IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL.

17, 2024 13261

Impact of Land Use/Land Cover Changes on Urban

Flooding: A Case Study of the Greater Bay
Area, China
Yuqi Lei , Hui Cao, Xuanyan Zhou , Jon Mills, and Wen Xiao , Member, IEEE

Abstract—Urban flooding caused by heavy rainfall is a common flooding induced by extreme heavy rainfall are continuously
natural hazard in cities globally. Impervious surfaces are often increasing [3], [4], [5], [6]. One of the reasons is that land
increased during urban development, but there is limited research use/land cover (LULC) changes in cities affect the surface water
on the impact of large-scale and long-term land use/land cover
(LULC) changes on urban flooding, while considering the influence flux, altering the runoff generation and convergence processes of
of using different remote sensing data sources. In this study, a watersheds, which may increase flood risk. The impact of LULC
framework to evaluate the correlation between LULC changes changes on urban flooding has gained extensive attention in the
and flooding extents is proposed, mainly comprising: 1) classifying related research communities [7], [8], [9], [10].
remote sensing time series, using different sources but adopting
LULC changes can be obtained through remote sensing ob-
the same classifier, to obtain the LULC of the Greater Bay Area,
China, over a one decade period; 2) designing flooding scenarios servations. As remote sensing technology has evolved over the
with different rainfall intensities, and using the soil conservation years, the availability of satellite data has become increasingly
service curve number (SCS-CN) model and local equal volume abundant [11]. Common optical satellite data sources, such as
method to extract the inundation extent of urban flooding; and the Landsat series, have the advantage of having long time series
3) analyzing the influence of different data sources on flood simu-
imagery, while Sentinel-2 optical data provides higher spatial
lation results, and calculating the correlations between LULC and
inundation area over the decade to analyze the impact of LULC resolution. In addition, microwave satellite data sources, e.g.,
changes on urban flooding. The resulting correlation coefficients of Sentinel-1, have the advantage of all-weather observation. At
water and built-up land are 0.93 and 0.42, and those of bare land, the same time, acquiring multitemporal multiplatform remote
grassland, orchard, and forest are −0.40, −0.61, −0.57, and −0.75, sensing images from the same area has become more conve-
respectively. The inundation derived by Sentinel and Landsat data
nient, which is conducive to multisource data fusion. Previous
showed around 99% consistency, while Landsat tends to derive
more inundation areas, with the differences mainly scattered in studies have demonstrated that the fusion of multiple remote
flat areas. sensing data sources for LULC classification, including optical
images and synthetic aperture radar (SAR) data, can improve
Index Terms—Guangdong–Hong kong–Macao greater bay area,
land use/land cover (LULC) change, landsat, soil conservation the classification accuracy based on pixel-based methods, and
service curve number (SCS-CN) model, Sentinel, urban flooding. the classification results are superior to those using only mul-
tispectral or microwave data [12], [13], [14], [15], [16], [17].
Different remote sensing data sources have their own strengths
I. INTRODUCTION and weaknesses, and it is necessary to evaluate the impact of
using different data on LULC classification and subsequent flood
RBAN flooding is a regional flooding phenomenon caused
U by excessive rainfall and low-lying terrain that cannot be
drained in time when short-term heavy precipitation exceeds the
modeling experiments.
Surface runoff is currently mainly simulated through various
hydrological models. Commonly used hydrological models in-
drainage capacity [1], [2]. Under the global process of climate
clude soil conservation service curve number (SCS-CN), Mike,
change and urban sprawl, the frequency and severity of urban
storm water management model (SWMM), InfoWorks ICM, and
Wallingford models [18], [19], [20], among which the SCS-CN
Manuscript received 6 February 2024; revised 5 June 2024 and 28 June 2024; model requires few parameters. It is easy to operate and is
accepted 11 July 2024. Date of publication 18 July 2024; date of current version commonly used to compute surface runoff in coastal areas [21],
5 August 2024. This work was supported by the National Natural Science [22], [23], [24]. Many studies have used it to simulate the
Foundation of China under Grant 42201485. (Corresponding author: Wen Xiao.)
Yuqi Lei, Hui Cao, and Xuanyan Zhou are with the School of Geography and generation and convergence processes of flood disasters in small
Information Engineering, China University of Geosciences, Wuhan 430074, watersheds [25], [26], [27]. Liu and Li [28] used the SCS-CN
China (e-mail: [email protected]). model to calculate the runoff process and the model accuracy
Jon Mills is with the School of Engineering, Newcastle University, NE1 7RU
Newcastle upon Tyne, U.K. was above 75%. The results proved that the SCS-CN model
Wen Xiao is with the School of Geography and Information Engineering, could simulate the runoff process in a typical subwatershed on
National Engineering Research Center of Geographic Information System, the Loess Plateau. The model was also combined with the local
China University of Geosciences, Wuhan 430074, China (e-mail: wen.xiao@
cug.edu.cn). equal volume method [29] and proved to be suitable for flooding
Digital Object Identifier 10.1109/JSTARS.2024.3430505 extent extraction in larger watersheds [30].
© 2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see
https://creativecommons.org/licenses/by-nc-nd/4.0/
13262 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 17, 2024

The Greater Bay Area, China, is a coastal region with ex-

panding built-up land, increasing population, and industrial
concentration [31]. The rapid urbanization there has led to an
increased risk of urban flooding [20], [32], [33], [34], [35]. Prior
studies of the Greater Bay Area have evaluated the effects of
urbanization on precipitation [36], [37], and calculated the depth
of surface runoff under extreme rainfall [38], [39], [40], but did
not further simulate the inundation extent after a rainstorm, or
explore the impact of LULC change on the inundation range. In
addition, LULC classification is an important input for runoff
calculation, whereas little research has studied the impact of
classification results from different remote sensing data sources
on urban flooding simulation.
In this study, multitemporal satellite imagery is used to obtain
LULC changes in the Greater Bay Area over the past decade. Fig. 1. Elevation map of the Greater Bay Area, China.
With the help of the SCS-CN model, surface runoffs under
various rainfall intensity scenarios are calculated, and the inun- and Zhaoqing in Guangdong Province, and Hong Kong and
dation extent of urban flooding is extracted. Then, the inundation Macao Special Administrative Regions, China. In 2022, the
extent of each administrative region and various LULC types are Greater Bay Area had a population of 86.62 million and a gross
statistically analyzed over a decade to calculate the correlation domestic product (GDP) of 1943.54 billion US dollars. With
coefficient between the extent of urban flooding and LULC its dense population and developed economy, urban flooding
changes. Additionally, Sentinel-1 and Sentinel-2 images are hazards in the region can cause significant losses. Unfortunately,
jointly used for LULC classification to compare with Landsat the region is vulnerable to urban flooding owing to geographical
data, and urban flooding simulation is conducted on both data and climatic conditions. The Greater Bay Area has a subtropical
sources to explore the influence of different data sources and maritime monsoon climate with concentrated and heavy rainfall.
feature selection on classification and flood simulations. The terrain of a high north and low south makes it susceptible
The main contributions of this study are threefold: 1) obtain to flash floods. The central urban areas have low-lying and flat
the LULC changes in the Greater Bay Area over a decade, terrain with little gradient, resulting in poor natural drainage
and simulate regional scale urban flooding under four rainfall capacity [41], [42], [43].
scenarios on this basis; 2) assess the influence of using different
remote sensing data sources, Landsat and Sentinel, on urban
flood simulation, revealing the consistency between the two in B. Data
both flooded and unflooded areas; 3) quantitatively evaluate The data used for the study include remote sensing imagery,
the correlation coefficient between each of the eight LULC soil texture, soil type, weather station locations, etc. Table I
changes and urban flooding extents for the first time, providing shows all utilized data sources.
more detailed and accurate references on the impact of urban Landsat images are sourced from the Landsat 8 Level 2,
development on flooding. The results reported in this article Collection 2, Tier 1 dataset on Google Earth Engine (GEE). The
provide scientific support for the Greater Bay Area to improve years of image acquisition are 2013, 2015, 2018, 2020, and 2022.
its capacity to mitigate urban flooding. Sentinel-1 images are sourced from Sentinel-1 SAR GRD: C-
The rest of this article is organized as follows. Section II of this band SAR Ground Range Detected, log scaling dataset on GEE.
article details the study area and the data utilized. In Section III, Sentinel-2 images are sourced from harmonized Sentinel-2 MSI:
the experimental procedures will be followed to introduce Multispectral Instrument, Level-2 A dataset on GEE. The year
preprocessing, LULC classification, runoff modeling, flooding of image acquisition is 2022.
extent estimation, and statistical methods. In Section IV, it The spatial distribution of soil texture in China, provided
will sequentially present the interannual changes in LULC, the by the Data Center of Resource and Environmental Sciences,
analysis of flood simulation, the impact of LULC changes on Chinese Academy of Sciences (RESDC), is divided according
urban flooding, and the effects of different data sources on to the content of sand, silt, and clay particles. The content of
the experimental results. In Section V, the deficiencies of the particles with different textures is reflected by percentage. Based
experiment will be summarized, and future prospects will be on this, a soil type map with a one-kilometer resolution was
outlined. Finally, Section VI concludes this article. obtained according to the International Standard for Soil Texture
Classification [44].
II. STUDY SITE AND DATA The locations of 33 weather stations in the study area were
obtained from the China National Meteorological Information
A. Study Site Center and NOAA National Centers for Environmental Infor-
As shown in Fig. 1, the Guangdong-Hong Kong-Macao mation. Among them, three stations are located in Hong Kong,
Greater Bay Area consists of the cities of Guangzhou, Zhuhai, one is in Macau, and 29 belong to the other cities in Guangdong
Shenzhen, Foshan, Dongguan, Huizhou, Jiangmen, Zhongshan, province.
LEI et al.: IMPACT OF LAND USE/LAND COVER CHANGES ON URBAN FLOODING 13263

TABLE I
RESEARCH DATA AND SOURCES

III. METHODOLOGY also performed coherent speckle filtering and shadow removal
As shown in Fig. 2, the research methodology in this article on mountainous areas. The terrain data required for processing
has four steps: LULC classification, SCS-CN runoff model- comes from SRTM30. The latest data for each month within
ing, flood extent estimation, and statistical analysis. In LULC the study period were obtained and combined to create monthly
classification, two different data sources are used for random images. Due to the cloudy weather conditions in the Greater Bay
forest classification. The first source is Landsat 8 imagery Area, there were few images that could be used for classification
from the years 2013, 2015, 2018, 2020, and 2022. The second most of the time.
source involves Sentinel-1 and Sentinel-2 imagery from 2022,
for comparison with Landsat results from the same year. The B. Land Use/Land Cover Classification
classification results are then used in subsequent experiments. 1) Classification Using Landsat 8 Data: Considering the
Next, in the runoff modeling, four rainfall intensity scenarios need to capture long-term LULC changes, the random forest
are simulated using the SCS-CN model to obtain surface runoff. method was used to classify five years of Landsat 8 data over
Then, the watersheds are divided based on the digital elevation the last decade. The LULC types were categorized into eight
model (DEM) and the locations of reservoirs, and flooded areas classes: farmland, orchard, forest, grassland, bare land, built-up
are extracted using the local equal volume method. Finally, land, water, and wetland. Random forest is an algorithm based
LULC and inundation areas over the decade are statistically on classification trees [46]. Among various classifiers, random
analyzed after normalization, i.e., calculating the correlation forest has demonstrated higher accuracy, efficiency, and rela-
coefficient between the flood area and each LULC. Moreover, tively lower computational complexity. It has become a reliable
a comparative analysis is conducted on the classification results choice for studying LULC changes [47].
and flooded areas obtained from the two satellite data sources. The random forest model in the study was trained using
sample data from the 2020 imagery. Due to the difficulty in
distinguishing some LULCs on the 30-m resolution images of
A. Preprocessing
Landsat 8, we first recorded the location of LULCs based on
Before LULC classification, we preprocessed the remote high-resolution images of Google Earth, and then confirmed and
sensing images. selected them as samples on the Landsat 8 images. The number
Landsat 8 images were obtained on GEE for 2013, 2015, 2018, of decision trees was tested at an interval of 5 within the range
2020, and 2022. The imagery has undergone preprocessing such of 0–150, and the one with the highest classification accuracy
as atmospheric correction and orthorectification. Further cloud was taken as the final number of decision trees. During training,
removal and fusion are needed to obtain the annual Landsat 8 all classes were treated as balanced, which helped improve
images for five epochs. A mask is made from the cloud shadows the extraction of classes with fewer samples. The model was
and cloud data of the QA_PIXEL band of each image, and then used to classify imagery from the years 2013, 2015, 2018,
the pixels blocked by clouds and cloud shadows are removed and 2022. The annual validation data were selected by visual
through the mask. After cloud removal, the median pixel values interpretation on Google Earth, ensuring that they did not overlap
of the annual images in the research area are extracted to form with the training samples. The overall accuracy represents the
five new images. proportion of correctly classified samples relative to the total
Seasonal Sentinel-2 imagery was acquired to create an image number of samples. Given the substantial variability in the areas
with temporal features. Cloud removal processing was applied of different LULC types within the study area, a completely
to each seasonal image, and the pixel values were fused by taking random selection of validation samples would result in a small
the median value of all pixel values at that location. For Sentinel- sample size for underrepresented LULC types. This could lead to
1 imagery, in addition to preprocessing steps such as eliminating an overall accuracy biased towards the dominant LULC types.
ground range detected (GRD) border noise and thermal noise, as Therefore, balancing the number of validation samples across
well as radiometric calibration and terrain correction, the study classes is preferable. To address the imbalance, we obtained
13264 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 17, 2024

Fig. 2. Framework of the research methodology.

samples for underrepresented LULC types as much as possible areas containing forests is approximately three times that of
and selected representative samples for dominant LULC types. those containing wetlands. As forests generally cover a larger
For example, in the case of wetlands, we selected samples within area, we therefore limited the number of forest samples in each
each catchment area, wherever available. Due to the absence of catchment area. The resulting sample difference is due to the
wetlands in some catchment areas, the number of catchment absence of wetlands in some catchment areas, which cannot be
LEI et al.: IMPACT OF LAND USE/LAND COVER CHANGES ON URBAN FLOODING 13265

further adjusted without compromising the representativeness of TABLE II

FEATURES EXTRACTED FROM SENTINEL 1/2 DATA FOR RANDOM
the forest samples across the study region. Therefore, even after FORESTS CLASSIFICATION
balancing the interclass quantity of validation samples, the gap
still exists. The forest sample was the largest, approximately 16
square kilometers, while the wetland sample was the smallest,
approximately 5.7 square kilometers. The total number of pixels
used for validation in each image should be greater than 100 000
and evenly distributed.
2) Classification Using Sentinel-1/2 Data: Sample selection
here is consistent with that of Landsat imagery, with eight
LULC types (farmland, orchard, forest, grassland, bare land,
built-up land, water, and wetland) of samples selected by visual
interpretation on Google Earth.
During preliminary feature selection, bands 2, 3, 4, 5, 6, 8,
11, and 12 were selected for the Sentinel-2 image, and then
ten spectral indices such as normalized difference vegetation
index (NDVI), normalized difference water index (NDWI),
soil adjusted vegetation index (SAVI), enhanced vegetation in-
dex (EVI), land surface water index (LSWI), bare soil index
(BSI), and 18 texture features such as angular second moment
(ASM), contrast, correlation, variance, inverse differential mo-
ment (IDM), sum average were calculated for each image [48].
The above indices were combined with the original bands to
form a new image. We selected the backscatter intensity of two
polarization modes, VV and VH, for the Sentinel-1 image, and
calculated six indices such as ratio and radar forest degradation
index (RFDI), which were combined into a new image. In
addition, we obtained four temporal features of each index in
the temporal image: mean, maximum, minimum, and standard
deviation.
Random forest can calculate out-of-bag (OOB) error to eval-
uate the importance of each feature [46]. Initially, features were method considers four main characteristics of runoff producing
inputted into the random forest in batches to calculate and rank basins, namely, land use, soil type, hydrological conditions, and
their importance. The ten least important features were replaced antecedent water conditions (AMC).
or removed in each iteration until the classification accuracy no The formula for calculating surface runoff through the SCS-
longer improved. CN model is
In summary, a total of 79 features were selected for Sentinel-2 
−Ia )2
optical imagery, including spectral, index, and texture features. Q = P(P−I a +S
(P > Ia )
(1)
For Sentinel-1 radar imagery, 16 features were selected, includ- Q=0 (P ≤ Ia )
ing terrain scattering characteristics and indices. Additionally,
where Q is the depth of surface runoff; P is the precipitation at
the DEM was included, resulting in a total of 96 features, as
a time; S is the maximum possible retention capacity; Ia is the
shown in Table II. All image bands were resampled to 10-m
initial loss value of rainfall. All units are in millimeters.
spatial resolution.
The formula for calculating the initial loss of rainfall is
In this study, the classification was carried out with the de-
cision tree number starting from 0 to 150 with a step size of 5. Ia = λS (2)
The decision tree number that yielded the best performance was
where λ is dimensionless, usually taken as 0.2.
selected as the final parameter for the random forest, and it was
The formula for calculating the maximum possible retention
set to 80 based on the test.
capacity is
25400
C. SCS-CN Runoff Modeling S= − 254 (3)
CN
The SCS-CN model, produced by the United States Soil and where CN is a dimensionless parameter that ranges from 0 to
Water Conservation Bureau, reflects the impact of varied land 100.
uses and soil types on surface runoff in watersheds [49]. The There are three grades of AMC, AMC III (wet), AMC II
model requires relatively few input data, and the simulation (normal), and AMC I (dry), which are based on the Antecedent
results are highly reliable [50], [51], hence it is widely used in the Precipitation Index (API), i.e., the total amount of rainfall (mm)
calculation of surface runoff in coastal areas [22]. The SCS-CN in the previous five days [52]. Based on the rainfall conditions
13266 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 17, 2024

TABLE III
CN VALUE OF ANTECEDENT MOISTURE CONDITION LEVEL 2 (AMC II)

in the study area, the normal AMC II conditions, which are D. Flooding Extent Estimation
commonly used in local studies, are chosen in this study to
The process of each runoff from the urban surface converging
calculate the CN value. The CN value is an empirical relationship to the corresponding runoff unit outlet is called confluence.
obtained through statistical analysis based on measured data.
As the study area is relatively large, watershed extraction and
The subsequent extensive application results have also proven watershed division are required [58]. Taking the river network
its rationality. This study refers to the study by Peng et al. [53] as input and combining with the locations of large-scale water
for the CN values set for different LULC types, which updates
projects such as reservoirs and dams within the watershed, the
the corrections of relevant research in the Pearl River Delta study site is divided into 160 catchment areas.
and surrounding regions based on the CN value lookup table Based on the runoff in each catchment area, the accumulated
provided by Mishra and Singh [54]. Finally, the CN values
water in that area is obtained. Next, the range and depth of flood-
corresponding to different underlying surface types under the ing were simulated using the local equal volume method. The
conditions of early medium soil moisture in the Greater Bay method treats rainfall as a passive submerged state, regardless
Area were determined as shown in Table III.
of the actual process of convergence, and fills each depression
Due to the significant impact of terrain slope changes on with the total amount of runoff according to the surface elevation
surface runoff, when using the SCS-CN model, slope correction
from high to low [59]. The formula is
is also necessary for the CN value. The Williams slope correction
formula [55] and Huang slope correction formula [56] are com- 
monly used, between which the Huang slope correction formula W = [Ew (x, y) − Eg (x, y)]dδ (6)
is more suitable for mountainous areas. Since the Greater Bay A
Area is not mountainous, the Williams slope correction formula
is adopted in this article, as follows: where W represents the total volume of water; A is the area of
the flooded area; Ew (x, y) is the water elevation of a certain
CN3 = CN2 ∗ exp[0.00673 ∗ (100 − CN2)] (4) point; Eg (x, y) is the elevation of a certain ground point; dδ the
flooded units.
where CN2 is the initial value of CN collected above; CN3 is Because the flow rate of flooding and ponding in urban rain-
the curve value at another degree of wetting storm is slow, the submerged water surface can be approximated
as a plane, and the formula is
(CN3 − CN2) 
CN2s = ∗ [1 − 2 ∗ e−13.86∗slope ] + CN2 (5)
3 W = [Ew − Eg (x, y)]dδ (7)
A
where CN2s is the CN value after slope correction; slope is
calculated using SRTM30 m data in the GEE platform; CN3 is where Ew is the unified water surface elevation of the region.
the curve value obtained from the previous formula for another Because the DEM data representing the surface elevation is a
degree of wetting. discrete regular grid, the formula is then discretized as follows:
Rainfall in the SCS-CN model is obtained through rainstorm
design. For the rainstorms in this article, the designed scenarios N

of rainfall recurrence periods are 10 years, 20 years, 50 years, W = [Ew − Eg (i)]Δσ (8)
and a century, and all last for 60 min. Note we only need the i=1
final flooded area but not the rainfall process and rainfall tem-
poral distribution. To improve the spatial rationality of rainfall, where Δσ is the area of the grid unit; N is the total grids number
Thiessen polygons are used to divide meteorological stations in the flooded area; Eg (i) is the elevation of the ith grid.
in each city [57]. The rainfall within each Thiessen polygon is After solving Ew for each catchment area, the submerged and
calculated using the rainstorm intensity formula corresponding nonsubmerged areas can be exported as elevation images. Then,
to the administrative region to which the meteorological station the correlation between LULC and inundation areas over the
belongs. decade can be statistically analyzed.
LEI et al.: IMPACT OF LAND USE/LAND COVER CHANGES ON URBAN FLOODING 13267

TABLE IV
LULC CLASSIFICATION ACCURACY AND KAPPA COEFFICIENT

E. Statistical Analysis TABLE V

LANDSAT 8 CLASSIFICATION RESULTS OF LULC (km2 )
The entire study area was divided into 11 sections by ad-
ministrative regions, and the LULC area and flooded area were
counted within each section. The area of each administrative
region varied greatly, hence the data needed to be normalized.
Therefore, the proportion of each type of LULC area and the
proportion of flooded area under each rainfall scenario were
calculated for each administrative region over the years. The area
proportions will be used to calculate the correlation coefficients
between the eight LULC categories and the flooded area under
the four rainfall scenarios.
Since our data do not follow a bivariate normal distribution, The area occupied by each classified category can be found in
it is not appropriate to use the Pearson correlation coefficient Table V. It shows that in the past decade, the farmland area in the
calculation [60]. In addition, we do not assume a linear relation- Greater Bay Area has shown a downward trend, while the areas
ship between LULC and flooded area, but speculate that there is of water, wetland, and bare land have remained relatively stable.
a monotonic correlation between the two. Thus, Spearman rank Although bare land may have undergone positional changes due
correlation is a more appropriate measurement in our case [60]. to human activities such as cultivation and construction, the total
The Spearman rank correlation coefficient [61] can be re- area has not changed much. The land use in the built-up area
garded as a nonparametric version of the Pearson correlation increased and peaked around 2018, and then decreased. Since
coefficient. It is calculated between the ranks of two variables 2015, the grassland area has been decreasing, while the forest
instead of their actual values [62]. The Spearman rank correla- and orchard areas have experienced significant fluctuations.
tion coefficient ranges from −1 to 1, with closer to 0 indicating
a lower correlation. B. Urban Flooding Analysis
As an example, the formula for calculating the correlation The designed rainfall scenarios are with recurrence periods
coefficient (rs ) between forest and the flooded area under the of 10 years, 20 years, 50 years, and a century. The runoff is
once in a century rainfall scenario is (9) shown at the bottom categorized into Level 1 (0–30 mm), Level 2 (30–50 mm), Level
of this page, where x is the proportion of forest area; y is the 3 (50–70 mm), Level 4 (70–85 mm), and Level 5 (85–100 mm).
proportion of flooded area, R(x) and R(y) are the rank of x and The runoff levels in the Greater Bay Area under the four rainfall
y, respectively; Ss is the sum of squared differences of ranked scenarios obtained by using Sentinel-1/2 data are shown in
variables x and y (xi , respectively, yi ); n is 55. Other correlation Fig. 4.
coefficients of LULC can be similarly calculated. From the runoff map of the 1-in-10-year rainfall scenario, the
areas with higher runoff depths in the Greater Bay Area are: the
IV. RESULTS AND ANALYSIS junction of Jiangmen, Zhongshan, and Zhuhai, the junction of
Foshan and Zhaoqing, the south coast of the Greater Bay Area,
A. LULC Changes and the area around the estuary of the Pearl River. In the runoff
Using the 2020 Landsat 8 image, the random forest classifier map of the 20-year scenario, the overall runoff is deepened, with
was trained and tested, with an overall accuracy and a kappa the runoff around the estuary of the Pearl River deepening into
coefficient of 89.97% and 0.87. The overall accuracy of each a circular pattern, while the runoff at the junction of Foshan and
Landsat image exceeded 87%, with kappa coefficients above Zhaoqing deepens into the cities. With the increase of rainfall,
0.83, as shown in Table IV. Based on Sentinel-1/2 images, the the runoff at the junction of Jiangmen, Zhongshan, and Zhuhai
overall accuracy of the LULC classification was 93.20%, and has the greatest depth, mostly reaching Level 5. Meanwhile, the
the kappa coefficient was 0.92. The LULC results obtained by distribution of deep runoff areas in the estuary of the Pearl River
the two data sources are shown in Fig. 3. spans a wide range, with runoff depths also reaching Level 4 to

n    
1
n i=1 R (xi ) − R(x) − R (yi ) − R(y)
rs =  (9)
n  2 n  2
1 1
n i=1 R (xi ) − R(x) ∗ n i=1 R (yi ) − R(y)
13268 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 17, 2024

Fig. 3. Classification results based on Landsat and Sentinel-1/2 data. (a)–(e) Results of Landsat 8 in 2013, 2015, 2018, 2020, and 2022, respectively. (f) Result
of Sentinel-1/2 in 2022.

Level 5 in most of Dongguan, Shenzhen, and Hong Kong in the TABLE VI

FLOODED AREA WITH RAINFALL SCENARIOS OF DIFFERENT RECURRENCE
eastern part of the estuary. PERIODS (km2 )
Using the local equal volume method within each watershed,
the flooded area under different rainfall scenarios was deter-
mined and presented in Fig. 5. It illustrates the inundation situa-
tion in various administrative regions. Zhaoqing in the northwest
and Huizhou in the northeast of the Greater Bay Area are affected
by hilly terrains, resulting in discrete blocks of flooded areas.
Jiangmen and the Pearl River estuary in the southwest, as well as
Dongguan, Shenzhen, Hong Kong, and Macau along the eastern declining in 2022. The flooded area of the once-in-100-year
coast, exhibit coastal and riverine distributions of affected areas, rainfall scenario only decreases from 2015 to 2018, with a larger
with less severe spreading and diffusion. The most severely area in 2015 and 2022.
affected areas are Guangzhou, Foshan, Zhongshan, and Zhuhai,
located in the central part and the western part of the Pearl River
estuary. They exhibit large-scale inundation, with numerous C. Analysis of the Influence of LULC Change on
small blocks of inundated areas forming extensive coverage. Urban Flooding
These areas are most affected by urban flooding. For the five LULC classification results obtained from Landsat
Table VI lists the flooded areas under the four rainfall scenar- 8 data over a decade, the flooded areas of 11 administrative
ios. As can be seen, over the past decade, flooding trends have regions were calculated, and 55 sets of data were obtained for
varied between scenarios, due to the fluctuation of the terrain and each rainfall scenario. Table VII shows the flooded area of each
changes in the land. The flooded area of the 1-in-10-year rainfall administrative region under the once in a century rainfall sce-
scenario increases abruptly from 2013 to 2015 and remains high. nario. After combination and normalization of the data from the
Then, until 2022, it decreases to the level of 2013. The flooded 5 epochs across the decade, the correlation coefficients between
area of the 1-in-20-year rainfall scenario is relatively stable. The the proportion of flood area and the proportion of each LULC
flooded area of the once-in-50-year rainfall scenario increases are calculated for each of the rainfall scenarios, as shown in
from 2013 to 2015, then stabilizes for the next three years before Table VIII. Fig. 6 shows the scatter charts of the proportions of
LEI et al.: IMPACT OF LAND USE/LAND COVER CHANGES ON URBAN FLOODING 13269

Fig. 4. Runoff levels in 2022 derived from Sentinel-1/2 data. (a)–(d) Results of once-in-10-year rainfall, once-in-20-year rainfall, once-in-50-year rainfall and
once-in-a-century rainfall, respectively. Level 1 (0–30 mm); Level 2 (30–50 mm); Level 3 (50–70 mm); Level 4 (70–85 mm); Level 5 (85–100 mm).

TABLE VII
FLOODED AREA (km2 ) AND PROPORTION (%) OF EACH ADMINISTRATIVE REGION UNDER A ONCE-IN-A-CENTURY RAINSTORM

flood area and LULC types under the once-in-a-century rainfall the highest. A positive correlation coefficient indicates that an in-
scenario. crease in the corresponding LULC type will lead to an increase in
The results show that water, built-up land, farmland, and the urban flooding area under the same precipitation conditions,
wetland have positive correlation coefficients, among which while a negative correlation coefficient indicates the opposite.
water has the highest one, whereas bare land, grassland, orchard, The larger the correlation coefficient, the stronger the effect of
and forest have negative correlation coefficients, and forest has the LULC type on urban flooding. In all the significance tests of
13270 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 17, 2024

Fig. 5. Flooded areas under different rainfall scenarios. (a)–(e) Results of the five years based on Landsat 8 data. (f) Result based on the 2022 Sentinel-1/2 data.

TABLE VIII
CORRELATION COEFFICIENT BETWEEN DIFFERENT LULC AND FLOODED AREA

correlation coefficients, only the p-value of farmland was greater It can be seen that bare land, grassland, orchard, and forest all
than 0.05. It is considered that there is no significant correlation have a certain capacity to reduce surface runoff and flooding.
between the area of farmland and the area of inundation. Beside Among them, forest has the best flood control and storage
that, the effect of water, bare land, built-up, grassland, orchard, capacity, followed by grassland and orchard, while bare land has
forest on flooded area was significant at the 0.01 level, and the the lowest capacity. The increase in water significantly increases
effect of wetland was significant at the 0.05 level. the area of urban floods, thus, the water level should be carefully
LEI et al.: IMPACT OF LAND USE/LAND COVER CHANGES ON URBAN FLOODING 13271

Fig. 6. Scatter plot of proportion of LULC and flooded area in each administrative region.

monitored. The expansion of built-up land also causes a rise in

the flooding area. There is no significant correlation between D. Comparison of Experimental Results From Different
farmland and flooding. The correlation of wetland was small, Data Sources
probably because it has a small area and few changes. However, A comparison is made between the classification results of
the role of wetland in regulating flooding cannot be denied. If Landsat and Sentinel data, as shown in Table IX. The comparison
water is converted into wetland, there can be a notable reduction of their flooded area results is shown in Table X. The columns
in surface runoff, which is obviously easier than converting water of these two tables represent the results based on Landsat data,
into forest. and the rows on Sentinel data.
13272 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 17, 2024

TABLE IX
COMPARISON OF CLASSIFICATION RESULTS FROM DIFFERENT DATA SOURCES (%)

TABLE X
COMPARISON OF FLOODED AREAS USING LANDSAT AND SENTINEL DATA (%)

Comparing the classification results of Landsat 8 with

Sentinel-1/2 data, it can be observed that there is high consis- V. DISCUSSION
tency in water, built-up land, orchard, and forest between the two. Previous studies have shown that built-up land increases the
However, the bare land, grassland, and wetland in the Sentinel impermeability of the surface, thereby increasing surface runoff
results are occasionally classified as farmland in the Landsat 8 and leading to more severe urban flooding, while forest can
results. This is because the Sentinel-1/2 data have higher tempo- reduce surface runoff and hence the risk of flooding [63], [64].
ral resolution and can utilize temporal features for classification, This is consistent with the results of this study. However, we
while the cloudy and rainy climate in southern China during the made a more fine-grained categorization of vegetation, revealed
summer limits the availability of sufficient temporal features the difference in the flood mitigation capacity of forest, orchard,
in Landsat 8 data. Additionally, during the harvesting period, and grassland, and also quantified the correlation between the
farmland and bare land appear similar, making them prone to eight LULC types and flooding. More detailed categorization
misclassification when Landsat 8 data are used. The consistency and quantification can be beneficial for accurate modeling and
in the classification of wetland between the two sets of results assessment of flood risk. For example, the impact of vegetation
is relatively low, as wetland can appear or disappear seasonally has been accounted for in some flood hazard risk assessments,
with water level changes, and better classification results can but these studies tend to lump all vegetation together [65], [66],
be achieved when temporal features are available. The built-up [67]. In this study, we have demonstrated that different vegeta-
land in the Sentinel-1/2 results is also prone to being classified tion resists flooding differently and quantitatively assessed the
as farmland in the Landsat 8 results. This is because most of the relationship between different vegetation types and flood hazard,
bands in Sentinel-1/2 have higher spatial resolution, allowing for which is useful for a more accurate assessment of flood risk. The
the differentiation of scattered built-ups near farmland, while at results of the relationship between LULC and urban flooding
30-m resolution, such built-ups are easily mixed within farmland serve as a reference for urban flood control decision-making
pixels in Landsat 8 data. and planning in the Greater Bay Area, while the analysis of
Table X shows that the two data sources have produced highly different remote sensing data sources provides evidence that the
consistent results, with more than 99% agreement for unflooded variation of input data sources has an insignificant impact on
areas and around 98% agreement for flooded areas under all four the final results. There is only an inconsistency of 1%–2% when
rainfall scenarios, with differences mainly observed in scattered using mixed data with apparent different spatial and spectral
flooded areas in relatively flat regions and at the edges of a few resolutions.
rivers and lakes in mountainous areas. There is a tendency to In terms of LULC classification, this study used the random
obtain more flooded areas with Landsat data than with Sentinel forest method, which has been widely used in experiments and
data. Sentinel data have higher spatial and spectral resolutions, has been proven to have high accuracy and stability [47]. Prasad
which can better distinguish complex and mixed LULC types in et al. [68] compared the classification accuracy of different
cities, such as grassland and farmland. algorithms when using Landsat 8 or fused Sentinel-1/2 data, and
LEI et al.: IMPACT OF LAND USE/LAND COVER CHANGES ON URBAN FLOODING 13273

the results showed that random forest classification is better than [7] G. Wang, J. Chen, C. Zhao, X. Zhou, and X. Deng, “Exploration of
ANN and most algorithms. However, with the rapid development the causality between area changes of green spaces and waterlogging
frequency in Beijing,” Phys. Chem. Earth, vol. 101, pp. 172–177, 2017.
of deep learning, complex neural networks have been applied to [8] Y.-F. Ning, W.-Y. Dong, L.-S. Lin, and Q. Zhang, “Analyzing the
LULC classification and achieved higher accuracy [69], [70]. causes of urban waterlogging and sponge city technology in China,” IOP
Future studies can try to use additional classification methods, Conf. Ser., Earth Environ. Sci., vol. 59, no. 1, 2017, Art. no. 012047,
doi: 10.1088/1755-1315/59/1/012047.
provided enough training data is available. [9] M. Ma et al., “Investigation of inducements and defenses of flash floods and
In terms of urban flooding area extraction, there is no urban urban waterlogging in Fuzhou, China, from 1950 to 2010,” Nat. Hazards,
drainage component in the model, limited by the difficulty in vol. 91, pp. 803–818, 2018.
[10] S. Du, A. Van Rompaey, P. Shi, and J. Wang, “A dual effect of urban
obtaining data. Urban drainage data in the study area are not expansion on flood risk in the Pearl River Delta (China) revealed by
publicly available. Quantitative assessment of urban drainage land-use scenarios and direct runoff simulation,” Nat. Hazards, vol. 77,
capacity on a large scale is still lacking and could be a future pp. 111–128, 2015.
[11] A. P. Cracknell, “The development of remote sensing in the last 40 years,”
study after overcoming the data source problem. Int. J. Remote Sens., vol. 39, pp. 8387–8427, 2018.
In addition, based on the proposed methodology in this study, [12] H. Huang, J. Legarsky, and M. Othman, “Land-cover classification using
the experiment can be applied to other regions, and the differ- radarsat and landsat imagery for St. Louis, Missouri,” Photogramm. Eng.
Remote Sens., vol. 73, no. 1, pp. 37–43, 2007.
ences in the impact of LULC on urban flooding in different [13] T. Zhou, Z. Li, and J. Pan, “Multi-feature classification of multi-sensor
climates can be analyzed. However, attention should be paid to satellite imagery based on dual-polarimetric sentinel-1A, landsat-8 OLI,
the subtle differences in surface infiltration relationships caused and hyperion images for urban land-cover classification,” Sensors, vol. 18,
no. 2, 2018, Art. no. 373.
by geological, soil, terrain slope, weather, and other factors in [14] D. L. Giandomenico, M. N. S. João, D. F. Salvatore, and M. Giuseppe,
different regions, taking into account the correction of CN values “Integrated use of sentinel-1 and sentinel-2 data and open-source machine
in different regions through field experiments. Expanding the learning algorithms for land cover mapping in a mediterranean region,”
Eur. J. Remote. Sens., vol. 55, no. 1, pp. 52–70, 2022.
scope of research could be a direction for future study. [15] B. Chen, B. Huang, and B. Xu, “Multi-source remotely sensed data
fusion for improving land cover classification,” ISPRS J. Photogramm.
Remote Sens., vol. 124, pp. 27–39, 2017. [Online]. Available: https://www.
VI. CONCLUSION sciencedirect.com/science/article/pii/S0924271616306463
This study examined the LULC changes in the Greater Bay [16] H. Gao et al., “A crop classification method integrating GF-3 Polsar and
Sentinel-2A optical data in the Dongting Lake Basin,” Sensors, vol. 18,
Area, China, over the past decade and simulated the extent of no. 9, 2018, Art. no. 3139. [Online]. Available: https://www.mdpi.com/
urban flooding under different intensities of rainfall. These two 1424-8220/18/9/3139
factors are combined to explore the influence of LULC change [17] J. Zhang, “Multi-source remote sensing data fusion: Status and trends,”
Int. J. Image Data Fusion, vol. 1, no. 1, pp. 5–24, 2010, doi: 10.1080/
on urban flooding and analyze the influence of using different 19479830903561035.
remote sensing data on the experiment results. It has been [18] P. Koudelak and S. West, “Sewerage network modelling in Latvia, use
found that water and built-up land show a positive correlation of infoworks CS and storm water management model 5 in Liepaja city,”
Water Environ. J., vol. 22, no. 2, pp. 81–87, 2008.
with urban flooding, while bare land, grassland, orchard, and [19] J. J. Wiles and N. S. Levine, “A combined GIS and HEC model for
forest exhibit a negative correlation. Meanwhile, wetland has a the analysis of the effect of urbanization on flooding; the Swan Creek
positive but insignificant correlation, and there is no significant watershed, Ohio,” Environ. Eng. Geosci., vol. 8, no. 1, pp. 47–61, 2002.
[20] H. Zhou and C. Wang, “Guangdong Dongguan city urban flooding causes
correlation for agricultural land. The Landsat and Sentinel based analysis and prevention measures,” China Flood Drought Manag., vol. 23,
urban flooding simulation results are highly consistent, with no. 4, pp. 70–71, 2013.
differences mainly in scattered flooded areas in flat terrain, [21] J.-W. Jung et al., “Water management practices and SCS curve numbers of
paddy fields equipped with surface drainage pipes,” Agric. Water Manage.,
ranging from 1% to 2%. The findings of this study can provide a vol. 110, pp. 78–83, 2012.
reference for the Greater Bay Area to make informed decisions [22] S. Jiao, X. Zhang, and Y. Xu, “A review of Chinese land suitability
for urban flooding mitigation. assessment from the rainfall-waterlogging perspective: Evidence from the
Su Yu Yuan area,” J. Clean. Prod., vol. 144, pp. 100–106, 2017.
[23] J. A. Reistetter and M. Russell, “High-resolution land cover datasets,
REFERENCES composite curve numbers, and storm water retention in the Tampa Bay,
FL region,” Appl. Geogr., vol. 31, no. 2, pp. 740–747, 2010.
[1] Z. Yin, J. Yin, S. Xu, and J. Wen, “Community-based scenario modelling [24] S. Verma, P. K. Singh, S. K. Mishra, V. P. Singh, V. Singh, and A.
and disaster risk assessment of urban rainstorm waterlogging,” J. Geogr. Singh, “Activation soil moisture accounting (Asma) for runoff estima-
Sci., vol. 21, no. 2, pp. 274–284, 2011, doi: 10.1007/s11442-011-0844-7. tion using soil conservation service curve number (SCS-CN) method,”
[2] X. Wu, D. Yu, Z. Chen, and R. L. Wilby, “An evaluation of the impacts J. Hydrol, vol. 589, 2020, Art. no. 125114. [Online]. Available: https:
of land surface modification, storm sewer development, and rainfall vari- //www.sciencedirect.com/science/article/pii/S0022169420305746
ation on waterlogging risk in Shanghai,” Nat. Hazards, vol. 63, no. 2, [25] X. Zhang, F. Meng, and N. Ding, “Application of SCS model to estimating
pp. 305–323, 2012, doi: 10.1007/s11069-012-0153-1. the quantity of runoff of small watershed in semi-arid or arid region,” Res.
[3] S. Roy, A. Bose, N. Singha, D. Basak, and R. C. Indrajit, “Urban wa- Soil Water Conserv, vol. 10, no. 4, pp. 172–174, 2003.
terlogging risk as an undervalued environmental challenge: An integrated [26] B. Xiao, Q.-H. Wang, J. Fan, F.-P. Han, and Q.-H. Dai, “Application
MCDA-GIS based modeling approach,” Environ. Challenges, vol. 4, 2021, of the SCS-CN model to runoff estimation in a small watershed with
Art. no. 100194. high spatial heterogeneity,” Pedosphere, vol. 21, no. 6, pp. 738–749,
[4] X. Wang, R. Jiang, J. Xie, Y. Zhao, F. Li, and J. Zhu, “Multiscale vari- 2011. [Online]. Available: https://www.sciencedirect.com/science/article/
ability of precipitation and their teleconnection with large-scale climate pii/S100201601160177X
anomalies,” J. Coast. Res., vol. 93, pp. 417–426, 2019. [27] A. Hagras, “Runoff modeling using SCS-CN and GIS approach in
[5] J. Yin, M. Ye, Z. Yin, and S. Xu, “A review of advances in urban flood the Tayiba valley basin, Abu Zenima area, south-west Sinai, Egypt,”
risk analysis over China,” Stoch. Environ. Res. Risk Assess., vol. 29, no. 3, Model. Earth Syst. Environ., vol. 9, pp. 3883–3895, 2023, doi: 10.1007/
pp. 1063–1070, 2015. s40808-023-01714-5.
[6] R.-S. Quan, “Rainstorm waterlogging risk assessment in central urban area [28] X. Liu and J. Li, “Application of SCS model in estimation of runoff from
of Shanghai based on multiple scenario simulation,” Nat. Hazards, vol. 73, small watershed in loess plateau of China,” Chin. Geogr. Sci., vol. 18, no. 3,
pp. 1569–1585, 2014. pp. 235–241, 2008.
13274 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 17, 2024

[29] S. Xie, W. Liu, Z. Yuan, H. Zhang, H. Lin, and Y. Wang, “Integrated risk [49] W. Boughton, “A review of the USDA SCS curve number method,” Soil
assessment of waterlogging in Guangzhou based on runoff modeling, AHP, Res., vol. 27, no. 3, pp. 511–523, 1989.
GIS and scenario analysis,” Water, vol. 14, no. 18, 2022, Art. no. 2899. [50] J. Du et al., “Assessing the effects of urbanization on annual runoff and
[Online]. Available: https://www.mdpi.com/2073-4441/14/18/2899 flood events using an integrated hydrological modeling system for Qinhuai
[30] Z. Chen, K. Li, J. Du, Y. Chen, R. Liu, and Y. Wang, “Three-dimensional river basin, China,” J. Hydrol., vol. 464-465, pp. 127–139, 2012.
simulation of regional urban waterlogging based on high-precision [51] Q. Weng, “Modeling urban growth effects on surface runoff with the
DEM model,” Nat. Hazards, vol. 108, no. 3, pp. 2653–2677, 2021, integration of remote sensing and GIS,” Environ. Manage., vol. 28,
doi: 10.1007/s11069-021-04793-8. pp. 737–748, 2001.
[31] C. Yang et al., “Spatiotemporal evolution of urban agglomerations in four [52] S. Fu, X. Wang, H. Wang, X. Wei, and A. Yuan, “Method of determining
major bay areas of US, China and Japan from 1987 to 2017: Evidence CN value in the SCS-CN method,” Arid Land Geogr., vol. 35, no. 3,
from remote sensing images,” Sci. Total Environ., vol. 671, pp. 232–247, pp. 415–421, 2012.
2019. [Online]. Available: https://www.sciencedirect.com/science/article/ [53] J. Peng, H. Wei, W. Wu, Y. Liu, and Y. Wang, “Storm flood disaster risk
pii/S0048969719311362 assessment in urban area based on the simulation of land use scenarios:
[32] M. Yan, J. C. L. Chan, and K. Zhao, “Impacts of urbanization on the A case of Maozhou watershed in Shenzhen city,” Acta Ecologica Sinica,
precipitation characteristics in Guangdong province, China,” Adv. Atmos. vol. 38, pp. 3741–3755, 2018.
Sci., vol. 37, no. 7, pp. 696–706, 2020, doi: 10.1007/s00376-020-9218-3. [54] S. K. Mishra and V. P. Singh, SCS-CN Method. Berlin, Germany: Springer,
[33] B. Liu, S. Chen, X. Tan, and X. Chen, “Response of precipitation to 2003, doi: 10.1007/978-94-017-0147-1_2.
extensive urbanization over the Pearl River Delta metropolitan region,” [55] J. R. Williams, “The erosion-productivity impact calculator (epic) model:
Environ. Earth Sci., vol. 80, no. 1, 2021, Art. no. 9, doi: 10.1007/ A case history,” Philos. Trans. R. Soc. Lond., B., Biol. Sci., vol. 329,
s12665-020-09299-3. no. 1255, pp. 421–428, 1990.
[34] D. Zhao, J. Zha, and J. Wu, “Changes in rainfall of different in- [56] M. Huang, J. Gallichand, Z. Wang, and M. Goulet, “A modification to the
tensities due to urbanization-induced land-use changes in Shenzhen, soil conservation service curve number method for steep slopes in the loess
China,” Clim. Dyn., vol. 56, no. 7, pp. 2509–2530, 2021, doi: 10.1007/ plateau of China,” Hydrol. Process., vol. 20, no. 3, pp. 579–589, 2006.
s00382-020-05601-y. [57] H. Derakhshan and N. Talebbeydokhti, “Rainfall disaggregation in non-
[35] L. Li et al., “Review of advances in urban climate study in the Guangdong- recording gauge stations using space-time information system,” Sci. Iran.,
Hong Kong-Macau greater bay area, China,” Atmos. Res., vol. 261, vol. 18, no. 5, pp. 995–1001, 2011.
2021, Art. no. 105759. [Online]. Available: https://www.sciencedirect. [58] J. F. O’Callaghan and D. M. Mark, “The extraction of drainage networks
com/science/article/pii/S016980952100315X from digital elevation data,” Comput. Vis. Graph. Image Process., vol. 27,
[36] K. M. Wai et al., “Observational evidence of a long-term increase in no. 2, pp. 323–344, 1984.
precipitation due to urbanization effects and its implications for sus- [59] L. Zhang, “Urban flooding analysis and modeling using GIS,” Acta Geogr.
tainable urban living,” Sci. Total Environ., vol. 599-600, pp. 647–654, Sinica, vol. 50, pp. 76–84, 1995.
2017. [Online]. Available: https://www.sciencedirect.com/science/article/ [60] P. Schober, C. Boer, and L. A. Schwarte, “Correlation coefficients:
pii/S0048969717311105 Appropriate use and interpretation,” Anesth. Analg., vol. 126, no. 5,
[37] Y. Li, W. Wang, M. Chang, and X. Wang, “Impacts of urbanization on pp. 1763–1768, 2018. [Online]. Available: https://journals.lww.com/
extreme precipitation in the Guangdong-Hong Kong-Macau greater bay anesthesia-analgesia/fulltext/2018/05000/correlation_coefficients__
area,” Urban Clim., vol. 38, 2021, Art. no. 100904. [Online]. Available: appropriate_use_and.50.aspx
https://www.sciencedirect.com/science/article/pii/S2212095521001346 [61] E. C. Fieller, H. O. Hartley, and E. S. Pearson, “Tests for rank correlation
[38] Z. Deng, Z. Wang, X. Wu, C. Lai, and Z. Zeng, “Strengthened tropical coefficients. I,” Biometrika, vol. 44, no. 3/4, pp. 470–481, 1957.
cyclones and higher flood risk under compound effect of climate change [62] G. Frapporti, L. A. M. Linnartz, and S. P. Vriend, “Spearmen-a dBase
and urbanization across China’s greater bay area,” Urban Clim, vol. 44, program for computation and testing of spearman rank correlation co-
2022, Art. no. 101224. [Online]. Available: https://www.sciencedirect. efficient distributions,” Comput. Geosci., vol. 17, no. 4, pp. 569–589,
com/science/article/pii/S2212095522001420 1991. [Online]. Available: https://www.sciencedirect.com/science/article/
[39] J. Zheng, H. Wang, and B. Liu, “Impact of the long-term precip- pii/009830049190115 T
itation and land use changes on runoff variations in a humid sub- [63] S. Y. Tola and A. Shetty, “Land cover change and its implication to
tropical river basin of China,” J. Hydrol. Reg. Stud, vol. 42, 2022, hydrological regimes and soil erosion in Awash river basin, Ethiopia:
Art. no. 101136. [Online]. Available: https://www.sciencedirect.com/ A systematic review,” Environ. Monit. Assess., vol. 193, no. 12, 2021,
science/article/pii/S2214581822001495 Art. no. 836, doi: 10.1007/s10661-021-09599-6.
[40] P.-J. Shi, Y. Yuan, J. Zheng, J.-A. Wang, Y. Ge, and G.-Y. Qiu, “The [64] P. E. Zope, T. I. Eldho, and V. Jothiprakash, “Impacts of urbanization
effect of land use/cover change on surface runoff in Shenzhen region, on flooding of a coastal urban catchment: A case study of Mumbai
China,” Catena, vol. 69, no. 1, pp. 31–35, 2007. [Online]. Available: https: city, India,” Nat Hazards (Dordr), vol. 75, no. 1, pp. 887–908, 2015,
//www.sciencedirect.com/science/article/pii/S0341816206000944 doi: 10.1007/s11069-014-1356-4.
[41] B. Liu, J. Chen, W. Lu, X. Chen, and Y. Lian, “Spatiotempo- [65] Y. Wang, J. Zhai, G. Gao, Q. Liu, and L. Song, “Risk assessment of
ral characteristics of precipitation changes in the Pearl River basin, rainstorm disasters in the Guangdong-Hong Kong-Macao greater bay area
China,” Theor. Appl. Climatol., vol. 123, no. 3, pp. 537–550, 2016, of China during 1990–2018,” Geomatics, Nat. Hazards Risk, vol. 13, no. 1,
doi: 10.1007/s00704-015-1375-4. pp. 267–288, 2022. doi: 10.1080/19475705.2021.2023224.
[42] J. Niu and J. Chen, “Terrestrial hydrological features of the Pearl River [66] Y. Li, Z. Zhang, S. Gong, M. Liu, and Y. Zhao, “Risk assessment of
Basin in South China,” J. Hydro-environ. Res., vol. 4, no. 4, pp. 279–288, rainstorm disasters under different return periods: A case study of Bohai
2010. [Online]. Available: https://www.sciencedirect.com/science/article/ Rim, China,” Ocean Coast. Manage., vol. 187, 2020, Art. no. 105107.
pii/S157064431000050X [67] Z. Zhu, S. Zhang, Y. Zhang, R. Yao, and H. Jin, “Integrating flood
[43] H. Wu, Z. Zhai, Z. Li, and Y. Liu, “Guangdong-Hong Kong-Macao greater risk assessment and management based on HV-SS model: A case study
bay area climate bulletin,” Guangdong Meteorological Bur., Guangzhou, of the pearl river delta, China,” Int. J. Disaster Risk Reduct., vol. 96,
China, 2019. [Online]. Available: https://www.hko.gov.hk/en/wxinfo/ 2023, Art. no. 103963. [Online]. Available: https://www.sciencedirect.
pastwx/2018/files/GD_HK_Mac_GBA_2018.pdf com/science/article/pii/S2212420923004430
[44] K. Wu and R. Zhao, “Soil texture classification and its application in [68] P. Prasad, V. J. Loveson, P. Chandra, and M. Kotha, “Evaluation and
China,” Acta Pedologica Sinica, vol. 56, pp. 227–241, 2019. comparison of the earth observing sensors in land cover/land use stud-
[45] Y. Shao and D. Shao, New Urban Storm Lntensity Formula in China. ies using machine learning algorithms,” Ecol. Inform., vol. 68, 2022,
Beijing, China: China Architecture & Building Press, 2014. Art. no. 101522. [Online]. Available: https://www.sciencedirect.com/
[46] L. Breiman, “Random forests,” Mach. Learn, vol. 45, no. 1, pp. 5–32, science/article/pii/S1574954121003137
2001. [69] M. U. Rasheed and S. A. Mahmood, “A framework base on deep neural
[47] M. Junaid, J. Sun, A. Iqbal, M. Sohail, S. Zafar, and A. Khan, “Mapping network (DNN) for land use land cover (LULC) and rice crop classification
LULC dynamics and its potential implication on forest cover in Malam without using survey data,” Clim. Dyn., vol. 61, pp. 5629–5652, 2023,
Jabba region with landsat time series imagery and random forest classifi- doi: 10.1007/s00382-023-06874-9.
cation,” Sustainability, vol. 15, no. 3, 2023, Art. no. 1858. [70] M. Digra, R. Dhir, and N. Sharma, “Land use land cover classification of
[48] R. M. Haralick, K. Shanmugam, and I. Dinstein, “Textural features for remote sensing images based on the deep learning approaches: A statistical
image classification,” IEEE Trans. Syst. Man Cybern, vol. SMC- 3, no. 6, analysis and review,” Arab. J. Geosci., vol. 15, no. 10, 2022, Art. no. 1003,
pp. 610–621, Nov. 1973. doi: 10.1007/s12517-022-10246-8.
LEI et al.: IMPACT OF LAND USE/LAND COVER CHANGES ON URBAN FLOODING 13275

Yuqi Lei received the B.S. degree in remote sensing Jon P. Mills received the B.Sc. (Hons.) degree in
science and technology in 2024 from China Uni- surveying science and the Ph.D. degree in digital pho-
versity of Geosciences, Wuhan, China, where she togrammetry from Newcastle University, Newcastle
is currently working toward the master’s degree in upon Tyne, U.K., in 1993 and 1996, respectively.
photogrammetry and remote sensing. He is currently a Professor of geomatic engineering
Her research interests include change detection, with the School of Engineering, Newcastle Univer-
image processing, laser scanning, and their applica- sity, where he leads the UKRI Centre for Doctoral
tions in smart cities and digital twins. Training (CDT) in Geospatial Systems. His research
interests include developing optimized geospatial so-
lutions for measurement, mapping, monitoring, and
parameterization of engineering models.
Dr. Mills is a Fellow of the International Society for Photogrammetry and
Remote Sensing (ISPRS, 2022) and has previously served as an elected Member
of ISPRS Council (2012–2016), as President of ISPRS Technical Commission
Hui Cao received the B.S. degree in geographical V on close range image measurement techniques (2008-.2012), and as Chair
information science from Central South University, of ISPRS Working Group I/4 on airborne digital sensor systems (2004–2008).
Changsha, China, in 2022. He is currently working During the period 2017–2023, he was the Chair of Commission 1, primary data
toward the master’s degree in software engineering acquisition, for European Spatial Data Research (EuroSDR).
with China University of Geosciences, Wuhan, China.
His research interests include deep learning,
3-D change detection, image processing, and LiDAR
point cloud processing. Wen Xiao (Member, IEEE) received the B.S. de-
gree in geodesy and geomatics from Wuhan Univer-
sity, Wuhan, China, the M.S. degree (cum laude) in
geoinformatics from the Faculty of Geo-Information
Science and Earth Observation (ITC), University of
Twente, Enschede, The Netherlands, and the Ph.D.
degree in geoinformation science and technology
Xuanyan Zhou received the B.S. degree in remote from the National Institute of Geographical Informa-
sensing science and technology from China Univer- tion and Forestry (IGN), Université Paris-EST, Paris,
sity of Geosciences, Wuhan, China, in 2024. He is France, in 2015.
currently working toward the master’s degree in re- He has been a Lecturer and a University Research
source and environment with Sun Yat-sen University, Fellow with Newcastle University, Newcastle upon Tyne, U.K. He is currently a
Guangzhou, China. Professor with the China University of Geosciences, Wuhan, China. His research
His research interests include related applications interests include 3-D mapping, laser scanning, photogrammetric computer vi-
such as digital image processing, hydrological remote sion, and their applications in smart cities and digital twins.
sensing, photogrammetry, etc. Dr. Xiao is also a Council Member of the Remote Sensing and Photogram-
metry Society (RSPSoc), U.K.