Agricultural Water Management 107 (2012) 74–85

Contents lists available at SciVerse ScienceDirect

Agricultural Water Management

journal homepage: www.elsevier.com/locate/agwat

Optimization of the irrigation water resources for agricultural sustainability in

Tarim River Basin, China
Y. Huang a,1 , Y.P. Li b,∗ , X. Chen c,2 , Y.G. Ma d,3
a
State Key Laboratory of Desert and Oasis Ecology, Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences, Urumqi 830011, China
b
MOE Key Laboratory of Regional Energy Systems Optimization, S-C Resources and Environmental Research Academy, North China Electric Power University, Beijing 102206, China
c
Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences, Urumqi 830011, China
d
Xinjiang Remote Sensing Center, Urumqi 830011, China

a r t i c l e i n f o a b s t r a c t

Article history: An integrated optimization method is developed for supporting agriculture water management and
Received 24 July 2011 planning in Tarim River Basin, Northwest China. The developed method couples two-stage stochastic
Accepted 16 January 2012 programming (TSP) with inexact quadratic program (IQP). The hydrological model is provided for fore-
Available online 8 February 2012
casting the available irrigation water. The simulation system is then embedded into an optimization
framework, where the objective is to maximize the system benefit for water resources management. The
Keywords:
developed method can not only deal with nonlinearities in the cost/benefit objective and uncertainties
Agriculture sustainability
expressed as probabilities and intervals, but also support the analysis of policy scenarios that are asso-
Optimization
Tarim River Basin
ciated with economic penalties when the promised water-allocation targets are violated. A case study
Two-stage is conducted for Kaidu-kongque watershed in Tarim River Basin. The results obtained can help gener-
Uncertainty ate desired policies for water resources management with maximized economic benefit and minimized
Water management system-failure risk.
© 2012 Elsevier B.V. All rights reserved.

1. Introduction To address the above concerns, the government and farmers should
strengthen water infrastructure construction for improving water
Water resources are lifeline of oasis agriculture development use efficiency, and thus increase the quantity and frequency of irri-
in arid region. The availability of water is an important factor for gation during agricultural cultivation (Zhou et al., 2010; Ping et al.,
agricultural production in arid areas (Sharma and Minhas, 2005). 2010).
However, excessive utilization of water resources strongly influ- Stochastic programming (SP) can deal with situations when
ences local agricultural development and plays a vital role in some random parameters or variables appear in the modeling for-
accelerating environmental degradation (Li et al., 2009). For exam- mulation of a program. Such random parameters are expressed
ple, in arid land of northwest China, agriculture water consumption as probability density functions (PDFs). This implies that SP tech-
accounts for approximately 90% of the total water uses (Li et al., nique can be employed when the quality of uncertain information
2010); but the average available water is less than 1635 × 108 m3 is comprehensive. The results can be interpreted under different
per year, only 5.8% of the China average level. On the contrary, level of probabilities (or risks). Preciously, a number of stochastic
increased population shifts and shrinking water supplies have programming methods were developed to support decision mak-
exacerbated competition among different users. This competitive- ing of water resources management and planning (Ferrero et al.,
ness can strengthen the agricultural water shortage, and serious 1998; Mylopoulos et al., 1999; Huang and Loucks, 2000; Seifi and
problems (e.g., agricultural sustainability concerned) can thus arise Hipel, 2001; Maqsood et al., 2005a,b; Li et al., 2008; Huang et al.,
from poorly planned water-management systems when merely 2010). Among them, two-stage stochastic programming (TSP) is
limited water resources are available for multiple competing users. proved effective for problems where an analysis of policy sce-
narios is desired and the right-hand-side coefficients are random
with known probability distributions (Li et al., 2006). For example,
Ferrero et al. (1998) examined hydrothermal scheduling of multi-
∗ Corresponding author. Tel.: +86 10 5197 1255; fax: +86 10 5197 1284. reservoir systems using a two-stage stochastic algorithm. Huang
E-mail addresses: [email protected] (Y. Huang), [email protected] (Y.P. Li), and Loucks (2000) proposed an inexact two-stage stochastic pro-
[email protected] (X. Chen), [email protected] (Y.G. Ma).
1
gramming model to address uncertainties expressed as probability
Tel.: +86 991 723131.
2
Tel.: +86 991 7885303; fax: +86 991 7885320. distribution functions and intervals. Luo et al. (2007) presented
3
Tel.: +86 991 3837564. a simulation-based inexact two-stage stochastic programming for

0378-3774/$ – see front matter © 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.agwat.2012.01.012
 Y. Huang et al. / Agricultural Water Management 107 (2012) 74–85 75

agricultural nonpoint source pollution control through land retire- serious problem due to increasing demand (Li et al., 2010). The
ment under uncertain conditions, where random parameters were results obtained will be used for helping local decision makers to
provided by the statistical analysis of the simulation outcomes of establish effective water exploitation and allocation policies, and
a distributed water quality approach. Li et al. (2010) developed an thus facilitate the local agriculture sustainability.
inexact two-stage water management model for planning agricul-
tural irrigation in the Zhangweinan River Basin, China; targeted 2. Study system
incomes, recourse costs, and net system benefits under different
policy scenarios are analyzed. Generally speaking, the above TSP Tarim River is located in northwest of China, which is the
methods were based on an assumption that the objective func- longest inland river all over the country. This basin is a typical
tion was linear, resulting in difficulties in dealing with such issues water-shortage area, with characteristics of low rainfall, high tem-
wherein marginal utility affect the revenue and/or cost coefficients perature, and high evaporation. The Tarim River is formed by the
in a TSP problem and thus make the relevant objective function unions of Aksu, Hotan, Yarkant and Kaidu-kongque rivers, and flows
nonlinear. east along the northern edge of the desert. The river usually refers
Quadratic program (QP) can reflect nonlinearity in the to the mainstream from Xiaojiake to the Taitmar Lake with a length
cost/benefit objectives, and has global optimum under a number of 1300 km. The region is suffering the extremely ecological degra-
of system conditions (Hillier and Lieberman, 1986). For exam- dation since 1970s. Nearly one-third chainage in the downstream
ple, Rockafellar and Wets (1986) proposed a Lagrangian finite of Tarim River has been dry out due to the extensive agriculture
generation technique for solving linear-quadratic problems in two- explore and improper irrigation method.
stage stochastic programming; Shilman (1992) used a stochastic Kaidu-kongque watershed is located in the middle reaches of
quasigradient method for quadratic optimization under dependent the Tarim River Basin. It is one of the most important bases of
observations. The stochastic quadratic programming can effec- cotton and grain in Xinjiang and China, and has an area of approx-
tively reflect random uncertainty; however, in many real-world imately 31,400 km2 . The region is of typical continental climatic
problems, the quality of available information about the uncer- characteristics: arid with little precipitation and intensive evapora-
tainties is often not good enough for establishing probability tion; the annual precipitation is usually from 200 m to 500 m in the
distributions. Chen and Huang (2001) introduced interval param- mountain, and less than 50 mm in the basin; annual mean temper-
eters into QP framework to handle uncertainty without knowing ature 10–11 ◦ C, accumulated temperature is up to 4000–4350 ◦ C;
probability distributions; a derivative algorithm method was pro- pan evaporation is about 1157 mm/year. There are 11 soil groups in
posed for solving the interval quadratic programming (IQP) model the region of Kaidu-kongque watershed; the main soil types cover
with a low computational effort. Although the IQP approach could floating aeolian sandy soil, semi-fixed aeolian sandy soil, salinized
handle the interval uncertainties in the model’s left- and/or right- branchy tamarisk forest land, salinized diversifolious poplar for-
hand sides as well as in the objective function, it had difficulties in est land, diversifolious poplar forest land, salification meadow soil,
dealing with uncertainties expressed as random variables and was meadow soil, meadow marsh soil, solonchak, residual solonchak,
lack of linkage to economic consequences of violated policies. residual peat swamp soil and artificial oasis irrigation arable soil.
Therefore, as an extension of the previous works, the objective The vegetation types cover warm-temperate-zone sparse shrub,
of this study is to develop a two-stage interval quadratic program- subshrub desert area, desert, spare shrub, subshrub desert area.
ming (TIQP) method for planning agriculture water management. According to administrative division, the watershed includes seven
The developed TIQP can not only handle uncertainties expressed irrigation-zones such as Hejing County, Hoxud County, Luntai
as interval values and random variables, but also tackle nonlin- County, Yanqi County, Bohu County, Korla City and Yuli County (Xu
earity in the objective function. The TIQP will be demonstrated et al., 2007; Huang et al., 2010). Fig. 1 shows the outline of Kaidu-
in the Kaidu-kongque watershed in northwest China for agricul- kongque watershed with the irrigation zones. The Kaidu-kongque
tural water planning and management. This watershed is one of the River is used for providing resources for regional water supply, agri-
main food and cotton producing regions in south Xinjiang province, culture irrigation, industrial production and stock breeding, as well
China; moreover, water shortage has become an increasingly as supplement to the lower reach of the Tarim River.

Fig. 1. Study system of the Kaidu-kongque watershed.

76 Y. Huang et al. / Agricultural Water Management 107 (2012) 74–85

Kaidu-kongque watershed is one of the most important agri-

cultural regions in Xinjiang Province, located in the middle-lower
reaches of the Tarim River. The intensified confrontation between
environmental protection and economic development of the region
has been inevitable as a result of exploitation and utilization of
the limited water resources for the local economic activities on
one hand, and great concerns for the deteriorating regional envi-
ronment on another. Sustainable development of the regional
socio-economy, therefore, has been seriously constrained. The
ecosystems in this region are extremely vulnerable to disturbances
and habitat degradation due to severe lack of water resources (Xu
et al., 2008). Poor management practices may lead to severe envi-
ronmental degradation at the early stages of the land reclamation, Fig. 2. Cumulated probability distribution function (CDF) of available water
resources.
and the problems may be further exacerbated by deforestation and
irrigation-induced salinity. In order to recover the ecological sys-
tem of the lower reach of the watershed, the local planers have model is used to estimate the time series of annual stream flow from
conducted ecological water releases from the Kaidu-konque River 1960 to 2005 for the DSK station. Then, we can get the probability
every year since 2000. Moreover, due to the shortage of precip- distribution of the stream flow based on the simulation time series.
itation in the oasis, more than 95% of the arable land requires When choosing the probability distributions of the time series of
irrigation, and about 92% of the total water-resources consump- stream flow, several criteria should be obeyed including (i) simple
tion in the watershed is contributed by agriculture irrigation. The calculation and high accuracy, (ii) the best fitting between simu-
main crop types in the Kaidu-kongque watershed are cereal, cotton, lation and observer data, and the shape of simulated curve should
oil bearing crops, vegetable, and forge. However, in Kaidu-kongque fit in with the physical property of hydrological phenomena (i.e.
watershed, planting area of the cereal crops, economic crops and the curve should have limit non-negative values). In this study,
breeding crops are not balance. Low water use efficiency of agricul- several probability distributions (i.e. gamma distribution, normal
tural irrigation has forced the planers to develop advance methods distribution, and logarithmic  distribution), which are often used
for agriculture water resources allocation in this region (Mainuddin in hydrological researches in northwest China, are chosen as the
et al., 1997; Zhang et al., 2009). A sound water-resources man- candidate distributions (Lai, 2007; Tao et al., 2007). As a result, the
agement planning with limited water resources is important for Gamma distribution is the best fit for stream flow. Fig. 2 shows
not only agricultural production but also economic development the cumulative probability distribution function (CDF)  of random
and ecology recovering in the basin. Moreover, because of the seasonal flow. The interval values of seasonal flow q±
j all
under
temporal variations of the available water resources, the optimal
different probability levels can then be calculated with this fitted
schemes for effective irrigation of the crops (i.e. optimal agricul-
distribution. Table 1 presents the probability levels correspond to
tural water-allocation schemes) can also vary correspondingly (Li
different interval values of the stream flow.
et al., 2011; Tran et al., 2011; Reca et al., 2001). The complexities can
become further compounded by not only interactions among many
3. Two-stage interval quadratic programming
uncertain system components but also their economic implica-
tions, which may affect the relevant water allocation optimization
When uncertain parameters (e.g., water availability) are
analysis and thus the associated decision. The advanced optimiza-
expressed as probability distributions and decisions need to be
tion model can be used for allocation and managing agriculture
made periodically over time, the problem can be formulated as
water in more efficient way.
a TSP model. A general TSP model can be formulated as follows
The Kaidu-kongque River is responsible for providing resources
(Huang and Loucks, 2000):
for agricultural water supply in the watershed. The random charac-
ters can be conducted through statistical analyses with simulation max f = cx − E[Q (x, )] (1a)
results of annual stream flow of the river which is responsible for
supplying water to the watershed. Two meteorological stations (i.e. subject to:
BYBLK and DSK stations) are used to derive the inputs of air tem- Ax ≥ b (1b)
perature, pan evaporation, and daily rainfall of the study area. A
rainfall runoff model named as Nedbor-Afstromnings Model (NAM) x≥0 (1c)
is proposed for the Kaidu-kongque watershed with the aim of
where x is the first-stage decision variable before the random vari-
representing observed hydrological conditions during a 48-month
able is observed,  is the random variable ( ∈ ˝), and Q(x, ) is the
period (1/1/1998–12/31/2001), where sufficient data are available
optimal value of the following nonlinear programming:
for model calibration and validation. Nash–Sutcliffe coefficient R2
(Nash and Sutcliffe, 1970) was used as the criteria to measure the min q(y, ) (2a)
simulation performance. The initial parameter values in the verifi-
cation process were approximated based on the NAM User’s Manual subject to:
(DHI, 2007). The corresponding simulated daily discharges were W ()W ()y = h() − T ()x (2b)
compared with observed data so that the values of R2 can reach
above or close to the desired value (0.7). The calibrated hydrological y≥0 (2c)

Table 1
Stream flow distribution (in 106 m3 ) and associated probability.

Activity/flow level Low (j = 1) Low-medium (j = 2) Medium (j = 3) High-medium (j = 4) High (j = 5)

 
Flow rate q±
j all
[2459, 2901] [2917, 3500] [3500, 4162] [4185, 4856] [5015, 5708]
Probability (pj ) 0.152 0.478 0.239 0.087 0.044
 Y. Huang et al. / Agricultural Water Management 107 (2012) 74–85 77

where y is the second-stage adaptive decision, which depends on xj± ≥ 0, j = 1, 2, . . . , n (7c)

the realization of random variable, q(y, ) denotes the second-
stage cost function, while T (), W (), h()   ∈ ˝ are random where a± ij
, b±
i
, cj± , dj± and xj± are interval parameters/variables, and
model parameters with reasonable dimensions, which are function the ‘−’ and ‘+’ superscripts represent the lower and upper bounds
of the random variable , the first-stage decision is made before of an interval parameter/variable, respectively. According to Chen
the random variable is observed. Then, when the random variable and Huang (2001), such a problem can then be transformed into
is observed, the discrepant that may exist between h() and T()x two deterministic submodels that correspond to lower and upper
corrected by recourse action that minimizes q(y, ) and satisfies bounds of the objective-function value. This transformation pro-
W() = h() − T()x, y ≥ 0. The pre-regulated cost and the potential cess is based on an interactive algorithm, which is different from
penalty can thus be taken into account. Model (1) can be reformu- the best/worst case analysis (Huang et al., 1995; Chen and Huang,
lated as follows: 2001).
  A manager is responsible for allocating water in an irrigation
max f = cx − E[min q(y, ) T ()x + W ()y = h() ] (3a)
y≥0 season from the upstream of the Kaidu-kongque River to five
main crops in seven subareas (i.e. irrigation zones): cereal, cot-
subject to: ton, oil bearing crops, vegetable and forge. Competition for water
Ax ≥ b (3b) exists among multiple crops when available water is less than
the demands. Farmers want to know where they stand by pro-
x≥0 (3c) viding foreseen information that is needed to make decisions for
Let the random variable  take discrete values l with a prob- various activities and investments. Given an irrigation target that
ability level pl , where l = 1, 2,. . ., n. It is assumed that pl > 0 and is promised to each crop in each subarea, if this water is deliv-
n ered, the resulting net benefit to the local economy per unit of
p = 1. The expected value of the second-stage optimization
l=1 l
problem can be expressed as: water allocated is estimated to be Bi . However, if the promised
water is not delivered, the user will have to either obtain water

n
from higher-priced alternatives or curb their irrigation plans. For
EQ (x) = E[Q (x, w)] = pl Q (x, l ) (4)
example, farmers may not be able to conduct irrigation for each
l=1 crop type as planned. These actions will result in increased costs
For each realization of random variable l, a second-stage deci- (due to increased water price) or decreased benefits (because of
sion is made, which is denoted by yl . The second-stage optimization reduced production) for regional agriculture. It is thus necessary
problem can then be written as: for the available water be effectively allocated to minimize any or
all associated penalties. Here, the associated penalties (Ci ) mean
min q(yl , l ) (5a) the acquisition of water from higher-priced alternatives, and the
subject to: negative consequences are generated from the curbing of the devel-
opment plans (Huang, 2011; Jing and Chen, 2011; Li et al., 2011).
W (l ) = h(l ) − T (l )x, ∀l = 1, 2, . . . , n (5b) The problem can be formulated as maximizing the expected value
yl ≥ 0 (5c) of net system benefit using the two-stage programming optimiza-
tion model. Moreover, the impacts of marginal utility are desired
Through combining models (4) and (5), model (6) can be refor- to reflect since the benefit and penalty are expressed as inexact
mulated as follows: linear functions of water demand and shortage. This may lead to

n the relevant objective function nonlinearity (i.e. quadratic forms).

max f = cx − pl q(yl , l ) (6a) Consequently, based on TIQP method, the study problem can be
l=1 formulated as follows:

subject to:

max f ± = (a± W ± + b±
ki ki ki
±
)Wki
Ax ≥ b (6b)
k=1 i=1
W (l ) = h(l ) − T (l )x, ∀l = 1, 2, . . . , n (6c)

tn m
± ± ± ±
x≥0 (6d) − pj (cki Wki + dki )Ski (8a)
k=1 i=1 j=1
yl ≥ 0 (6e)
subject to:
The TSP model can handle uncertainties expressed as probability
distributions; however, it has difficulties in deal with uncertainties

expressed as discrete intervals and nonlinearities in cost objective q±

j
≥ (Wi± − Sij± ) ∀k, i, j (8b)
to reflect the economies of scale in many environmental problems k=1 i=1
(Li et al., 2009). IQP allows the uncertainties expressed as inter- ± ±
Wki max ≥ Wki ≥ Skij ∀k, i, j (8c)
val values to be communicated into the optimization process and
resulting solution. An IQP problem can be formulated as follows ±
Skij ≥ 0 ∀k, i, j (8d)
(Chen and Huang, 2001):
±
where f± is net system benefit (RMB¥); Bki is net benefit to user i

n
2
max f ± = [cj± xj± + dj± (xj± ) ]
3 3
per m of water allocated (RMB¥/m ) for subarea k (first-stage rev-
(7a)
±
j=1
enue parameter); Wki is targeted water allocation that is promised
3 ±
to user i (m ) for subarea k (first-stage decision variable); Cki is
subject to: 3 ± ±
loss to user i per m of water not delivered for subarea k, Cki > Bki

n 3 ±
(RMB¥/m ) (second-stage cost parameter); Ski is shortage of water
a± x± ≤ b±
ij j i
, i = 1, 2, . . . , m (7b) to user i under flow level j for subarea k, the amount by which Wki ±

j=1 3
is not met when the seasonal flow is qj (m ) (second-stage decision
78 Y. Huang et al. / Agricultural Water Management 107 (2012) 74–85

Rainfall Uncertainty data

Allocation method
Runoff Model (Spatial and temporal)

Probability distribution Discrete intervals Quadratic program

Interval Two-stage Two-stage

Two-stage Two-stage Stochastic interval-quadratic model
Stochastic
Programming Programming
Programming (TIQP)

TIQP upper bound

submodel
Solution
(allocation methods)
TIQP lower bound
submodel

Fig. 3. Schematic of two-stage interval-quadratic model (TIQP).

variable); qj is amount of seasonal flow under flow level j (m3 ) (ran- Wki max ≥ Wki − + Wki zki ≥ Skij
±
≥ 0 ∀k, i, j (9c)
dom variable); Wki max is maximum allowable allocation amount of
user i (m3 ) for subarea k; pj is the probability of occurrence of flow 0 ≤ zki ≤ 1 ∀k, i (9d)
level j; m is the total number of water users; n is the total number
of flow levels; k is the subareas, with k = 1 for Hejing County, k = 2 When Wi± are known as deterministic values, model (9) can be
for Hoxud County, k = 3 for Luntai County, k = 4 for Yanqi County, transformed into two deterministic sub-models which correspond
k = 5 for Bohu County, k = 6 for Korla City and k = 7 for Yuli County; i to the upper and lower bounds of the desired objective-function
is water user, with i = 1 for cereal, i = 2 for cotton, i = 3 for oil bearing value, respectively. The transformation process is based on the
crops, i = 4 for vegetable, i = 5 for forge; j is flow levels, with j = 1 methods of interactive and derivative algorithms (Huang et al.,
representing low flows, j = 2 representing low-medium flows, j = 3 1995; Chen and Huang, 2001). In this study, since the penalty coeffi-
representing medium flows, j = 4 representing medium-high flows, cients of ci± and di± that are related to the interval decision variables
and j = 5 representing high flows. a± i
is the slope of the benefit curve (Sij± ) have same sign (i.e. ci± > 0 and di± > 0), according to Chen
for water demand for user i (m3 ) during period t (a± i
< 0); b± i
is the and Huang (2001), all Sij± correspond to f+ , and all Sij+ correspond
intercept on the benefit axis for user i during period t (b± i
> 0); ci± is to f− . Then, sub-model (10) corresponding to f+ (i.e. most desirable
3
the slope of penalty curve for water shortage for user i (m ) during system objective) can be firstly formulated as follows:
period t (ci± > 0); di± is the intercept on the penalty axis for user i
during period t (di± > 0).

n
In model (8), water allocation target Wi± (i.e. water-allocation max f + = [a+ −
(Wki
2
+ Wki zki ) + b+ (Wki − + Wki zki )]
ki ki
target that is promised to user i by the authority) need to be deter- k=1 i=1
mined at the first-stage before the random inflows are known.
Let Wit± = Wit− + Wit zit , where Wit = Wit+ − Wit− and zit ∈ [0, 1],

tm n
− −2 − −
where zit are decision variables that are used for identifying an − pj (cki Skij + dki Skij ) (10a)
optimized set of target values (Wit± ) in order to support the related k=1 i=1 j=1

policy analyses. Thus, model (8) can be converted to:

subject to:

n
2
max f ± = [a± −
(Wki + Wki zki )
ki

n
k=1 i=1
q+
j
≥ (Wki − + Wki zki − Skij
−
) ∀k, i, j (10b)

t n m
k=1 i=1
+ b±
ki
(Wki − + Wki zki )] − ± ±2
pj (cki ± ±
Skij + dki Skij ) (9a)
k=1 i=1 j=1 Wki max ≥ Wki − + Wki zki ≥ Skij
−
≥ 0 ∀k, i, j (10c)

subject to:
0 ≤ zki ≤ 1 ∀k, i (10d)

n
q±
j
≥ (Wki − + Wki zki − Skij
±
) ∀k, i, j (9b) Let Sij− opt
and zi opt be solutions of sub-model (10). The
k=1 i=1 optimized water-allocation targets are Wi± opt = Wi − + Wi zi opt .
 Y. Huang et al. / Agricultural Water Management 107 (2012) 74–85 79

Table 2
Allowable water allocations (in 106 m3 ) and related economic data (in RMB¥/106 m3 ).

Activity/user Cereal Cotton Oil bearing crops Vegetable Forge

Maximum allowable allocation (Wi max )

Hejing County 48 34 12 75 14
Hoxud County 36 160 3 26 3
Luntai County 75 220 3 19 30
Yanqi County 75 9 15 65 22
Bohu County 30 60 7 55 4
Korla City 50 330 3 15 1
Yuli County 2 330 3 3 2

Water allocation target (Wi± )

Hejing County [36.40,47.77] [13.33,33.04] [6.69,11.50] [23.65,71.10] [11.72,12.99]
Hoxud County [21.27,35.10] [50.76,157.82] [1.64,2.70] [13.92,25.30] [0.42,2.12]
Luntai County [56.60,74.50] [92.32,219.83] [0,0.29] [6.59,17.89] [0,29.46]
Yanqi County [36.40,73.42] [2.07,7.70] [6.16,14.55] [27.25,61.75] [14.43,21.56]
Bohu County [23.04,28.49] [8.85,58.21] [3.64,5.16] [25.30,54.42] [0,3.22]
Korla City [19.12,45.08] [165.27,320.35] 0 [8.46,14.45] [0,0.42]
Yuli County [0,1.84] [170.40,320.19] 0 [1.65,2.54] [0,1.53]

Net benefit when water demand is satisfied (Bi± )

Lower bound −374x + 186,023 −594x + 1,261,734 −5220x + 506,394 −2389x + 1,110,176 −2301x + 451,937
Upper bound −449x + 223,228 −712x + 1,514,082 −6264x + 607,673 −2867x + 1,332,211 −2761x + 542,324

Reduction of net benefit when demand is not delivered (Ci± )

Lower bound 674x + 284,615 1069x + 1,926,358 9396x + 774,783 4300x + 1,685,441 4141x + 691,463
Upper bound 898x + 379,487 1425x + 2,568,478 12529x + 1,033,043 5734x + 2,247,254 5522x + 921,951

+
Then, we have the sub-model corresponding to f− : Let Skij opt
be solutions of sub-model (11). Thus, interval solu-
tions for the ITQP-primal model under the optimized water

n
2
allocation plan can be obtained:
max f − = [a− −
(Wki + Wki zki ) + b− (Wki − + Wki zki )] ± − +
ki ki Skij opt
= [Skij opt
, Skij opt
] ∀k, i, j (12a)
k=1 i=1
± − +

m fopt = [fopt , fopt ] (12b)

+ +2 + +
− pj (cki Skij + dki Skij ) (11a)
A±
kij opt
±
= Wki opt
±
− Skij opt
∀k, i, j (12c)
k=1 i=1 j=1

where A±kij opt

is the optimal water allocation to user i under
subject to: flow level j, which is obtained by subtracting the optimal water
±
shortages Skij opt
(the second-stage decision variables) from the
±

n optimized target water allocation Wki opt

(the first-stage decision
q−
j
≥ (Wki − + Wki zki − Skij
+
) ∀k, i, j (11b) variables). Fig. 3 illustrates the general framework of the TIQP. The
k=1 i=1 modeling approach is based on the optimization techniques and
stream flow simulation model.
Wki max ≥ Wki − + Wki zki ≥ Skij
+
≥ 0 ∀k, i, j (11c) In the study area, as the irrigation water consists of both sur-
face water and ground water, the surface water irrigation target is
a proportion of the total irrigation target. The proportion in each
0 ≤ zki ≤ 1 ∀k, i (11d) subarea is a deterministic value, which is equal to the average

Fig. 4. Optimized water-allocation targets under basic scenario (Symbols “1, 2, 3, 4, 5, 6, and 7” denote “Hejing County, Hoxud County, Luntai County, Yanqi County, Bohu
County, Korla City and Yuli County”, respectively).
80 Y. Huang et al. / Agricultural Water Management 107 (2012) 74–85

Fig. 5. Water allocation patterns under basic scenario [(a) low inflow; (b) medium inflow; (c) high inflow].

proportion of surface water to total irrigation water in subarea demand (R) of the lower reaches of the Tarim River Basin is about
in recent year. Modeling inputs are estimated based on the sta- 10.51 × 106 m3 in the planning period (Tarim Basin Management
tistical yearbook of Xinjiang Uygur Autonomous Region in 2009 Bureau, 2009). Thus, we have: q±
j
= (q±
j all
− R) × E (Table 2).
and presented in Table 1. The irrigation quotas are [347,512] m3 /ha
for cereal, [326,552] m3 /ha for cotton, [273,391] m3 /ha for oil bear-
ing crops, [336,499] m3 /ha for vegetable, and [415,566] m3 /ha for 4. Results and discussion
forge. In the Kaidu-kongque watershed, about 92% of the water
consumption was contributed by the agriculture user. However, 4.1. Result analysis
due to relatively backward water infrastructure, the agricultural
irrigation rate (E) is very low in most irrigation districts in South Different scenarios for water diversion can be defined based
Xinjiang, China. In the Kaidu-kongque watershed, water efficiency on different water-allocation targets (Wit± ) which could reflect the
of canal system (EC ) is 0.4, and water efficiency in field (EF ) is 0.68. local decision makers’ attitudes to water resources availabilities.
Therefore, the agricultural irrigation rate (E) is 0.27 (i.e. E = EC × EF ) In scenario 1, the priority for different crops during the irrigation
(Dong et al., 2005). There are great potentialities on water saving process was not considered in the TIQP (denoted as base condi-
in this irrigation region. In addition, ecosystem recovering water tion, S1); in scenario 2, food security problem is firstly considered.
 Y. Huang et al. / Agricultural Water Management 107 (2012) 74–85 81

Fig. 6. Optimized water allocation patterns under low-medium inflow.

Thus, we give priority to the cereal sector and 20% of lower demand of z21 opt = 0 indicates that the optimized water-allocation target
is required to be satisfied (denoted as S2); in scenario 3, give pri- reaches its lower bound (i.e. 21.27 × 106 m3 ); this also implies that
ority to the development of animal husbandry, and 20% of lower the manager has a conservative attitude toward water allocation to
demand from the forage sector is required to be satisfied (denoted this crop despite of water flow level. Optimized water-allocation
as S3); in scenario 4, give priority to both cereal and forage sec- targets for oil bean crop and vegetable would be 1.99 × 106 and
tors and 10% of lower demands from the two sectors are required 22.34 × 106 m3 (i.e. z23 opt = 0.33 and z24 opt = 0.74).
to be satisfied (denoted as S4). The results indicate that solutions Deficits would occur if the available water could not satisfy
± −
for the objective function value and most of the non-zero decision the total demand. The lower-bound of Skij (i.e. Skij ) corresponds
variables are intervals. Generally, solutions presented as intervals to a higher system benefit; in comparison, the upper-bound of
demonstrate that the related decisions should be sensitive to the ± +
Skij (i.e. Skij ) corresponds to a lower system benefit. For exam-
uncertain modeling inputs. In case of insufficient water, the allot- ple, in the Hoxud County, under high flow, there would be zero
ment to the crops should be first decreased but guaranteeing the shortage for all water users; under medium flow, the water short-
minimum promised target. The cotton should be first guaranteed age would be [9.40,21.27] × 106 , 0, [0,1.99] × 106 , [0,10.21] × 106
since it brings the highest benefit when water demand is satisfied; and [0,2.12] × 106 m3 for cereal, cotton, oil bean crops, vegetable
meanwhile, it is subject to the highest penalty if the promised water and forage sectors, respectively; under low flow, the shortage
is not delivered. would be 21.27 × 106 , [25.70,58.41] × 106 , 1.99 × 106 , 22.35 × 106
Fig. 4 presents solutions of the water allocation targets under and 2.12 × 106 m3 for cereal, cotton, oil bean crop, vegetable and
the basic scenario (i.e. S1). The optimized targets for the five forage, respectively. The actual water allocated (as shown in Fig. 5)
crops in seven irrigation zones could be obtained through formula would vary with the optimized target and the probabilistic short-
Wi± = Wi− + Wi zi . For example, in the Hoxud County, the results age under a given stream inflow with an associated probability
of z22 opt = z25 opt = 1.0 indicate that the optimized water-allocation (i.e. A± = Wi± opt − Sij± opt ). For example, the actual water allo-
ij opt
targets for the cotton and forage would equal 157.82 × 106 m3 and
cated to cereal in Hoxud County would be 0, 0, [0,11.88] × 106 ,
2.12 × 106 m3 , corresponding to their upper-bound target values.
[13.36,21.27] × 106 and 21.27 × 106 m3 under low to high flows,
This means that the managers are optimistic for water supply to
respectively. It is indicate that, the managers can make aggres-
the cotton and forge users, and thus promises an upper-bound
sive investment plans for cereal sector under the medium-high and
water quantity to the users. Conversely, for cereal sector, the result
82 Y. Huang et al. / Agricultural Water Management 107 (2012) 74–85

Fig. 7. Optimized water allocation patterns under medium-high inflow.

high flow levels; the investment plans for cereal sectors would be For example, in Hejing county, for the cereal, cotton, oil bear-
reduced under the low to medium flow levels. The other results ing crops, vegetable and forge the water allocated would be (i)
could be analyzed based on the solutions of optimized water allo- 0, [6.87,33.04] × 106 , 0, [0,13.44] × 106 and 0 m3 under scenario
cation targets and shortages. 1; (ii) 7.28 × 106 , [4.07,33.04] × 106 , 0, [0,10.86] × 106 and 0 m3
±
The solution of fopt = RMB ¥ [518, 1248] × 106 provides two under scenario 2; (iii) 0, [6.49,33.04] × 106 , 0, [0,13.09] × 106 and
extremes of the system benefit under the optimal water alloca- 2.34 × 106 m3 under scenario 3; (iv) 3.64 × 106 , [5.27,33.04] × 106 ,
tion pattern. As the actual value of each variable or parameter 0, [0,11.97] × 106 and 1.17 × 106 m3 under scenario 4. Fig. 7 shows
varies within its two bounds, the system benefit may change cor- the comparison result of optimized water allocation patterns
− +
respondingly between fopt and fopt with varied reliability levels. for different agricultural sectors under medium-high inflow. The
Planning for the lower-bound of the objective function value will results indicate that, interactions exist among supplies for multi-
lead to a lower system benefit with a lower risk of violating the ple competing users when water deficits occur, due to benefit and
water-allocation target. Conversely, planning with a higher sys- penalty effects on the water-allocation patterns. Under scenario 1,
tem benefit will correspond to a higher possibility of violating the the water allocation to cereal would firstly be decreased in case of
water-allocation target when approaching the upper-bound of the insufficient water (due to its lowest benefit); then shortage would
objective-function value. Therefore, there is a tradeoff between the be passed to the forge. However, under scenario 2 (i.e. priority to
water-allocation benefit and the system-failure risk. the cereal sector were considered), 20% of lower demand of the
cereal sector would be satisfied firstly and, then, water allocations
4.2. Policy analysis to cereal and forge would be decreased in case of insufficient water;
water shortages for other sectors would become increasing due
The main advantage of the two-stage programming is its to the irrigation priority of cereal were considered under scenario
capability of incorporating multiple policies of water resources 2. The priorities of irrigation were contributed to the forge sector
management within the optimization framework. The solutions under scenario 3 and cereal and forage sectors under scenario 4;
for the optimized water-allocation patterns under different sce- amount of promised water must be allocated to the priority sectors
narios were also obtained. Fig. 6 shows the comparison of the firstly. Under scenarios 2–4, some of water would still be allo-
optimized water allocation patterns under low-medium inflow. cated to the cereal or forge crops even under very disadvantageous
 Y. Huang et al. / Agricultural Water Management 107 (2012) 74–85 83

Table 3
Solutions of ITSP model under basic scenario.

User Cereal Cotton Oil bearing crops Vegetable Forge

Target (Wi± opt ) (106 m3 )

Hejing 36.40 33.04 11.50 23.65 12.98
Hoxud 21.27 157.81 2.69 13.92 2.12
Luntai 56.60 219.83 0.29 6.59 29.46
Yanqi 36.40 7.70 14.54 27.24 21.56
Bohu 23.04 58.21 5.16 25.30 3.22
Korla 19.12 320.35 0 8.46 0.42
Yuli 0 320.19 0 1.65 1.53

Allocation (A±
ij opt
) under a flow level of (106 m3 )
Hejing Low (j = 1) 0 0 0 0 0
Low-medium (j = 2) 0 0 0 0 0
Medium (j = 3) 0 33.04 [0,11.49] 0 0
High-medium (j = 4) [0,36.40] 33.04 11.49 [0,23.65] [0,12.98]
High (j = 5) 36.40 33.04 11.49 23.65 12.98

Hoxud Low (j = 1) 0 0 0 0 0
Low-medium (j = 2) 0 20.01 0 0 0
Medium (j = 3) 0 157.82 [0,2.70] 0 0
High-medium (j = 4) [0,21.27] 157.82 2.70 [0,13.92] [0,2.12]
High (j = 5) 21.27 157.82 2.70 13.92 2.12

Luntai Low (j = 1) 0 8.34 0 0 0

Low-medium (j = 2) 0 219.83 0 0 0
Medium (j = 3) 0 219.83 [0,0.29] 0 0
High-medium (j = 4) [0,41.00] 219.83 0.29 [0,6.59] [0,29.46]
High (j = 5) 56.60 219.83 0.29 6.59 29.46

Yanqi Low (j = 1) 0 7.70 0 0 0

Low-medium (j = 2) 0 7.70 0 0 0
Medium (j = 3) 0 7.70 [0,14.55] [0,15.36] 0
High-medium (j = 4) 0 7.70 14.55 [24.26,27.25] [0,21.56]
High (j = 5) 36.40 7.70 14.55 27.25 21.56

Bohu Low (j = 1) 0 58.21 0 0 0

Low-medium (j = 2) 0 58.21 0 0 0
Medium (j = 3) 0 58.21 [0,5.16] [0,25.30] 0
High-medium (j = 4) [0,23.04] 58.21 5,16 25.30 [0,3.22]
High (j = 5) 23.04 58.21 5.16 25.30 3.22

Korla Low (j = 1) 0 320.35 0 0 0

Low-medium (j = 2) 0 320.35 0 0 0
Medium (j = 3) 0 320.35 0 [0,8.46] 0
High-medium (j = 4) [0,19.12] 320.35 0 8.46 [0,0.42]
High (j = 5) 19.12 320.35 0 8.46 0.42

Yuli Low (j = 1) 0 [149.49,320.19] 0 0 0

Low-medium (j = 2) 0 [85.88,320.19] 0 0 0
Medium (j = 3) 0 [149.35,320.19] 0 [0,1.65] 0
High-medium (j = 4) 0 320.19 0 1.65 [0,1.53]
High (j = 5) 0 320.19 0 1.65 1.53

Net benefit (f± ) (RMB¥106 ) [7.14,1763]

condition (i.e. low inflow). This demonstrates that water- willingness to accept a low system benefit could guarantee meet-
management policies that consider irrigation priorities for cereal ing the minimum water supply for cereal and forge sectors, while a
and forge crops are useful for minimizing unfair competition among strong desire to acquire a high system benefit could run into a high
multiple agricultural users, and thus can protect the farmland risk of water shortage for cereal and forge sectors. The above analy-
resources even though merely single objective (i.e. economic ben- ses demonstrate that different water-allocation policies would lead
efit) was considered by the modeling formulation. to varied economic consequences and system-failure risks.
Solutions obtained can not only provide an effective evaluation
for the pre-regulated water policies, but also obtain the associ- 4.3. Comparison with ITSP
ated economic implications. Different water management policies
would result in varied system benefits (when targeted water was The problem can also be solved through the ITSP method with-
allocated to users) and penalties (when the promised water was out considering the effects of marginal utility on benefit and penalty
not delivered). System benefits would be RMB¥[473,1212] × 106 , (Huang and Loucks, 2000). The results for ITSP are presented in
RMB¥[512,1243] × 106 and RMB¥[910,1229] × 106 under sce- Table 3. The net system benefit would be RMB¥[7.1,1763] × 106
narios 2–4; however, when the irrigation priority was not through the ITSP model, which has a higher middle value and wider
considered (under scenario 1), the system benefit would be interval than that through the TIQP (i.e. RMB¥[518,1248] × 106 ).
RMB¥[518,1248] × 106 . The results indicate that both lower- and Generally, ITSP is based on an assumption that the effects of
upper-bound benefits under scenarios 2 and 3 would be smaller marginal utility are negligible and, thus, the relevant objective func-
±
than those under scenario 1; system benefit fopt under scenario 4 tion is represented as linear form. However, the impacts of marginal
has narrower interval range, compared with scenario 1. Therefore, utility may be significant when the limited water resources will be
84 Y. Huang et al. / Agricultural Water Management 107 (2012) 74–85

allocated among multiple competing users. This may make the ITSP editors and the anonymous reviewers for their insightful comments
method less realistic, leading to over-simplification of the relation- and suggestions.
ship between benefit and demand, as well as relationship between
penalty and shortage. In comparison, TIQP can effectively deal with
such issues where in marginal utility affect the benefit/cost coeffi- References
cients in a mathematic programming problem.
Chen, M.J., Huang, G.H., 2001. A derivative algorithm for inexact quadratic program-
application to environmental decision making under uncertainty. European
Journal of Operational Research 128, 570–586.
5. Conclusions DHI, 2007. MIKE 11 User’s and Reference Manual. DHI Water & Environment, Hor-
sholm, Denmark.
A two-stage interval-quadratic programming method was Dong, X.G., Deng, M.J., Zhou, J.L., Zhong, J.P., 2005. On exploitation of water resources
and soil salinization in irrigation area of XinJiang Plain. Jorunal of Irrigation and
developed for agricultural water resources planning and man-
Drainage 24 (5), 14–17.
agement. TIQP is based on two-stage stochastic programming, Ferrero, R.W., Rivera, J.F., Shahidehpour, S.M., 1998. A dynamic programming
interval-quadratic programming, and a lumped rainfall runoff two-stage algorithm for long-term hydrothermal scheduling of multireservoir
systems. IEEE Transactions on Power Systems 13, 1534–1540.
model. It can deal with nonlinearities in the cost/benefit objec-
Hillier, F.S., Lieberman, G.J., 1986. Introduction to Operations Research, fourth ed.
tive and uncertainties expressed as probability density functions Holden-Day, Oakland, CA.
and discrete intervals, and can also support the analysis of pol- Huang, G.H., Baetz, B.W., Patry, G.G., 1995. Gery quadratic programming and its
icy scenarios that are associated with economic penalties when application to municipal waste management planning under uncertainty. Engi-
neering Optimization 23, 210–223.
the promised targets are violated. The optimization and simu- Huang, G.H., Loucks, D.P., 2000. An inexact two-stage stochastic programming model
lation models are coupled with the random parameters in the for water resources management under uncertainty. Civil Engineering and Envi-
optimization model provided by the simulation outcomes, and ronment System 17, 95–118.
Huang, G.W., 2011. Role change of in-channel vegetation with regard to sediment
through the assumed nonlinear relationships in the system. The retention at the decadal scale. Journal of Environmental Informatics 18 (1),
developed TIQP model represents a new effort to obtain optimal 22–27.
water-allocation scheme through maximizing the long-term eco- Huang, Y., Chen, X., Li, Y.P., Willems, P., Liu, T., 2010. Integrated modeling system for
water resources management of Tarim River Basin. Environmental Engineering
nomic and environmental benefits. The developed TIQP is applied Science 27 (3), 255–269.
to a real case of planning agricultural water management in Tarim Jing, L., Chen, B., 2011. Field investigation and hydrological modelling of a subarctic
River Basin, China, where a number of scenarios based on differ- wetland—the Deer River Watershed. Journal of Environmental Informatics 17
(1), 36–45.
ent water resources management policies are analyzed. The results
Lai, Z.X., 2007. Analysis of the frequency curves of annual runoffs at Sanmenxia Dam.
indicate that reasonable solutions have been generated, which pro- Water Conservancy & Electric Power Machinery 29, 216–221.
vide combined information of discrete intervals and probabilistic Li, Y.P., Huang, G.H., Nie, S.L., 2006. An interval-parameter multistage stochastic pro-
gramming model for water resources management under uncertainty. Advances
distributions. Moreover, different policies for water resources plan-
in Water Resources 29, 776–789.
ning would lead to varied allocation targets, shortages, system Li, Y.P., Huang, G.H., Yang, Z.F., Nie, S.L., 2008. Interval-fuzzy multistage pro-
benefits, and penalties. They will help generate desired policies for gramming for water resources management under uncertainty. Resources,
water resources management with maximized economic benefit Conservation and Recycling 52, 800–812.
Li, Y.P., Huang, G.H., Wang, G.Q., Huang, Y.F., 2009. FSWM: a hybrid fuzzy-stochastic
and minimized system-failure risk. water management model for agricultural sustainability under uncertainty.
When a manager is responsible for allocating water from a Agricultural Water Management 96, 1807–1818.
river or an unregulated reservoir to multiple users, he/she has to Li, Y.P., Huang, G.H., Nie, S.L., Chen, X., 2011. A robust modeling approach for regional
water management under multiple uncertainties. Agricultural Water Manage-
obtain related social and economic data firstly, such as the water ment 98, 1577–1588.
demand from various users, water allocation target, and net ben- Li, W., Li, Y.P., Li, C.H., Huang, G.H., 2010. An inexact two-stage water management
efit from water allocation. These values are mainly derived from model for planning cultural irrigation under uncertainty. Agricultural Water
Management 97, 1905–1914.
statistic data and local survey. However, some parameters could be Luo, B., Maqsood, I., Huang, G.H., 2007. Planning water resources systems with
impossible to be acquired with daily interval. On the other hand, interval stochastic dynamic programming. Water Resources Management 21,
when analyzing and evaluating various water management plans 997–1014.
Mainuddin, M., Das Gupta, A., Raj Onta, P., 1997. Optimal crop planning model
designed to distribute the natural unregulated flows over time
for an existing groundwater irrigation project in Thailand. Agricultural Water
and space, it is usually sufficient to consider average conditions Management 33, 43–62.
within discrete time periods. Commonly in river basin model- Mylopoulos, Y.A., Theodosiou, N., Mylopoulos, N.A., 1999. A stochastic optimiza-
tion approach in the design of an aquifer remediation under hydrogeologic
ing, weekly, monthly or seasonal flows are used as opposed to
uncertainty. Water Resources Management 13, 335–351.
daily flows. One main shortage of the developed method is that Maqsood, I., Huang, G.H., Huang, Y.F., Chen, B., 2005a. ITOM: an interval-parameter
it can hardly reflect the dynamic variation of system conditions, two-stage optimization model for stochastic planning of water resources sys-
especially when sequential structure and interaction exist. When tems. Stochastic Environmental Research and Risk Assessment 19, 125–133.
Maqsood, I., Huang, G.H., Yeomans, J.S., 2005b. An interval-parameter fuzzy two-
handling the water allocation problems in a short interval within stage stochastic program for water resources management under uncertainty.
one year (e.g. monthly), many practical problems often involve a European Journal of Operational Research 167, 208–225.
sequence of decisions that interact with periodical outcomes that Nash, J.E., Sutcliffe, J.V., 1970. River flow forecasting through conceptual models.
Part 1—a discussion of principles. Journal of Hydrology 10, 282–290.
evolve over time. Consequently, future works can continue to focus Ping, J., Chen, Y., Chen, B., Howboldt, K., 2010. A robust statistical analysis approach
on developing multistage stochastic programming method, which for pollutant loadings in urban rivers. Journal of Environmental Informatics 16
can extend by permitting revised decisions in each time stage based (1), 35–42.
Reca, J., Roldan, J., Alcaide, M., Lopez, R., Camacho, E., 2001. Optimization model
on the sequentially realized uncertain events. for water allocation in deficit irrigation systems: I. Description of the model.
Agricultural Water Management 48, 103–116.
Rockafellar, R.T., Wets, R.J.B., 1986. A Lagrangian finite generation technique for
Acknowledgements solving linear-quadratic problems in stochastic programming. Mathematical
Programming Study 28, 63–93.
Seifi, A., Hipel, K.W., 2001. Interior-point method for reservoir operation with
This research was supported by the “Western Light” Dr. Program
stochastic inflows. Journal of Water Resources Plan and Management 127,
of Chinese Academy of Sciences (XBBS201010), the Natural Sci- 48–57.
ences Foundation of China (50979001 and 51190095), the Program Sharma, B.R., Minhas, P.S., 2005. Strategies for managing saline/alkali waters for sus-
for New Century Excellent Talents in University (NCET-10-0376), tainable agricultural production in South Asia. Agricultural Water Management
78, 136–151.
and the Program for Changjiang Scholars and Innovative Research Shilman, S.V., 1992. Stochastic quasigradient method for quadratic optimization
Team in University (IRT1127). The authors are grateful to the under dependent observations. Automation and Remote Control 53, 1881–1896.
 Y. Huang et al. / Agricultural Water Management 107 (2012) 74–85 85

Tao, H., Song, Y.D., Zou, S.P., 2007. Variation characteristics of mountainous Xu, H.L., Ye, M., Li, J.M., 2008. The water transfer effects on agricultural development
runoff in the Kaidu river basin, Tianshan mountains. Arid Land Geography 30, in the lower Tarim River, Xinjiang of China. Agricultural Water Management 95,
43–48. 59–68.
Tran, L.D., Schilizzi, S., Chalak, M., Kingwell, R., 2011. Optimizing competitive uses Zhang, X.D., Huang, G.H., Nie, X.H., 2009. Optimal decision schemes for agricul-
of water for irrigation and fisheries. Agricultural Water Management 101, tural water quality management planning with imprecise objective. Agricultural
42–51. Water Management 96, 1723–1731.
Xu, H.L., Ye, M., Li, J.M., 2007. Changes in groundwater levels and the response of Zhou, H.H., Chen, Y.N., Li, W.H., 2010. Soil properties and their spatial pattern in
natural vegetation to transfer of water to the lower reaches of the Tarim River. an oasis on the lower reaches of the Tarim River, northwest China. Agricultural
Journal of Environmental Sciences 19, 1199–1207. Water Management 97, 1915–1922.