Journal of Hydrology 573 (2019) 524–533

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

Research papers

Rainfall-runoff simulation in karst dominated areas based on a coupled T

conceptual hydrological model
⁎
Qing Zhoua, Lu Chena, , Vijay P. Singhb, Jianzhong Zhoua, Xiaohong Chenc, Lihua Xiongd
a
College of Hydropower & Information Engineering, Huazhong University of Science & Technology, Wuhan 430074, PR China
b
Dept. of Biological and Agricultural Engineering and Zachry Dept. of Civil Engineering, Texas A & M University, TAMU, College Station, TX 77843-2117, USA
c
State Key Laboratory of Water Resources and Hydropower Engineering Science, Wuhan University, Wuhan 430072, PR China
d
Center for Water Resource and Environment, Sun Yat-Sen University, Guangzhou 510275, PR China

A R T I C LE I N FO A B S T R A C T

This manuscript was handled by G. Syme, Rainfall-runoff processes in the karst dominated regions are of great importance. However, not all areas in a
Editor-in-Chief, with the assistance of Craig T. karst-dominated river basin are covered by typical limestone and both limestone and soil cover may exist. This
Simmons, Associate Editor study proposes a coupled conceptual hydrological model for simulating the rainfall-runoff processes in karst-
Keywords: dominated areas. The model, named as K-XAJ, couples the traditional Xinanjiang (XAJ) model and a two re-
Rainfall-runoff simulation servoir-based karst model for simulating runoff in both the karst area and the non-karst area in the Lijiang River
Karst areas basin. Simulated results demonstrated that the proposed K-XAJ model satisfactorily simulated the rainfall-runoff
Xinanjiang model processes. Compared with the traditional XAJ model, the K-XAJ model produced better forecasts. The peak flow
Karst reservoirs
predicted by the K-XAJ model was larger than that by the XAJ model, but the interflow routed by the K-XAJ
Lijiang River basin
model was significantly smaller than that by the XAJ model. The magnitude of groundwater predicted by the K-
XAJ model was greater. This study thus provides a new way to simulate rainfall-runoff processes in karst areas.

1. Introduction flow, generating temporal variability of recharge and hydraulic con-

nectivity (Bonacci, 2001; Bailly-Comte et al., 2009). Generally, a karst
Karst is formed from the dissolution of soluble rocks, such as aquifer is composed of three interacting systems: The uppermost layer
limestone, dolomite, and gypsum, and is also characterized by under- of the carbonate rock is referred to as the epikarst where fractures
ground drainage systems with sinkholes and caves. Karst areas cover develop, followed by an infiltration zone, to a few meters below the
about 15% of the world’s land area or about 2.2 million km2 and are ground surface, where both unsaturated and saturated flow may occur,
home for around 1 billion people (17% of the world’s population) (Yuan and finally, the low permanently saturated karst zone composed of a
and Chai, 1988). About 25% of the world’s population received its highly organized and hierarchized drainage conduit system in connec-
water supply largely or entirely from karst waters, including deep tion with a network of secondary drains (Labat et al., 1999).
carbonate aquifers (Ford and Williams, 2007). These areas are located To simulate the discharge from a karst area, a variety of karst
in Africa, Asia, Europe, South and North America and Oceania. The models have been developed and applied around the world, which can
karst terrain covers nearly 35% of the whole of Europe, so the karst be classified into three kinds: (1) empirical, (2) distributed, and (3)
processes are significant components of the physical geography of conceptual. Empirical (black-box) models regard the whole karst
Mediterranean basins (Daly et al., 2002). One of the largest continuous system as a black box without considering the structure and char-
karst areas in the world is located in Yunnan-Guizhou Plateau of acteristics of the karst aquifer, and develop a relationship between
southwest China. It covers about 540,000 km2 and has a population of rainfall and runoff. Linear as well as non-linear functions have been
nearly 100 million people (Zhang et al., 2010). used to establish the relationship between rainfall and runoff (Dreiss,
Since the karst areas are so large, it is of great importance to study 1989a,b; Labat et al., 1999; Lambrakis et al., 2000). Neural networks
their rainfall-runoff processes. Because of the characteristics of soil, have also been used to simulate flow dynamics in karst areas
rocks and so on, the rainfall-runoff process in karst areas is very dif- (Paleologos et al, 2013; Chen et al., 2014a,b). The advantages of em-
ferent from that in non-karst areas,. The complex geometry of the pirical models are their ease to implement with fewer parameters and
network of conduits is sometimes the cause of surface and groundwater applicability to areas where information is lacking. The drawbacks are

⁎
Corresponding author.
E-mail address: [email protected] (L. Chen).

https://doi.org/10.1016/j.jhydrol.2019.03.099
Received 11 February 2019; Received in revised form 27 March 2019; Accepted 28 March 2019
Available online 29 March 2019
0022-1694/ © 2019 Published by Elsevier B.V.
Q. Zhou, et al. Journal of Hydrology 573 (2019) 524–533

Fig. 1. Structure of the K-XAJ model.

that the models have no physical meaning and the parameters in the same river basin where runoff generation mechanisms in limestone and
function also can also vary and may themselves be functions of char- soil cover areas are entirely different. Therefore, the runoff generation
acteristics of the aquifer. Due to the lack of process representation, process in these kinds of areas should be considered separately.
black-box models cannot be used to evaluate and manage the impact of This paper aims at developing a conceptual hydrological model,
human activities on spring flows. named Karst-Xinanjiang (K-XAJ) model, which is suitable for karst
The second type is the distributed models, which consider temporal areas and reflects the physical characteristics of the karst aquifer for
and spatial variability. Wang and Brubaker (2014) proposed a non- runoff simulation. This hydrological model coupled the traditional
linear modification of groundwater algorithm in SWAT (ISWAT) model Xinanjiang (XAJ) model and a two- reservoir karst model, which can
that improved recession and low-flow simulation. Nikolaidis et al. simulate runoff in both karst area and non-karst areas. The parameters
(2013) developed a reservoir model and linked it with SWAT to simu- of the proposed model were calibrated using a multi-objective optimi-
late karst behavior and recharge of springs of Koiliaris basin in Crete. zation algorithm. The Lijiang River basin in southwestern China was
Recently, Malago et al. (2016) used the KSWAT model (SWAT model selected as a case study and the proposed K-XAJ model and the tradi-
and karst-flow model) for the quantification of a spatially and tempo- tional XAJ model were used to simulate rainfall-runoff in karst-domi-
rally explicit hydrologic water balance of karst-dominated geomor- nated areas. Simulated results were compared with observed values.
phology in Crete Island. However, this model needs more data sets and
contains more parameters that need to be calibrated. It is often difficult 2. Review of XAJ model
to obtain enough meteorological data in karst areas which restrict its
application in practice. The Xinanjiang (XAJ) model is a conceptual hydrologic model de-
The third type is the conceptual models, which conceptualizes a veloped by Zhao et al. (1980) for flood simulation in the Xinan River
karst aquifer as a configuration of internal storages and pathways, basin, China and has since been widely used in China for flood simu-
while physical relationships are not considered explicitly but are re- lation in humid and semi-humid regions (Zhao, 1992). This model has
presented in general terms through conceptualization of the aquifer also been used for almost all large river basins in China, including the
(Shi et al., 2013). Fleury et al. (2007) used a three-reservoir model to Yellow River, Yangtze River, Huaihe River, and so on. Usually, a large
simulate soil, slow discharge, and rapid discharge of a karstic aquifer in watershed is divided into a set of sub-basins to capture the spatial
southern France. Tritz et al. (2011) proposed a two-reservoir con- variability of precipitation, evaporation, and the underlying surface.
ceptual model with a hysteretic, non-linear transfer function to model The inputs of the XAJ model are the average areal rainfall and eva-
the behavior of a karst catchment. Charlier et al. (2012) proposed a poration in each sub-basin. The output of the model is streamflow at the
numerical karst model to simulate discharge as well as concentration of watershed outlet for each sub-basin.
non-conservative tracers during flood events in French Jura. Using a The XAJ model is composed of four main modules, namely eva-
conceptual rainfall–runoff model, Jukic et al. (2009) simulated poration module, runoff generation module, runoff partition module,
groundwater flow in the Jadro Spring catchment. A conceptual model and runoff routing module. The structure of the XAJ model is given in
appears to be a good compromise between an empirical model and a Fig. 1. The physical meanings of the model parameters are given in
distributed model, which has relatively fewer parameters than the Table 1.
distributed model and can generally reflect physical relationships of the
karst aquifer. (1) Evaporation module
Not all karst areas are covered by typical limestone. For example,
the Island of Crete in Greece has an area of 8336 km2 of which about In this model, a three-layer evapotranspiration scheme is used. The
2730 km2 representing 30% of the total area, is carbonate rocks soil is divided into three layers, namely upper layer, lower layer and
(Chartzoulakis et al., 2001). Koiliaris River watershed, with the area of deep layer, of which actual storage are WU, WL, WD respectively. And
132 km2, is located in the northwestern part of Crete, of which the karst the storage capacity of the three-layer soil are UM, LM, and DM. The
area is at least 50 km2 (Moraetis et al., 2010). Guizhou Province in actual evapotranspiration E is the sum of evapotranspiration in the
Southwest China has one of the largest and continuous karst areas in the upper layer (EU), evapotranspiration in the lower layer (EL) and eva-
world, of which 73% areas are covered by limestone (Li et al., 2003; potrans-piration in the deep layer (ED). EU occurs at the potential rate;
and Li et al., 2006). Generally, both limestone and soil cover exist in the EL occurs at the modified rate which is calculated by multiplying the

525
Q. Zhou, et al. Journal of Hydrology 573 (2019) 524–533

Table 1
Calibrated parameters of the K-XAJ and XAJ models.
Parameter Physical meaning K-XAJ XAJ

UM(mm) Averaged soil moisture storage capacity of the upper layer 19.79 20.00
LM(mm) Averaged soil moisture storage capacity of the lower layer 82.08 75.73
DM(mm) Averaged soil moisture storage capacity of the deep layer 52.80 10
B Exponential of the distribution to tension water capacity 0.40 0.39
IM Percentage of impervious and saturated areas in the catchment 0.04 0.03
K Ratio of potential evapotranspiration to pan evaporation 0.78 0.84
C Coefficient of the deep layer 0.20 0.20
SM(mm) real mean free water capacity of the surface soil layer 35.59 32.42
EX Exponent of the free water capacity curve influencing the development of the saturated area 1.40 1.10
KI Outflow coefficients of the free water storage to interflow relationships 0.25 0.20
KG Outflow coefficients of the free water storage to groundwater relationships 0.45 0.43
CI Recession constants of the lower interflow storage 0.99 0.63
CG Recession constants of the groundwater storage 0.80 0.95
CS Recession constant in the lag and route method for routing through the channel system within each sub-basin 0.24 0.24
KE Muskingum time constant for each sub-reach 0.12 0.10
XE Muskingum weighting factor for each sub-reach 0.01 0.01
KM(mm) Storage capacity of the karst reservoir 39.25 ——
HK(mm) Threshold value 30.00 ——
IK Percentage of karst areas in the catchment 0.41 ——
KKB Outflow coefficients of the karst water storage to direct flow relationships 0.45 ——
KKG Outflow coefficients of the karst water storage to groundwater relationships 0.15 ——
CK Recession constants of the direct flow storage 0.96 ——

potential rate and the ratio of WL to LM; ED occurs at the modified rate are the recession constants of the lower interflow storage and ground-
which is computed by multiplying the potential rate and the coefficient water storage respectively, and CS represents recession constant in the
of the deep layer C. lag and route method for routing through the channel system within
each sub-basin.
(2) Runoff generation module The outflow from each sub-basin is finally routed by the Muskingum
successive-reaches model to produce the flow at the outlet of the entire
The XAJ model divides the basin into permeable and impervious catchment. The river reach is normally subdivided into sub-reaches in
areas by IM which is the percentage of impervious and saturated areas this successive routing scheme. There are two parameters to be de-
in the catchment. In the impervious areas, runoff equals the rainfall termined for each sub-reach, including the Muskingum time constant
excess without further loss. In the permeable areas, the main feature of (KE) and weighting factor (XE).
the XAJ model is the concept of runoff formation on the depletion of
storage, which means that runoff is not produced until the soil moisture 3. K-XAJ model
content of the aeration zone reaches the field capacity, and thereafter
runoff equals the rainfall excess without further loss. The runoff-pro- In the following, two different methods were used to simulate runoff
ducing area is critical for calculating runoff. Runoff distribution is generation in karst areas and non-karst areas in a karst-dominated river
usually nonuniform across a region because the soil moisture deficit is basin.
heterogeneous. In order to accommodate the nonuniformity of the soil
moisture deficit or the tension water capacity distribution, the XAJ
3.1. Runoff generation in karst area
model adopted the storage capacity curve (Zhao et al., 1980) to cal-
culate the total runoff, while the B represents the exponential of the
In order to simulate discharge of river basins which show significant
distribution to tension water capacity.
karst characteristics, a karst reservoir (V1) and a karst regulating re-
servoir (V2) were proposed by Cheng et al. (1991) to simulate runoff
(3) Runoff partition module
generation in karst areas. The structure and parameters of the reservoirs
are shown in Fig. 2 and Table 1. Each reservoir represents physical
In the XAJ model, the total runoff is subdivided into three compo-
nents by a free reservoir, including surface runoff (RS), interflow (RI)
and groundwater runoff (RG). SM represents the real mean free water
capacity of the surface soil layer, and KI, KG are outflow coefficients of
the free water storage to interflow relationships and groundwater re-
lationships respectively. Fresh water capacity is usually nonuniform
across a region, because the runoff-producing area of saturated slope
flow is constantly changing. And the XAJ model adopted the free water
capacity distribution curve to accommodate the nonuniformity of the
fresh water capacity while EX represents the exponent of the free water
capacity curve influencing the development of the saturated area.

(4) Runoff routing module

The surface runoff is routed directly to the watershed outlet of each

sub-basin in a source-to-sink form (Moore et al., 2006) by means of the
lag-and-route method, while the interflow and groundwater runoff are
routed through linear reservoirs (Lu et al., 2014). In the module, CI, CG Fig. 2. Structure of the karst reservoirs.

526
Q. Zhou, et al. Journal of Hydrology 573 (2019) 524–533

runoff partition module of the traditional XAJ model.

3.3. Runoff concentration in karst area

A combination model, namely Karst-Xinanjiang (K-XAJ) model, as

shown in Fig. 1, was proposed in this study. In order to consider runoff
generation in both karst area and non-karst area in a karst-dominated
basin, the K-XAJ model was developed, where parameter IK was used to
denote the weight of the karst area in the whole basin. Based on the
areal weight IK, the K-XAJ model can capture the runoff generation
characteristics of both karst and non-karst areas.
In the K-XAJ model, RS calculated by the XAJ model was combined
with RKS calculated by the karst reservoir and Rb in the impermeable
area to form the total surface runoff (QS). RG calculated by the XAJ
model and RKG calculated by the karst reservoir model formed the total
Fig. 3. Physical process of a karst system.
groundwater QG. Since the physical processes are very different for
generating RI and RKB in karst and non-karst areas, RI and RKB were
regulated by the linear reservoir in the XAJ model and karst reservoir
storage over the catchment. Reservoir V1 represents the karst system, (V2), respectively. RI and RKB formed interflow (QI) and direct karst
aiming at dividing the runoff into three kinds, namely rapid karst runoff flow (QK), respectively. The equations for calculating the surface flow
(RKS), direct karst runoff (RKB), and karst groundwater runoff (RKG). (Qs), interflow (QI), direct karst flow (QK), and groundwater (QG) were
Fig. 3 shows the physical process of each runoff component, where RKS defined as:
represents the surface runoff of carbonate rock; RKB represents the
transverse runoff formed in the fissures and fractures of epikarst (the QS = (RS + Rb + RKS ) × U (5)
uppermost layer of the carbonate rock); and RKG refers to runoff in a QI = QI 0 × CI + RG × (1 − CI ) × U (6)
karst underground river or conduit. Reservoir V2 represents epikarst,
which is used to regulate the direct karst flow. QK = QK 0 × CK + RKB × (1 − CK ) × U (7)
Reservoir V1 has a balance equation:
QG = QG0 × CG + (RG + RKG ) × (1 − CG ) × U (8)
SK = P − E + SK 0 − RKS − RKB − RKG (1)
where QI0, QK0, and QG0 are the flow QI, QK, and QG in the previous
where SK0 is the initial water storage of V1, and SK is the water storage period; CI, CK, and CG are the recession constants of interflow, direct
in reservoir 1. flow, and groundwater, respectively; and U is the unit conversion
The difference between rainfall (P) and evaporation (E) was used as coefficient.
the input of Reservoir V1. If the current water storage of Reservoir V1
was greater than the water storage capacity of Reservoir V1, the excess 4. Parameter calibration
parts would generate rapid runoff RKS, and the remaining part would
generate RKB and RKG according to the linear outflow coefficients KKB 4.1. Model calibration
and KKG, respectively. However, due to the development of larger
fractures and swallowing holes on the surface and in the epikarst, The parameter calibration of hydrological models is normally done
larger-scale karst systems usually have high infiltration capacities, considering only a single objective function, such as in genetic algo-
which make the flow in reservoir 1 very small, and more water occurs rithm (GA) (Holland et al., 1975), particle swarm optimization algo-
in the form of groundwater. Thus, in order to simulate the feature of rithm (PSO) (Kennedy, 2010), shuffled complex evolution algorithm
karst area, a threshold value (HK) is considered, which reflects the de- (SCE-UA) (Duan et al., 1992), and so on. However, practical applica-
velopment of vertical fractures in the shallow aquifer. While SK is tions indicate that a single-objective calibration method is often sub-
greater than HK, the direct karst runoff RKB is generated. The equations jective and inadequate to properly account for all the characteristics of
for calculating RKS, RKB and RKG are given as follows: the hydrological system.
The flood peak, volume and hydrograph shape are three basic ele-
S − KM if SK ⩾ KM
RKS = ⎧ K ments for describing a flood event (Zhang et al., 2018). These elements
⎩0
⎨ if SK < KM (2) are an important criterion for calibrating model parameters (Chen
et al., 2015, 2016; Huang et al., 2018a,b). In this study, the flood vo-
⎧ KKB × KM if SK ⩾ KM
lume error, Nash–Sutcliffe efficiency, and relative flood peak error were
RKB = KKB × (SK − HK ) if HK ⩽ KM ⩽ SK
⎨ selected as objective functions, which are given as follows. The objec-
⎩0 if SK < HK (3) tive functions, Eqs. (9) and (10), were evaluated at a daily time step,
while Eq. (11) was evaluated for a number of discrete flood events.
RKG = KKG × SK if SK ⩾ 0 (4)

where KKM is the coefficient of the karst reservoir outflow to direct flow; (1) Objective function in terms of the flood volume
and KKG is coefficient of the karst reservoir outflow to groundwater. n n
Obj1 = ∑ (Qobs,i − Qsim,i⎞⎟ ⎛ ⎞
⎜∑ Qobs, i⎟
3.2. Runoff generation in non-Karst area i=1 ⎠ ⎝ i = 1 ⎠ (9)

Runoff generation was calculated using the traditional XAJ model in (2) Objective function in terms of Nash–Sutcliffe efficiency
non-karst areas. In the impermeable area, the precipitation minus n n
evaporation formed the surface runoff (Rb). In the permeable area, the ⎡ ⎤ ⎡ ¯ 2⎤
NSE = 1 − ⎢∑ (Qobs, i − Qsim, i )2⎥ ⎢∑ (Qobs, i − Qobs ) ⎥
generation of surface runoff (RS), interflow (RI), and groundwater ⎣ i=1 ⎦ ⎣ i=1 ⎦ (10a)
runoff (RG) was simulated separately according to the procedures given
in the evapotranspiration module, the runoff generation module, and Obj2 = 1 − NSE (10b)

527
Q. Zhou, et al. Journal of Hydrology 573 (2019) 524–533

Fig. 4. Study area and locations of the gauging stations.

(3) Objective function in terms of flood peak the hydrological model parameter optimization. Detailed information
N of the Pareto optimization algorithm has been given by Khu and
1 Q′obs, i − Q′sim, i ⎞
Obj3 =
N
∑⎛ ⎜
Q′obs, i
⎟
Madsen (2005).
i=1 ⎝ ⎠ (11)
where Qobs, i is the observed flow at time i; Qsim, i is the simulated flow at 4.2. Performance index
time i; n is the length of sequence data; Qobs is the average value of
observed data; Qobs′ , i is the observed data of flood peak at time i; Qsim
′ , i is In order to evaluate the performance of the hydrological model,
the simulated value of flood peak at time i; n is the length of the data three performance indices, including Nash–Sutcliffe coefficient of effi-
set; and N is the number of flood events. ciency (NSE) (as shown in Eqs. (10)), relative flood peak error (RPE) (as
Guo et al. (2013) proposed a multi-objective shuffled complex dif- shown in Eq. (11)) and root mean square error (RMSE), were used to
ferential evolution (MOSCDE) for parameter calibration considering evaluate the K-XAJ model.
multi objectives, in which MOSCDE replaced the simplex search used in The root mean square error was defined as
SCE-UA with the differential evolution (DE) algorithm. The MOSCDE n
1
was used to calibrate the parameter of the hydrological models in- RMSE =
n
∑ (Qsim,i − Qobs )2
cluding Tank and HYMOD models by Ye et al. (2014) and Lu et al. i=1 (13)
(2016), and the results proved that the multi-objective method can be
In addition, the Chinese flood simulation standard recommends the
used for parameter estimation.
use of the qualified rate (QR) to evaluate the flood simulation perfor-
The multi-objective function was defined as
mance. A predicted peak value is regarded as “qualified” when the
Min {Obji (θ), Obj2 (θ), ...,Objm (θ)} (12) relative absolute error (RAE) between the predicted and the observed
streamflow values is within a given threshold value ε (Chen et al., 2013,
where m is the number of objective functions, Obji (θ) is the value of
2018c). QR can be calculated by
objective function for event i, and θ is the model parameter. The
n
MOSCDE method was used in this study for parameter calibration.
∑ numi
It is well known that the difficult problem of multi-objective cali- i=1
QR = × 100%
bration is how to get optimal parameters in a series of non-inferior n (14a)
optimization solutions. Generally, a weighted sum approach is used to
solve this problem where different objective functions are assigned where
different weights to convert a multi-objective problem to single-objec- 1, if (RAEi ⩽ ε ) |Qobs, i − Qsim, i |
tive problem (Schoups et al., 2005). However, the subjective prior se- numi = ⎧ , RAEi =
⎨
⎩ 0, otherwise Qobs, i (14b)
lection of weights typically affects the calibrated parameter values.
Thus, in this paper, the Pareto optimization algorithm was adopted for where RAEi is the relative absolute error (RAE) of the ith datum; numi is

528
Q. Zhou, et al. Journal of Hydrology 573 (2019) 524–533

set to 1 when RAE is less than or equal to the predefined threshold value relative flood peak error (RPE) of K-XAJ model for the validation period
(ε ), which is regarded as qualified simulation. ε was set as 20% in ac- were calculated based on the predicted and observed continuous daily
cordance with the accuracy standard established by the Ministry of streamflow series. The results were 90.4% (NSE), 298 (RMSE), 98.5%
Water Resources of the People’s Republic of China (2008). (QR), and 18.3% (RPE), respectively. Second, the model performance
was evaluated in terms of a number of discrete flood events. The flood
events were selected on the basis of the threshold flow, which is set to
5. Case study
3000 m3/s. When the peak flow is above this threshold value, this flood
event was selected. 13 flood events were chosen in seven years, in
The Lijiang River basin (as shown in Fig. 4), located in the south-
which 8 events are in calibration period and 5 in validation period. We
western part of China, was selected as a case study. The Lijiang River
calculated the performance indices of each flood event in terms of the
belongs to the Pearl River basin and is a tributary of the Xijiang River.
flood peak and volumes, and then computed the average value of each
The basin is in a subtropical monsoon region with distinct seasons,
performance index in the calibration and validation period respectively.
humid climate, sufficient sunshine and abundant rain. The average
The performance results corresponding to the flood events are shown in
annual precipitation is about 1700 mm. However, the annual distribu-
Table 2 as well, in which the NSE, QR, RPE of K-XAJ model for the flood
tion of rainfall is uneven, and the rainy season is from April to Sep-
events in the validation period are 89.4%, 100%, and 19%, respec-
tember, which accounts for more than 70% of the yearly precipitation.
tively. These results demonstrate that the K-XAJ model simulated runoff
In the Lijiang River basin, the exposed rock formations on the sur-
well and the model thus reflected the features of the karst area.
face of the area are mainly carbonate rocks, and the climate of the basin
is humid and hot climate. This basin has a large area of carbonate-based
tropical karst landforms. 6.2. Comparisons of the K-XAJ model with the traditional XAJ model
Pingle gauging station is the control station of the Lijiang River
basin, and the catchment area is 12,617 km2. The mainstream of the 6.2.1. Comparisons of the performances of the K-XAJ and XAJ models
Lijiang River is the Darong River, and there are four gauging stations Both the XAJ and K-XAJ models were used to simulate rainfall-
along the river, namely Darong, Guilin, Yangshuo, and Pingle. The main runoff in the Lijiang River basin. The performance indices were calcu-
tributaries are the Gongcheng River and the Lipu River, and there are lated, as shown in Table 2. It can be seen that the proposed K-XAJ
two gauging stations along the two tributaries, namely Gongcheng and model was better than the traditional XAJ model for both calibration
Lipu gauging stations, respectively. The precipitation and evaporation and validation periods. The predicted results of the K-XAJ and XAJ
data from the Guilin, Yangshuo, Gongcheng, and Pingle stations, were models are shown in Fig. 6 which demonstrates that the flood peak
used as input of the model, and the data from the Pingle gauging station predicted by the XAJ model is larger than the observed one. Compared
was used as the output of the hydrological model. The data length was with the traditional XAJ model, the proposed K-XAJ model gave better
from 2001 to 2007, which was divided into two parts. The data from simulation results.
the year 2001 to 2005 was used for model calibration, and the data
from the year 2006 and 2007 was used for model validation.
6.2.2. Comparisons of parameters of the K-XAJ and XAJ models
The traditional XAJ model was used for flood simulation in the
6. Results and discussion Lijiang River basin. The MOSCDE algorithm was adopted to calibrate
the parameters of the proposed K-XAJ model. In the XAJ model, there
6.1. Flood simulation results of the K-XAJ model are sixteen parameters, which are divided into sensitive and insensitive
parameters. The sensitive parameters were K, SM, KG, KI, CG, CI, and CS;
The proposed XAJ model was used for flood simulation in Lijiang and the insensitive parameters were UM, LM, DM, B, IM, C, EX, KE, and XE.
River basin. The MOSCDE algorithm was adopted to calibrate the In the K-XAJ model, there are six more parameters, of which IK, KKG,
parameters of the proposed K-XAJ model and the calibrated parameters KKB, and CK are sensitive parameters, and HK and KM are less sensitive
are given in Table 1. The observed and runoff simulated by K-XAJ parameters. Sensitive parameters have a significant influence on the
model are shown in Fig. 5. First, the model performance was evaluated simulation results. Table 1 presents the calibrated parameter values of
at a daily time step. The performance indices, including NSE, RMSE, QR the K-XAJ and XAJ models, respectively. The sensitive parameters CI
and RPE according to Eqs. (10), (11), (13), (14) were applied to eval- and CG showed obvious fluctuations. Compared with the traditional
uate the simulation results. The performance indices were calculated, as XAJ model, CI increased and CG decreased, which demonstrated that
shown in Table 2. The Nash–Sutcliffe efficiency (NSE), root-mean- the karst dominated area produced less interflow and much more
square error (RMSE), qualified rate of the flood volume (QR), and groundwater runoff. These results are consistent with the features of

0
12000

10000 50
Rainfull
Observe
Runoff(mm /s)

8000 100
Rainfall(mm)
3

K-XAJ
6000
150
4000
200
2000

0 250
2001/1/1 2002/1/1 2003/1/1 2004/1/1 2005/1/1 2006/1/1 2007/1/1 2008/1/1
Days
Fig. 5. Predicted results of the K-XAJ model (2001–2007).

529
Q. Zhou, et al. Journal of Hydrology 573 (2019) 524–533

Table 2
Comparison of performances of K-XAJ and the traditional XAJ models.
Models Daily streamflow series Flood events

NSE (%) RMSE RPE (%) QR (%) NSE (%) RPE (%) QR (%)

K-XAJ Calibration 85.7 296 12 93.7 90.7 11.2 100

Validation 90.4 298 18.3 98.5 89.4 19 100

XAJ Calibration 84.8 303 13.8 91.7 89.2 12.2 87.5

Validation 89.4 314 22.7 97.9 86.5 23 80

convergence in the karst region. coefficients of the free water storage to groundwater. With the increase
of karst area, the total underground runoff should increases. According
6.2.3. Comparisons of performances and parameter of the K-XAJ and XAJ to the calculated results, those two parameters also increase. CG is the
models due to different proportion of karst areas coefficient of groundwater regression. The larger the karst area is, the
The data from Guilin and Gongcheng gauging stations which is also faster the groundwater regress, and the smaller CG is. CI and CS are the
the control stations for two sub-basins, were used as case studies. The regression coefficient of interflow and surface runoff, respectively.
proportions of the karst area in these two sub-catchments are about When the karst area increases, the interflow should decrease and the
10% and 35%, respectively. The daily streamflow data from the year surface runoff should increase in the basin. According to Eqs. (6) and
2001 to 2007 was used for model calibration and validation. The per- (8) and results in Table 4, CI increases and CS decreases. Considering the
formance indices for K-XAJ and traditional XAJ models were calculated parameters evaluation in K-XAJ model due to the change of the pro-
and given in Table 3. portion of the karst areas, it is concluded that the proposed K-XAJ
It can be seen from Table 3 that for the sub-basin controlled by model can reflect the hydrological characteristics of the karst areas. The
Gongcheng gauging station, the performance of the proposed K-XAJ parameter evolution characteristics of K-XAJ model are consistent with
model is better than that of the traditional XAJ model for both the the features of runoff generation in the karst region.
calibration and validation periods. In the other sub-basin controlled by
Guilin gauging station, the karst area accounts only approximately 10% 6.2.4. Comparisons of simulated surface flow, interflow and groundwater
of the total areas. Compared with the traditional XAJ model, the per- by the K-XAJ and XAJ models
formance measured by NSE indicates that the K-XAJ model is slightly Several typical flood hydrographs for the year 2003 were selected
better than the traditional XAJ model. Generally, with the higher pro- and simulated by the proposed K-XAJ and XAJ models. Surface runoff,
portion of karst area, the K-XAJ model shows a better simulation effi- interflow, and groundwater were calculated, as shown in Figs. 7–9,
cacy. To the contrary, corresponding to the lower proportion of the respectively. Fig. 7 shows that the surface flood peak calculated by the
karst area, the performance of the two models do not show significant K-XAJ model was significantly higher than that by the XAJ model. This
difference. was due to the thin soil and little evaporation in the karst area. When a
Table 4 presents the sensitive parameters of the K-XAJ and XAJ heavy rainfall occurs, the rainfall minus the evaporation directly flows
models in the three basins. The proportions of karst area corresponding into the karst reservoir. If the current water storage (SK) of the karst
to the three basins are approximately 10%, 35% and 40% respectively Reservoir 1 is greater than the karst storage capacity (KM), the excess
based on the observed geographical data. K is the ratio of potential water will generate the rapid karst flow, which finally forms surface
evapotranspiration to pan evaporation, which is related to the air runoff and causes the flood peak to increase rapidly.
temperature and observed evaporation data, so there is no obvious Fig. 8 shows that the interflow calculated by the K-XAJ model is
difference. IK reflects the ratio of the karst area, which is equal to 7%, significantly less than that by the XAJ model. In the karst area, due to
37% and 41% in the K-XAJ model for the three basins respectively. The the exposure of multiple bedrocks and thin soil, the water storage ca-
estimated IK (7%, 37% and 41%) is more or less the same with the pacity of soil layer decreases and the ability of soil layer to redistribute
observed data (10%, 35% and 40%), which proves the rationality of the rainfall is weak, resulting in a significant decrease in interflow.
proposed method. KI and KKB are outflow coefficients of the free water In Fig. 9, the magnitude of groundwater simulated by the K-XAJ
storage to interflow and the karst water storage to direct flow. With the model was greater than by the traditional XAJ model. The groundwater
increase of the karst areas, the interflow and direct karst runoff should concentration speed was faster, and the volume was larger in the initial
decrease. According to the results present in Table 4, KI and KKB de- period of rainfall; and similarly in the later period the groundwater
crease with the increase of the karst areas. KG and KKG are outflow receded faster. Due to the development of vertical fractures and

12000 0

10000 Rainfall 50
Observe 100
8000 XAJ
Runoff (m3/s)

Rainfall(mm)

K-XAJ 150
6000
200
4000
250
2000 300
0 350
0 10 20 30 40 50 60 70
Days
Fig. 6. Simulation of flood hydrograph in 2002 by the K-XAJ and XAJ models.

530
Q. Zhou, et al. Journal of Hydrology 573 (2019) 524–533

Table 3
Comparisons of performances of K-XAJ and the traditional XAJ models due to different proportion of karst areas.
Gongcheng Guilin

NSE (%) RMSE RPE (%) QR (%) NSE (%) RMSE RPE (%) QR (%)

K-XAJ Calibration 85.9 66 14 97.3 84.9 110 16.8 93.9

Validation 87.7 62 13.9 94.8 86.4 71 21.5 84.9

XAJ Calibration 84.9 68 16.3 95.6 84.8 111 16.2 95.2

Validation 85.4 65 17.1 92.6 84.7 82 20.8 82.7

Table 4 7. Conclusion
Calibrated sensitive parameters of the K-XAJ and XAJ models in three basins.
Parameter Guilin Gongcheng Pingle A conceptual coupled hydrological model, namely K-XAJ model, has
been proposed for modelling rainfall-runoff generation in the karstic
K-XAJ XAJ K-XAJ XAJ K-XAJ XAJ catchments. The conceptual model appears to be a good compromise
between an empirical model and a distributed model, which has rela-
K 0.40 0.40 0.40 0.73 0.78 0.84
KI 0.35 0.36 0.28 0.30 0.25 0.20
tively fewer parameters than the distributed model, and can generally
KG 0.37 0.37 0.44 0.43 0.45 0.43 reflect physical relationships of the karst aquifer. The Lijiang River
CI 0.80 0.82 0.98 0.80 0.99 0.63 basin was selected as a case study. Both the proposed K-XAJ and tra-
CG 0.98 0.98 0.82 0.90 0.80 0.95 ditional XAJ models were used for flood simulation in this region and
CS 0.50 0.48 0.48 0.34 0.24 0.24
the simulated results were compared and analyzed. The main conclu-
IK 0.07 —— 0.37 —— 0.41 ——
KKB 0.60 —— 0.59 —— 0.45 —— sions are given as follows.
KKG 0.09 —— 0.10 —— 0.15 —— The main features of the proposed model are: (a) The model is
composed of the traditional XAJ model and karst reservoirs to simulate
rainfall-runoff for a karst dominated area; (b) the proposed model can
swallowing holes in the epikarst, the surface flow was often connected simulate runoff of both the karst area and the non-karst area in the karst
to the underground. Thus, precipitation at the beginning of rainfall can dominated basin; (c) the proposed karst reservoir can separate runoff
rapidly infiltrate into the underground, and recharge groundwater. into three components, which match the physical process of runoff
Since the developed underground multi-channels and caves provided generation in karst areas; and (d) the parameters in the K-XAJ model
sufficient space for the storage of groundwater, the development of can reflect the characteristics of karst aquifers.
underground rivers and conduits increased the speed of groundwater The performance indices and simulated flood hydrographs showed
recession. that the proposed K-XAJ model simulated the rainfall-runoff processes
well. The estimated parameters have physical meaning. Compared with
the traditional XAJ model, the proposed K-XAJ model better simulated
6.2.5. Comparison of flood recession period simulated by the K-XAJ and runoff.
XAJ models Comparison of surface flow, interflow, and groundwater, simulated
Fig. 10 shows the flood hydrograph predicted in 2003. As shown in by the proposed and traditional XAJ models demonstrated that the peak
the dash box of Fig. 10, the flood recession of the K-XAJ model was surface flow predicted by the K-XAJ model was larger than that by the
faster than that of the XAJ model, which was closer to the observed XAJ model. On the contrary, the interflow routed by the K-XAJ model
flood recession. The reasons are given as follows. First, the increase in was significantly less than that by the XAJ model. The magnitude of
surface runoff and the decrease in the interflow in karst area reduced groundwater predicted by the K-XAJ model was greater.
the time of flood recession. Second, the development of underground Finally, the simulated flood hydrograph in the flood recession
rivers and conduits in karst area accelerated the velocity of flood re- period was analyzed. Results indicated that the flood recession of the K-
cession. XAJ model was faster than that of the XAJ model, which was closer to
the observed flood recession.

4000
K-XAJ
3500
XAJ
3000

2500
Runoff(mm /s)
3

2000

1500

1000

500

0
0 20 40 60 80 100
Days
Fig. 7. Surface runoff simulation by the proposed K-XAJ and XAJ models.

531
Q. Zhou, et al. Journal of Hydrology 573 (2019) 524–533

800

700 K-XAJ
XAJ

Subsurface runoff (m 3 /s)

600

500

400

300

200

100

0
0 20 40 60 80 100
Days
Fig. 8. Interflow simulation by the proposed K-XAJ and XAJ models.

1000
K-XAJ
XAJ
Groundwater runoff (m /s)

800
3

600

400

200

0
0 20 40 60 80 100
Days
Fig. 9. Groundwater simulation by the proposed K-XAJ and XAJ models.

4000

3500 Karst-XAJ
Observe
3000 XAJ

2500
Runoff (m3/s)

2000

1500

1000

500

0
0 10 20 30 40 50 60 70 80
Days
Fig. 10. Simulation results of floods recession (2003).

Declaration of interests Acknowledgments

The authors declare that they have no known competing financial This study is supported by National Key R&D Program of China
interests or personal relationships that could have appeared to influ- (2017YFC0405900), National Natural Science Foundation of China
ence the work reported in this paper. (51679094; 51879109; 91547208).

532
Q. Zhou, et al. Journal of Hydrology 573 (2019) 524–533

References conceptual rainfall-runoff model. J. Hydrol. 373, 302–315.

Kennedy, J., 2010. Particle swarm optimization. In: Encyclopedia of Machine Learning.
Springer, US, pp. 760–766.
Bailly-Comte, V., Jourde, H., Pistre, S., 2009. Conceptualization and classification of Khu, S.T., Madsen, H., 2005. Multiobjective calibration with Pareto preference ordering:
groundwater-surface water hydrodynamic interactions in karst watersheds: case of an application to rainfall-runoff model calibration. Water Resour. Res. 41, 477–490.
the karst watershed of the Coulazou river (southern France). J. Hydrol. 376, Labat, D., Ababou, R., Mangin, A., 1999. Linear and nonlinear input/output models for
456–462. karstic spring flow and flood prediction at different time scales. Stoch. Env. Res. Risk
Bonacci, O., 2001. Analysis of the maximum discharge of karst springs. Hydrogeol. J. 9, A 13, 337–364.
328–338. Lambrakis, N., Andreou, A.S., Polydoropoulos, P., 2000. Nonlinear analysis and fore-
Charlier, J.B., Bertrand, C., Mudry, J., 2012. Conceptual hydrogeological model of flow casting of a brackish Karstic spring. Water Resour. Res. 36, 875–884.
and transport of dissolved organic carbon in a small Jura karst system. J. Hydrol. Li, Y.B., Wang, S.J., Wei, Z.F., 2006. The spatial distribution of soil loss tolerance in
460–461, 52–64. carbonate area in Guizhou province. Earth Environ. 36–40 (in Chinese).
Chartzoulakis, K.S., Paranychianakis, N.V., Angelakis, A.N., 2001. Water resources mana- Li, R., Wang, S., Zhou, D., 2003. The correlation between rock desertification and li-
gement in the island of Crete, Greece, with emphasis on the agricultural use. Water thology in karst area of Guizhou. Acta Geograph. Sin. 314–320 (in Chinese).
Pol. 3, 193–205. Lu, M.J., Li, X., 2014. Time scale dependent sensitivities of the XinAnJiang model
Chen, L., Singh, V.P., Lu, W., Zhang, J., Zhou, J., Guo, S., 2016. Streamflow forecast parameters. Hydrol. Res. Lett. 8, 51–56.
uncertainty evolution and its effect on real-time reservoir operation. J. Hydrol. 540, Lu, W.W., Zhou, J.Z., 2016. Study on the Short-term Hydrological Forecasting of Zhexi
712–726. Watershed. Doctoral dissertation.
Chen, L., Singh, V.P., Guo, S., Zhou, J., Zhang, J., 2015. An objective method for parti- Malagò, A., Efstathiou, D., Bouraoui, F., 2016. Regional scale hydrologic modeling of a
tioning the entire flood season into multiple sub-seasons. J. Hydrol. 528, 621–630. karst-dominant geomorphology: the case study of the island of crete. J. Hydrol. 540,
Chen, L., Singh, V.P., Guo, S.L., 2013. Measure of correlation between river flows using 64–81.
the copula-entropy method. J. Hydrol. Eng. 18 (12), 1591–1606. Ministry of Water Resources of the People’s Republic of China (MWR), 2008. Standard for
Chen, L., Singh, V.P., Guo, S., Zhou, J., Ye, L., 2014a. Copula entropy coupled with ar- Hydrological Information and Hydrological Forecasting (GB/T22482-2008).
tificial neural network for rainfall-runoff simulation. Stochastic Environ. Res. Risk Standards Press of China, Beijing (in Chinese).
Assess. 8 (7), 1755–1767. Moore, R.J., Cole, S.J., Bell, V. A., 2006. Issues in flood forecasting: ungauged basins,
Chen, L., Sun, N., Zhou, C., 2018c. Flood forecasting based on an improved extreme extreme floods and uncertainty. pp. 103–122.
learning machine model combined with the backtracking search optimization algo- Moraetis, D., Efstathiou, D., Stamati, F., 2010. High frequency monitoring for the iden-
rithm. J. Water. 10, 2–17. tification of hydrological and biochemical processes in a Mediterranean river basin. J.
Chen, L., Ye, L., Singh, V.P., Zhou, J., Guo, S., 2014b. Determination of input for artificial Hydrol. 389, 127–136.
neural networks for flood forecasting using the copula entropy method. J. Hydrol. Nikolaidis, N.P., Bouraoui, F., Bidoglio, G., 2013. Hydrologic and geochemical modeling
Eng. 11 (11), 19. of a karstic Mediterranean watershed. J. Hydrol. 477, 129–138.
Cheng, G.W., 1991. The Xinanjiang model in Karst area. Hydropower Energy Sci. 9, Paleologos, E., Skitz, i.I., Katsifarakis, K., 2013. Neural network simulation of spring flow
139–144 (in Chinese). in karst environments. Stoch. Envrion. Res. Risk A 27, 1829–1837.
Daly, D., Dassargues, A., Drew, D., Dunne, S., Goldscheider, N., Neale, S., Popescu, C., Schoups, G., Addams, C.L., Gorelick, S.M., 2005. Multi-objective calibration of a surface
Zwhalen, F., 2002. Main concepts of the ‘‘European Approach” for (karst)ground- water-groundwater flow model in an irrigated agricultural region: Yaqui Valley,
water vulnerability assessment and mapping. Hydrogeol. J. 10 (2), 340–345. Sonora Mexico. Hydroll. Earth Syst. Sci. 9, 549–568.
Dreiss, S.J., 1989a. Linear kernels for karst aquifers. Water Resour. Res. 18, 865–876. Shi, P., Zhou, M., Qu, S., 2013. Testing a conceptual lumped model in karst area,
Dreiss, S.J., 1989b. Regional scale transport in a Karst Aquifer: 2. Linear systems and time southwest china. J. Appl. Mathemat. 2, 1–10.
moment analysis. Water Resour. Res. 25, 126–134. Tritz, S., Guinot, V., Jourde, H., 2011. Modelling the behaviour of a karst system catch-
Duan, Q.Y., Sorooshian, S., Gupta, V., 1992. Effective and efficient global optimization for ment using non-linear hysteretic conceptual model. J. Hydrol. 397, 250–262.
conceptual rainfall-runoff models. Water Resour. Res. 28 (1015–1031), 1992. Wang, X., Liang, Z., Wang, J., 2014. Simulation of runoff in Karst-influenced Lianjiang
Fleury, P., Plagnes, V., Bakalowicz, M., 2007. Modelling of the functioning of karst Watershed using the SWAT Model. Sci. J. Earth Sci. 4, 85–92.
aquifers with a reservoir model: application to Fontaine de Vaucluse (South of Ye, L., Zhou, J., Zeng, X., 2014. Multi-objective optimization for construction of predic-
France). J. Hydrol. 345, 38–49. tion interval of hydrological models based on ensemble simulations. J. Hydrol. 519,
Ford, D.C., Williams, P.W., 2007. Karst Hydrogeology and Geomorphology. Wiley, 925–933.
Chichester. Yuan, D., Chai, G., 1988. The Science of Karst Environment. Chongqing Press, Chongqing,
Guo, J., Zhou, J., Zou, Q., 2013. A novel multi-objective shuffled complex differential China.
evolution algorithm with application to hydrological model parameter optimization. Zhang, Z., Chen, X., Ghadouani, A., 2010. Modelling hydrological processes influenced by
Water Resour. Manage. 27, 2923–2946. soil, rock and vegetation in a small karst basin of southwest China. Hydrol. Process.
Holland, J.H., 1975. Adaptation in Natural and Artificial Systems. University Michigan 25, 2456–2470.
Press, Michigan. Zhang, C., Chen, X., Li, Y., Ding, W., Fu, G., 2018. Water-energy-food nexus: concepts,
Huang, K.D., Ye, L., Chen, L., Wang, Q.S., Dai, L., Zhou, J., Singh, V.P., Huang, M., Zhang, questions and methodologies. J. Clean. Prod. 195, 625–639.
J., 2018a. Risk analysis of flood control reservoir operation considering multiple Zhao, R.J., 1992. Xinanjiang model applied in China. J. Hydrol. 135, 371–381.
uncertainties. J. Hydrol. 565, 672–684. Zhao, R.J., Zuang, Y., Fang, L., 1980. The xinanjiang model. IAHS AISH Publ. 129,
Huang, K.D., Chen, L., Zhang, J.H., et al., 2018b. Flood hydrograph coincidence analysis 351–356.
for mainstream and its tributaries. J. Hydrol. 565, 341–353.
Jukic, D., Denic-Jukic, V., 2009. Groundwater balance estimation in karst by using a

533