossRef]
10. Berg, W.; Chase, R. Determination of mean rainfall from the Special Sensor Microwave/Imager (SSM/I) using a mixed lognormal
distribution. J. Atmos. Ocean. Technol. 1992, 9, 129–141. [CrossRef]
11. Xie, P.; Arkin, P.A. Global precipitation: A 17-year monthly analysis based on gauge observations, satellite estimates, and
numerical model outputs. Bull. Am. Meteorol. Soc. 1997, 78, 2539–2558. [CrossRef]
12. Huffman, G.J.; Adler, R.F.; Arkin, P.; Chang, A.; Ferraro, R.; Gruber, A.; Janowiak, J.; McNab, A.; Rudolf, B.; Schneider, U. The
global precipitation climatology project (GPCP) combined precipitation dataset. Bull. Am. Meteorol. Soc. 1997, 78, 5–20. [CrossRef]
13. Ziqiang, M.; Jintao, X.; Kang, H.; Xiuzhen, H.; Qingwen, J.; TseChun, W.; Wentao, X.; Yang, H. An updated moving window
algorithm for hourly-scale satellite precipitation downscaling: A case study in the Southeast Coast of China. J. Hydrol. 2020, 581,
124378.
14. Gao, Y.; Xu, H.; Liu, G. Evaluation of the GSMaP Estimates on Monitoring Extreme Precipitation Events. Remote sensing
Technology and Application. Remote Sens. Technol. Appl. 2019, 34, 1121–1132.
15. Michaelides, S.; Levizzani, V.; Anagnostou, E.; Bauer, P.; Kasparis, T.; Lane, J.E. Precipitation: Measurement, remote sensing,
climatology and modeling. Atmos. Res. 2009, 94, 512–533. [CrossRef]
16. Zhang, J.; Howard, K.; Langston, C.; Kaney, B.; Qi, Y.; Tang, L.; Grams, H.; Wang, Y.; Cocks, S.; Martinaitis, S. Multi-Radar
Multi-Sensor (MRMS) quantitative precipitation estimation: Initial operating capabilities. Bull. Am. Meteorol. Soc. 2016, 97,
621–638. [CrossRef]
17. Shen, Y.; Zhao, P.; Pan, Y.; Yu, J. A high spatiotemporal gauge-satellite merged precipitation analysis over China. J. Geophys. Res.
Atmos. 2014, 119, 3063–3075. [CrossRef]
18. Alharbi, R.; Hsu, K.; Sorooshian, S. Bias adjustment of satellite-based precipitation estimation using artificial neural networks-
cloud classification system over Saudi Arabia. Arab. J. Geosci. 2018, 11, 1–17. [CrossRef]
19. Xu, G.; Wang, Z.; Xia, T. Mapping Areal Precipitation with Fusion Data by ANN Machine Learning in Sparse Gauged Region.
Applied Sciences. 2019, 9, 2294. [CrossRef]
20. Shen, Y.; Pan, S.; Xu, B.; Y, J. Parameter Improvements of Hourly Automatic Weather Stations Precipitation Analysis by Optimal
Interpolation over China. J. Chengdu Univ. Technol. 2012, 27, 219–224.
21. Kunwei, L.; Xiong, Y.; Xin, Z.; Fen, T. Multi-source Precipitation Data Fusion Method Based on Filtersim. J. Syst. Simul. 2019,
31, 1232.
22. Wu, H.; Yang, Q.; Liu, J.; Wang, G. A spatiotemporal deep fusion model for merging satellite and gauge precipitation in China. J.
Hydrol. 2020, 584, 124664. [CrossRef]
23. Chen, S.; Xiong, L.; Ma, Q.; Kim, J.; Chen, J.; Xu, C. Improving daily spatial precipitation estimates by merging gauge observation
with multiple satellite-based precipitation products based on the geographically weighted ridge regression method. J. Hydrol.
2020, 589, 125156. [CrossRef]
24. Delrieu, G.; Wijbrans, A.; Boudevillain, B.; Faure, D.; Bonnifait, L.; Kirstetter, P. Geostatistical radar–raingauge merging: A novel
method for the quantification of rain estimation accuracy. Adv. Water Resour. 2014, 71, 110–124. [CrossRef]
25. Sideris, I.V.; Gabella, M.; Sassi, M.; Germann, U. Real-Time Spatiotemporal Merging of Radar and Raingauge Precipitation
Measurements in Switzerland. In Proceedings of the 9th International Workshop on Precipitation in Urban Areas, St. Moritz,
Switzerland, 6–9 December 2012.
26. Azimi-Zonooz, A.; Krajewski, W.F.; Bowles, D.S.; Seo, D.J. Spatial rainfall estimation by linear and non-linear co-kriging of
radar-rainfall and raingage data. Stoch. Hydrol. Hydraul. 1989, 3, 51–67. [CrossRef]
27. Zhang, G.; Tian, G.; Cai, D.; Bai, R.; Tong, J. Merging radar and rain gauge data by using spatial–temporal local weighted linear
regression kriging for quantitative precipitation estimation. J. Hydrol. 2021, 601, 126612. [CrossRef]
Remote Sens. 2022, 14, 1750 23 of 23

28. Chen, H.; Chandrasekar, V.; Cifelli, R.; Xie, P. A Machine Learning System for Precipitation Estimation Using Satellite and Ground
Radar Network Observations. IEEE Trans. Geosci. Remote 2019, 58, 982–994. [CrossRef]
29. Sønderby, C.K.; Espeholt, L.; Heek, J.; Dehghani, M.; Oliver, A.; Salimans, T.; Agrawal, S.; Hickey, J.; Kalchbrenner, N. Metnet: A
neural weather model for precipitation forecasting. arXiv 2020, arXiv:2003.12140.
30. Hazra, A.; Maggioni, V.; Houser, P.; Antil, H.; Noonan, M. A Monte Carlo-based multi-objective optimization approach to merge
different precipitation estimates for land surface modeling. J. Hydrol. 2019, 570, 454–462. [CrossRef]
31. Pang, Y.; Shen, Y.; Yu, J.; Xiong, A. An experiment of high-resolution gauge-radar-satellite combined precipitation retrieval based
on the Bayesian merging method. Acta Meteorol. Sin. 2015, 73, 177–186.
32. Wehbe, Y.; Temimi, M.; Adler, R.F. Enhancing precipitation estimates through the fusion of weather radar, satellite retrievals, and
surface parameters. Remote Sens.-Basel 2020, 12, 1342. [CrossRef]
33. Li, J.; Yu, R.; Sun, W. Duration and seasonality of the hourly extreme rainfall in the central-eastern part of China. Acta Meteorol.
Sin. 2013, 71, 652–659.
34. Trenberth, K.E.; Dai, A.; Rasmussen, R.M.; Parsons, D.B. The changing character of precipitation. Bull. Am. Meteorol. Soc. 2003, 84,
1205–1218. [CrossRef]
35. Li, D.; Chen, W.; Ye, A. Climatic characteristics and forecast focus of heavy rain in Qingyuan. Guangdong Meteorol. 1999, 2, 8–10.
36. Roe, G.H. Orographic precipitation. Annu. Rev. Earth Planet. Sci. 2005, 33, 645–671. [CrossRef]
37. Huffman, G.J.; Bolvin, D.T.; Braithwaite, D.; Hsu, K.; Joyce, R.; Xie, P.; Yoo, S. NASA global precipitation measurement (GPM)
integrated multi-satellite retrievals for GPM (IMERG). Algorithm Theor. Basis Doc. ATBD Version 2015, 4, 26.
38. Shige, S.; Yamamoto, T.; Tsukiyama, T.; Kida, S.; Ashiwake, H.; Kubota, T.; Seto, S.; Aonashi, K.; Okamoto, K. The GSMaP
precipitation retrieval algorithm for microwave sounders—Part I: Over-ocean algorithm. IEEE Trans. Geosci. Remote 2009, 47,
3084–3097. [CrossRef]
39. Hou, A.Y.; Kakar, R.K.; Neeck, S.; Azarbarzin, A.A.; Kummerow, C.D.; Kojima, M.; Oki, R.; Nakamura, K.; Iguchi, T. The global
precipitation measurement mission. Bull. Am. Meteorol. Soc. 2014, 95, 701–722. [CrossRef]
40. Ushio, T.; Sasashige, K.; Kubota, T.; Shige, S.; Okamoto, K.; Aonashi, K.; Inoue, T.; Takahashi, N.; Iguchi, T.; Kachi, M. A
Kalman filter approach to the Global Satellite Mapping of Precipitation (GSMaP) from combined passive microwave and infrared
radiometric data. J. Meteorol. Soc. Jpn. Ser. II. 2009, 87, 137–151. [CrossRef]
41. Kyriakidis, P.C. A geostatistical framework for area-to-point spatial interpolation. Geogr. Anal. 2004, 36, 259–289. [CrossRef]
42. Chen, T.; Guestrin, C. Xgboost: A Scalable Tree Boosting System. In Proceedings of the 22nd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, 13–17 August 2016; pp. 785–794.
43. Ke, G.; Meng, Q.; Finley, T.; Wang, T.; Chen, W.; Ma, W.; Ye, Q.; Liu, T. Lightgbm: A highly efficient gradient boosting decision
tree. Adv. Neural Inf. Processing Syst. 2017, 30, 3146–3154.
44. Breiman, L. Bagging predictors. Mach. Learn. 1996, 24, 123–140. [CrossRef]
45. Zhang, R. Spatial Variation Theory and Applications; Science Press: Beijing, China, 2005.
46. Huang, X.; He, L.; Zhao, H.; Huang, Y.; Wu, Y. Prediction model based on the Laplacian eigenmap method combined with a
random forest algorithm for rainstorm satellite images during the first annual rainy season in South China. Nat. Hazards 2021,
107, 331–353. [CrossRef]
47. Chao, L.; Zhang, K.; Li, Z.; Zhu, Y.; Wang, J.; Yu, Z. Geographically weighted regression based methods for merging satellite and
gauge precipitation. J. Hydrol. 2018, 558, 275–289. [CrossRef]
48. Li, X.; Wei, Z.; Shaoping, H.; Weihua, D.; Xueying, Z. Analysis of fusion test results on hourly precipitation from meteorological
and hydrological stations and radar. Torrential Rain Disasters 2020, 39, 276–284.