Science of the Total Environment 574 (2017) 34–45

Contents lists available at ScienceDirect

Science of the Total Environment

journal homepage: www.elsevier.com/locate/scitoten v

The influence of changes in land use and landscape patterns on soil

erosion in a watershed
Shanghong Zhang a,⁎, Weiwei Fan a, Yueqiang Li a, Yujun Yi b
a
Renewable Energy School, North China Electric Power University, Beijing 102206, China
b
State Key Laboratory of Water Environment Simulation and Pollution Control, School of Environment, Beijing Normal University, Beijing 100875, China

H I G H L I G H T S G R AP HI C A L A B S T R AC T

• The relation between soil erosion and

land use patterns was studied.
• Multiple linear regression method was
used to obtain the relation.
• The four main contributing landscape
indices were highlighted by regression
analysis.
• The large patch index was the most im-
portant landscape index affecting soil
erosion.
• It is not feasible to obtain the soil ero-
sion amount from landscape metrics
alone.

a r t i c l e i n f o a b s t r a c t
Article history: It is very important to have a good understanding of the relation between soil erosion and landscape patterns so that
Received 13 May 2016 soil and water conservation in river basins can be optimized. In this study, this relationship was explored, using the
Received in revised form 22 August 2016
Liusha River Watershed, China, as a case study. A distributed water and sediment model based on the Soil and Water
Accepted 4 September 2016
Assessment Tool (SWAT) was developed to simulate soil erosion from different land use types in each sub-basin of
Available online 12 September 2016
the Liusha River Watershed. Observed runoff and sediment data from 1985 to 2005 and land use maps from 1986,
Editor name: D. Barcelo 1995, and 2000 were used to calibrate and validate the model. The erosion modulus for each sub-basin was calculated
from SWAT model results using the different land use maps and 12 landscape indices were chosen and calculated to
Keywords: describe the land use in each sub-basin for the different years. The variations in instead of the absolute amounts of
Landscape pattern index the erosion modulus and the landscape indices for each sub-basin were used as the dependent and independent var-
Soil erosion iables, respectively, for the regression equations derived from multiple linear regression. The results indicated that the
SWAT model variations in the erosion modulus were closely related to changes in the large patch index, patch cohesion index, mod-
Liusha River Watershed ified Simpson's evenness index, and the aggregation index. From the regression equation and the corresponding land-
Regression analysis
scape indices, it was found that watershed erosion can be reduced by decreasing the physical connectivity between
patches, improving the evenness of the landscape patch types, enriching landscape types, and enhancing the degree
of aggregation between the landscape patches. These findings will be useful for water and soil conservation and for op-
timizing the management of watershed landscapes.
© 2016 Elsevier B.V. All rights reserved.

⁎ Corresponding author.
E-mail address: zhangsh928@126.com (S. Zhang).

http://dx.doi.org/10.1016/j.scitotenv.2016.09.024
0048-9697/© 2016 Elsevier B.V. All rights reserved.
35 S. Zhang
S. Zhang
et al.
et /al.
Science
/ Science
of the
of the
Total
Total
1. Introduction
Water and soil loss in catchments leads to a range of problems, such
as extensive sediment deposition in river channels and reservoirs, de-
clines in soil fertility, and environmental pollution (Munro et al.,
2008). It is, therefore, important to study the various factors that influ-
ence soil and water loss to develop and implement effective measures
for soil and water conservation at the watershed level. Watershed ero-
sion is related to many factors, including rainfall processes and variation
in the watershed surface, such as land use, soil, terrain, and topography
(Ni et al., 2008; Ochoa et al., 2016; Saedi et al., 2016). Rainfall, soil types,
and topography in a watershed do not change significantly in the short-
term, which means that human activities, through changes in land use,
are the main influence on changes in watershed erosion.
There are two main ways by which land use can modify watershed
erosion. Watershed erosion may be changed by modifying the land
use type; for example, cultivated land may be returned to forest and
grassland, so that erosion decreases (Zhang et al., 2014). The land use
distribution pattern, known as landscape and land use morphology,
may also be changed, resulting in changes in erosion. Land use types
and landscape morphology both play an important role in soil erosion
(Qi et al., 2012; Wang et al., 2009a).
Many studies have examined changes in watershed erosion related
to changes in land use types. For example, Hao et al. (2004) used the
Soil and Water Assessment Tool (SWAT) in the Yellow River basin and
found that forestation decreased the sediment yield, and that increases
in the amount of land used for agriculture increased the sediment yield.
Ouyang et al. (2010) established the relation between soil erosion and
sediment yield with the normalized difference vegetation index and
showed that the vegetation status had a significant impact on soil ero-
sion and transport. Ciampalini et al. (2012) studied soil erosion in re-
sponse to historical changes in agricultural and soil conservation
practices in the Aksum area, and found that large areas of arable land
were converted into grazing land resulting in significant increases in
soil loss. Durán-Zuazoa et al. (2013) explored the impact of land-cover
types on soil erosion and runoff, and, by comparing different land use
types, demonstrated that erosion and runoff were lower from forest
dominated by Pinus than from abandoned farmland. There is general
agreement on how changes in land use types influence erosion; numer-
ous studies have shown that soil and water losses are generally lower
from forest land than from cultivated land, grassland, and pasture
(Feng et al., 2010; Nunes et al., 2011; Alatorre et al., 2012).
Various researchers have studied the influence of land use morphol-
ogy on erosion, and have provided case studies that report analysis of
landscape patterns (Xiao and Ji, 2007; Liu and Yu, 2010; Huang et al.,
2013; Liu and Lu, 2011; Silva et al., 2015). Wei et al. (2006) studied
the relationship between landscape patterns and soil erosion in an agri-
cultural watershed in northern China where Mollisols dominated and
found that the multiple correlation coefficients between the nine select-
ed landscape metrics and erosion were higher than the single-factor
coefficients between any of the individual landscape indices and ero-
sion. Wang et al. (2009b) used GIS and analysis of landscape indices to
explore the relation between soil erosion and landscape patterns in
the Wuyuer River and concluded that for soil erosion control and man-
agement, it was very important to regulate landscape patterns in crop-
land, grassland, and forestland. Ouyang et al. (2009) selected landscape
metrics at the patch level to study the response of soil erosion dynamics
to landscape patterns and indicated that contiguous grassland patches
reduced soil erosion yield and that larger areas of water led to more
soil erosion. Shi et al. (2013) used partial least-squares regression to
link land cover patterns to soil erosion and sediment yield in a water-
shed and found that landscape characteristics, such as the Shannon di-
versity index (SHDI), aggregation index (AI), largest patch index (LPI),
contagion (CONTAG), and patch cohesion index (COHESION) were the
main indices that influenced watershed soil erosion and sediment
yield at the landscape level. Landscape indices have been widely applied
Environment
Environment
574574
(2017)
(2017)
34–45
34–45 35
to examine the response of soil erosion to landscape patterns. However,
there is no consistent standard method of selecting landscape metrics,
which means that studies arrive at different conclusions, and to date
there is no universal conclusion. For example, since soil erosion is affect-
ed by multiple factors, such as precipitation, land use, soil, and topogra-
phy, the question remains as to whether soil erosion can be quantified
directly from landscape indices alone.
The objectives of this study were to (1) determine the influence
of land use patterns on watershed erosion, (2) derive a relation be-
tween watershed erosion and landscape metrics at the landscape
level by combining ecologically significant landscape metrics, and
(3) establish a robust basis for soil and water conservation in
watersheds.
2. Study area
The Liusha River is in the transition zone of the Sichuan Basin and the
Western Sichuan Plateau. It is a tributary of the Dadu River, and it runs
from north to south through the city of Ya'an, Sichuan Province, China.
The Liusha River Watershed covers 1150 km2 and lies between 102°
16′–102°45′E and 29°19′–29°43′N. Its source is in the Shanzi Moun-
tains, which has an elevation of 3300 m, and an elevation difference of
2545 m between the highest and lowest points in the watershed
(Fig. 1).
The Liusha River Watershed has a subtropical climate. The water-
shed is characterized by mountain ranges, steep slopes, and deep valleys
(Xiang and Tang, 2006). The climate is dry and windy, and there is abun-
dant annual sunshine. The average annual precipitation is 726 mm, of
which 80%–90% falls during the monsoon season (June–September).
The average annual runoff is 469 mm, the average annual sediment
transport rate is 59.9 kg/s, and topsoil losses are as much as
2.9 million m3/year. The valley area is the focus of agriculture in the wa-
tershed, and long-term soil erosion has resulted in decreases in soil fer-
tility; ongoing land reclamation and human activities in the basin have
resulted in severe soil degradation (Yang, 2012).
3. Methods
3.1. General study framework
Regression analysis was used to explore the impact of landscape pat-
terns on soil erosion. The sub-basin was used as the analysis unit to en-
sure there was sufficient sample data. The soil erosion of each sub-basin
was simulated using the SWAT model, which is a continuous-time,
semi-distributed, process-based watershed scale model that was devel-
oped to predict the impact of land management practices on water, sed-
iment, and chemicals (Arnold et al., 1998; Neitsch et al., 2011). The
landscape metrics of each sub-basin were calculated with Fragstats ver-
sion 4.0, a widely accepted tool for quantifying landscape indices
(McGarigal et al., 2012). Multiple linear regression was used in Statisti-
cal Product and Service Solutions (SPSS) to fit the sample data, and then
the implications of the regression equation were interpreted and
discussed. Particular consideration was given to the following topics in
this study:
(1) Soil erosion simulation
The ArcGIS-ArcView extension and graphical user input inter-
face of SWAT, ArcSWAT, was used to establish and validate the dis-
tributed hydrological model of the Liusha River Watershed for
runoff and sediment data from 1986 to 2005, and the soil for erosion
of each sub-basin was simulated on the basis of land use maps that
related to three different years (1986, 1995, and 2000). The erosion
modulus (= erosion/area) was used to eliminate the influence of
area on sub-basin soil erosion.
(2) Calculation of the sub-basin landscape pattern
The key landscape pattern metrics were selected at the landscape
level, and then the landscape pattern indices of the three land use
36 S. Zhang
S. Zhang
et al.
et /al.
Science
/ Science
of the
of the
Total
Total
Environment
Environment
574574
(2017)
(2017)
34–45
34–45 36

Fig. 1. Liusha River Watershed, China.

scenarios (1986, 1995, and 2000) were calculated for each sub-basin
with Fragstats version 4.0 software.
(3) Multiple linear regression analysis
The variation in the erosion modulus and the landscape metrics was
calculated for each sub-basin. The variation in the erosion modulus was
taken as the dependent variable and the variation in the landscape indi-
ces was considered as the independent variable. The regression equa-
tions from multiple linear regressions were used to examine the
relation between erosion and landscape metrics.
3.2. Data collection
Sources of the spatial and attribute data and details about the accu-
racy of the analysis are provided in Table 1. Land use maps for three dif-
ferent years were used in this study (Fig. 2). On these maps, the land use
was reclassified as Forest-Mixed (forest land and open forest land),
Range-Brush (shrub land), Range-Grasses (high, middle, and low cover-
age grassland), Water (washland), Urban (urban land), Rice (mountain
paddy field and Pingba paddy field), and Winter-Wheat (mountain dry
land and Pingba dry land) in ArcGIS. So the land use types in the study
area were reclassified to correspond with the criteria of the SWAT
model database (Winchell et al., 2013). Before and after the reclassifica-
tion, the scale and resolution of the land use maps were unchanged.
3.3. Model setup
The meteorological and soil databases were assembled. There was a
threshold (8000 ha) for the source area in SWAT to derive the stream
network and the Liusha River Watershed was divided into 63 sub-
basins by calculation of sub-basin parameters. The 1995 land use map
was set as the baseline case when the SWAT model was built. The
model was calibrated and validated with 26 years of observed data at
the watershed outlet. The first five years were used as the warm-up pe-
riod (1980–1984) of the model to ensure that the simulation was pre-
cise and stable. The calibration and validation simulation periods for
runoff were from 1985 to 1994 and from 1995 to 2005, respectively.
The land use map was unchanged during these periods until a
satisfactory model was built. The results of the calibration and valida-
tion of the runoff and sediment simulations are shown in Figs. 3 and 4,
respectively.
The coefficient of determination (R2) was used to evaluate the rela-
tion between the measured and the simulated runoff and sediment

Table 1
SWAT input data and measured runoff and sediment yield data.

Data type Name Accuracy Source

Spatial data (measured) DEM 30 m resolution http://www.gscloud.cn

Meteorological data (simulated)
Hydrological data (measured)
Soil attribute database (simulated)
Land use map of Sichuan Province 1:100,000 http://www.geodata.cn
National soil type map 1:1,000,000 http://westdc.westgis.ac.cn
Precipitation, Max & Min temperature, Daily Rainfall data: Six rainfall stations;
Relative humidity, Wind speed Meteorological data: Global Weather
Data for SWAT: http://globalweather.tamu.edu/
Runoff, Sediment Daily Liusha River station
Volume weight of soil, Harmonized World Soil Database
Wilting coefficient, Hydraulic conductivity
37 S. Zhang
S. Zhang
et al.
et /al.
Science
/ Science
of the
of the
Total
Total
Environment
Environment
574574
(2017)
(2017)
34–45
34–45 37

Fig. 2. Land use maps for 1986, 1995, and 2000 for the Liusha River Watershed.
yield data, and the Nash–Sutcliffe efficiency (Ens) was used to compare and + 22%, respectively, which indicates model underestimation bias.
the modeled values with the measured values (Nash and Sutcliffe, While during the calibration and validation periods for sediment yield,
1970). These are computed as follows: the PBIAS was + 22% and − 15%, respectively, which was a satisfactory
result.
2 32
n 2 From the results of three evaluation index, the simulation results in-
2 6 ∑ i¼1 Q oi −Q avg Q pi −Q pavg dicated that the hydrological and sediment yield models developed in
R ¼ 7
n 2
∑
4qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi5 ð1Þ
n 2
∑ i¼1 Q oi −Q avg i¼1 Q pi −Q pavg
this study were suitable for analyzing variability in sediment yield
from this watershed.
n 2
∑i ¼1 Q −Q
oi pi
Ens ¼ 1− 3.4. Calculation of landscape pattern metrics
∑ni¼1 Q oi −Q avg ð2Þ
2
Fragstats version 4.0 was used to calculate the landscape pattern in-
where Qoi and Qpi are the observed and predicted values, respectively,
dices from the three grid files of land use for the three periods. More
Qavg and Qpavg are the mean values of the observed and predicted data
than 110 landscape metrics can be calculated for each land use grid.
respectively, and n is the number of the data values.
However, if all of the possible landscape pattern metrics had been con-
The monthly time scale was used to evaluate the model performance
sidered as effective variables for regression analysis, it would have
and the evaluation indicators showed that the results of the simulation
caused workload redundancy and introduced the curse of dimensional-
were good. The R2 and Ens values for runoff reached 0.96 and 0.71 for the
ity because of the linear relations between various metrics (Hargis et al.,
calibration period, and 0.95 and 0.80 for the validation period, respec- 1998). It is, therefore, important to select the most representative met-
tively. Using the same approach, the R2 and Ens values for sediment rics and those that are highly correlated with erosion. There are four
reached 0.93 and 0.79 for the calibration period, and 0.87 and 0.75 for kinds of landscape metrics at the landscape level in soil erosion re-
the validation period, respectively. In addition, the evaluation indicator search: aggregation metrics, area and edge metrics, shape metrics, and
percent bias (PBIAS) was introduced to assess the accuracy of sediment diversity metrics. Based on the landscape patch distribution, land use,
simulation. PBIAS measures the average tendency of the simulated data and soil boundary conditions, 12 suitable metrics were selected for
to be larger or smaller than their observed counterparts (Gupta et al.,
this study (Table 2). Changes in the landscape metrics were calculated
1999). PBIAS is computed as follows:
to correspond with changes in land use and were used as the indepen-

" n # dent variables for the regression analysis.

∑ Q −Q
i¼ 1 oi pi Correlation analysis for the variation of 12 landscape metrics was
PBIAS ¼ ð3Þ
n ∙100
done and the correlation coefficients between metrics are listed in
∑ i¼1 Q oi Table 3. The preliminary analysis revealed that most of the landscape
38 S. Zhang
S. Zhang
et al.
et /al.
Science
/ Science
of the
of the
Total
Total
Environment
Environment
574574
(2017)
(2017)
34–45
34–45 38

In general, model simulation can be judged as satisfactory if PBIAS metrics were weakly correlated, and only 9% of the correlation coeffi-
± 55% for sediment and ± 25% for runoff (Moriasi et al., 2007). During cients were N 0.9. In general, the shape metrics (e.g., SHAPE_MN,
the calibration and validation periods for runoff, the PBIAS was + 40% FRAC_MN, and PAFRAC) were more independent than the other groups
39 S. Zhang
S. Zhang
et al.
et /al.
Science
/ Science
of the
of the
Total
Total
Environment
Environment
574574
(2017)
(2017)
34–45
34–45 39

Fig. 3. Calibration and validation of the runoff simulation.

of metrics. Many of the diversity metrics (e.g., SHDI, MSIEI, and SHEI)
were correlated among themselves. The diversity metrics includes two
categories of the Shannon's index and the Simpson's indices, and the
Shannon's indices were more sensitive to the rare patch type in the
landscape, while Simpson's index was more sensitive to the common
land cover types. In addition, both of the diversity metrics were highly
correlated with the contagion metric (CONTAG), the former use the pro-
portions of the cover types and the number of classes to describe the
landscape diversity, while the latter describes the diversity by the adja-
cency matrix in the landscape. However, these two kinds of metrics
expressed the different aspects of landscape heterogeneity in the
study area, so the selected metrics reasonably represent landscape var-
iations and, thus, were used as the variables for the multiple linear
regressions.
3.5. Multiple linear regression analysis
3.5.1. Principle of multiple regression analysis
Regression is used to explore the relations between different vari-
ables and a regression equation can quantitatively express the depen-
dence relation. Multiple linear regression can express the relation
between a dependent variable and more than one independent variable.

Fig. 4. Calibration and validation of the sediment simulation.

40 S. Zhang
S. Zhang
et al.
et /al.
Science
/ Science
of the
of the
Total
Total
Environment
Environment
574574
(2017)
(2017)
34–45
34–45 40

The mathematical model can be expressed as follows: respectively. The F statistic was determined as follows:

8
> F¼ Sreg =m
ΔS ¼ β0 þ β1 ΔPD þ ⋯ þ βp ΔAI þ ε ð4Þ S = n−m−1
>
> res ð Þ
< n 2
d
Sreg ¼ ∑ ΔSi −Δ S ð6Þ
> i¼1

where the variation in the erosion modulus (ΔS) is a random observa- > n 2
d
: Sres ¼ ∑ ΔSi −Δ
>
tion, β0 is a constant term, and β1, …, βp are partial regression coeffi- Si
i¼1
cients. βi(i = 1,2, …, p) relates the variation in the landscape indices to

the variation in the erosion modulus. A change of one unit in ΔPD ⋯

where Sreg is the sum of squares of the regression, S res is the sum of
ΔAI causes an average change in ΔS when the other independent vari-

ables remain unchanged. The number of landscape indices is p (here:
p ≤ 12), the observed number of samples is n, and the set of observed
values for the first i (i = 1, 2…n) column (ΔSi, ΔPDi, …, ΔAIi). It is as-
sumed there is a linear relation between (ΔPDi, …, ΔAIi) and ΔSi, the in-
dependent variable:
ΔSi ¼ ΔSbi þ εi ¼ b0 þ b1 ΔPD þ ⋯ þ bp ΔAI þ εi
where εi is the residual error, which is the difference between the de-
pendent variable ΔSi and its estimated value ðΔSbi Þ. If the residual error
is a function of the magnitude of the dependent variable and approxi-
mates a normal distribution, it can be used to assess whether the regres-
sion equation has been established. The estimated sample values of β0,
β1, …, and βp are b0, b1, …, and bp.
3.5.2. Significance test
squares of the residuals, dΔSi is the predicted value of the dependent var-
iable, ΔS is the mean value of the dependent variable, ΔSi is the observed
value of the dependent variable, m is the degrees of freedom of the re-
gression, n is the number of sample data, and n – m − 1 is the degrees
of freedom of the residuals. The results of variance analysis indicate that
the null hypothesis is rejected if F N F α (m, n-m-1), α = 0.15. It is assumed
there is a linear relation between the dependent variable and the inde-
ð5Þ pendent variable if the null hypothesis cannot be rejected, and the
established regression equation is assumed to be feasible.
(2) Hypothesis test of the partial regression coefficient
Once a statistically significant regression model is achieved, the rela-
tion between the independent variables and the dependent variables
should be tested for linearity. This approach allows elimination of vari-
ables that are not statistically significant, and results in simpler regres-
sion equations. The t-test is used to determine whether the partial
regression coefficient is significantly different from to zero. The null hy-
pothesis of the T was H0: βi = 0 (i = 1, 2, 3, 4), and the null hypothesis
would be rejected if │T│ N T α/2(n − m − 1), α = 0.15.

For regression analysis, the significance test of the equations and the ¼ b i −0 bi
T ¼ ð7Þ
variables can be used to judge the accuracy of the regression model. The Sbi Sbi
regression equation and the contribution of each independent variable

should be statistically significant to ensure that the multiple regression
equation is consistent with the characteristics of the data. So hypothesis
testing is necessary for the established equation.
(1) Hypothesis test for the regression equation
The F test was used as the hypothesis test for the regression equa-
tion. The null hypothesis is H0: β1 = β2 = …. = βp = 0, that is, no re-
gression model can be established. Note, the number of the variable in
the regression equation increases from 1 in the process of stepwise re-
gression, and in this study the number of variable (p) is 1, 2, 3, and 4,
In Eq. (7), bi is the partial regression coefficient of the independent
variable, and Sbi is the standard error of bi.
3.5.3. Selection of independent variables
The stepwise regression method in the SPSS software was used to
input the key variables into the regression equation and to select the in-
dependent variables. Stepwise regression combines the characteristics
of both the forward and backward regression methods; the difference
between this method and multiple linear regression is that the variables

Table 2
Landscape metrics used in this study.

No. Metrics Abbreviation Description Category

1 Patch Density PD Number of patches per unit area (number per 100 ha) Aggregation
2 Largest Patch Index LPI Percentage of the landscape in the largest patch (units: percent) Area and
Edge
3 Edge Density ED Total length of all edge segments per hectare for the considered landscape (units: none) Area and
Edge
4 Mean Patch Size SHAPE_MN Patch perimeter divided by the minimum perimeter possible for a maximally compact patch (units: none) Shape
5 Mean Patch Fractal FRAC_MN Based on the fractal dimension of each patch (units: none) Shape
Dimension
6 Perimeter Area Fractal PAFRAC Reflects shape complexity across a range of spatial scales (patch sizes) (units: none) Shape
Dimension
7 Aggregation Index AI Number of like adjacencies involving the corresponding class, divided by the maximum possible number of like Aggregation
adjacencies involving the corresponding land use type (units: percent)
8 Contagion CONTAG Tendency of the patch types to be aggregated (units: percent) Aggregation
41 S. Zhang
S. Zhang
et al.
et /al.
Science
/ Science
of the
of the
Total
Total
Environment
Environment
574574
(2017)
(2017)
34–45
34–45 41

9 Patch Cohesion Index COHESION Measures the physical connectedness of the corresponding patch type (units: none) Aggregation
10 Shannon's Diversity SHDI Based on information theory, indicates the patch diversity in a landscape (units: none) Diversity
Index
11 Modified Simpson's MSIEI Modified Simpson's Evenness Index is expressed such that an even distribution of area among patch types results Diversity
Evenness Index in maximum evenness (units: none)
12 Shannon's Evenness SHEI Shannon's Evenness Index is expressed such that an even distribution of area among patch types results in Diversity
Index maximum evenness (units: none)
42 S. Zhang
S. Zhang
et al.
et /al.
Science
/ Science
of the
of the
Total
Total
Environment
Environment
574574
(2017)
(2017)
34–45
34–45 42

in the equation are tested after each new variable is introduced, and the
variables that match the exclusion criteria are eliminated one by one
(Neter et al., 1996; Fritz and Berger, 2015). The following method was
used to select the independent variables:
(1) The simple linear regression models (only include one variable)
were fitted with 12 candidate independent variables (ΔPAFRAC, ΔSHEI,
…, ΔPD, respectively). Out of these, k = (k ≤ 12) linear regression
models passed the F test and the corresponding independent variable
ΔLPI resulted in the model with the smallest p value and ΔLPI was intro-
duced into the model first.
(2) Once the model using the variable ΔLPI was fitted, the linear re-
gression models of the 11 extra independent variables were fitted, that
is, the independent variables were combined as ΔLPI and ΔPAFRAC,
ΔLPI and ΔSHEI, ΔLPI and ΔPD, and so on. The p value of ΔLPI and
ΔCOHESION was the smallest, and the linear regression model met
the standard of the F test, so the ΔCOHESION variable was added to
the model.
(3) The t-test was used to examine whether the independent vari-
able (ΔLPI) still had statistical significance, and if ΔLPI did not pass the
t-test, it was removed from the model. The regression models of the ad-
ditional 10 independent variables (except ΔLPI) including ΔCOHESION
were fitted and the combination of independent variables with the low-
est p values was added to the regression model. The operation was ter-
minated if no variables were statistically significant.
(4) If ΔLPI passed the t-test, it was still meaningful and the analysis
continued to fit the model with the other 10 variables, starting with the
ΔLPI and ΔCOHESION variables. The rest of the variables were examined
for statistical significance and the variable yielding the smallest p value
was added to the regression model. This process was repeated until the
independent variables not included in the model were not statistically
significant.
The flow chart of the regression analysis is shown in Fig. 5.
4. Results
4.1. Sub-basin erosion calculation
The soil erosion of each sub-basin for the three land use periods
(1986, 1995, and 2000) was calculated with the established hydrologi-
cal and sediment yield model and the values for the three years were
converted into the erosion modulus (Fig. 6). There were 126 groups
(1986–1995 and 1995–2000, respectively) of changes in soil erosion
for the 63 sub-basins over the three periods of land use. The historical
land use of the area indicated that there were no significant changes
at the sub-basin scale of the Liusha River Watershed, and some land
use types didn't change at all over the 15 years study period (1986–
2000). If the changes in the sub-basin erosion modulus and landscape
pattern both were zero, these sub-basins would be eliminated from
the analysis, and so 20 sub-basins were selected for the regression
analysis in this study (3 samples from 1986 to 1995 and 17 samples
from 1995 to 2000).
As shown in Fig. 6, there were few substantial changes in the soil
erosion modulus between the first two periods of land use (1986,
1995), but there were significant changes between 1995 and 2000.
The most serious area of soil erosion was in the middle-lower reaches
of the Liusha River, where the erosion modulus was above 4 t/ha
(such as sub watershed No. 24, 39, 50, 55 and 58) and reached 28 t/ha
in the most serious area (sub watershed No. 53 sub-basin) over the 15
study period. The soil erosion in 1986 and 1995 were more serious
than 2000, and the average variation in the erosion modulus of the
sub-basin was 0.46 t/ha, and the maximum variation in the erosion
modulus of the sub-basins over the 15 years (1986–2000) was 3.52 t/ha.
4.2. Regression analysis of changes in the erosion modulus
Regression analysis was done in SPSS using the change in the 12 wa-
tershed landscape metrics as the independent variable and the change
in the erosion modulus in the 20 selected groups of sub-basins as the de-
pendent variable, and four regression models were obtained (Table 4).
The R2 coefficient reflects the proportion of the variance of the original
data that was explained by the regression equation. The R2 values for
the outputs of the four models were 0.601, 0.830, 0.854, and 0.883,
and these values indicate that the fourth model was the best. However,
the R2 coefficient has limitations when evaluating the model: the value
of R2 may still increase even if the introduced variable is not statistically
significant. Adjustments in the R2 value to account for the number of pa-
rameters in the model provides guidance on the selection of new vari-
ables; the adjusted R2 value decreases when the introduced variable is
not statistically significant, so it compensates for limitations in R2 and
is also an important indicator of the merits of the regression equation.
In this study, the R2 and adjusted R2 both increased as the number of in-
dependent variables increased, which was a reasonable result.
Eight variables were excluded by model 4 and their test results are
listed in Table 5. In this study, the t-test was applied to the independent
variables at a significance level of 0.15 for α. As shown in Table 5, the
probability values (Sig. T) of the eight variables added to model 4 are
greater than 0.15, The variance inflation factor (VIF, the reciprocal of tol-
erance ) of the collinearity statistics is greater than 5 for some variables,
which indicates significant collinearity between variables, so these var-
iables were excluded from the model.
The model coefficient table obtained from SPSS is listed in Table 6,
including the coefficients of the model variables, the probability of the
t-test, and the collinearity statistics (tolerance and VIF). The probabili-
ties of the t-test for the constant and the four independent variables
were much b 0.15, which indicates that the variables are statistically sig-
nificant. The VIF values of the four variables are b 5, which indicates that
there was no obvious collinearity between the four independent vari-
ables. Because the influence of the different units and dimensions of
the original variables were eliminated when the coefficients were

Table 3
Correlation analysis for 12 selected metrics in this study.

Metrics PD LPI ED SHAPE_MN FRAC_MN CONTAG COHESION SHDI MSIEI AI SHEI PAFRAC

PD 1
LPI −0.060 1
ED 0.259 −0.324 1
SHAPE_MN −0.499 −0.165 0.219 1
FRAC_MN −0.138 0.036 0.192 0.861 1
CONTAG −0.050 0.596 −0.520 −0.316 −0.319 1
COHESION −0.429 0.585 −0.124 −0.256 −0.364 0.318 1
SHDI −0.027 − 0.634 0.397 0.293 0.262 − 0.985 −0.308 1
MSIEI −0.101 − 0.699 0.424 0.330 0.221 − 0.905 −0.343 0.919 1
AI −0.249 0.319 − 0.998 −0.204 −0.162 0.476 0.106 −0.355 −0.395 1
SHEI −0.004 −0.588 0.373 0.305 0.308 − 0.986 −0.318 0.995 0.905 −0.326 1
PAFRAC − 0.617 −0.521 0.233 0.350 −0.090 −0.358 0.162 0.421 0.454 −0.244 0.354 1

Note: The bold-faced numerical values indicate that the correlation coefficient is greater than 0.6.
43 S. Zhang
S. Zhang
et al.
et /al.
Science
/ Science
of the
of the
Total
Total
Environment
Environment
574574
(2017)
(2017)
34–45
34–45 43

Fig. 5. The flow chart of the regression analysis.

standardized, standardized coefficients are used in this study to com- in this study was 0. The final multiple linear regression equation is
pare the contribution of the independent variables to the dependent
variables. The mean value of the standardized variable was 0 and the
ΔS ¼ −1:272ΔLPI þ 0:597ΔCOHESION−0:189ΔAI−0:285ΔMSIEI ð8Þ
standard deviation was 1, so the constant term of the fitted regression
44 S. Zhang
S. Zhang
et al.
et /al.
Science
/ Science
of the
of the
Total
Total
Environment
Environment
574574
(2017)
(2017)
34–45
34–45 44

Fig. 6. Soil erosion modulus and variation for the three time periods.

The regression equation contains four landscape indices, which were △COHESION, △MSIEI, and △AI indices played an important role in the
calculated by Fragstats 4.0 using the following formulas: changes in the erosion modulus. Further, the change in the erosion
modulus was positively correlated with △COHESION, but negatively
max aij
correlated with △LPI, △MSIEI, and △AI.
LPI ¼ 100 ð9Þ
A Landscapes are composed of elements and spatial components. A
2 3 convenient and popular model for conceptualizing and representing
∑
m n −1 the elements in a categorical map pattern (or patch mosaic) is known
6 i¼1 ∑ j¼1 1
Pij
COHESION 1− qffi ffiffiffiffi5 1− pffi ffiffi 100 ð10Þ as the patch-corridor-matrix model (Forman, 1995). Under this
4 m n
a 7 Z
¼ ∑
model, three major landscape elements are typically recognized, and
i¼1 ∑ j¼1 Pij ij the extent and configuration of these elements defines the pattern of
the landscape. The calculation and analysis of landscape pattern indices
gii
AI ¼ ∑i¼1
m
Pi 100 ð11Þ
max→gii are based on this framework. Patches are the nonlinear surface region
with a homogeneous internal spatial structure and which is different
m
− ln ∑ P 2i
i ¼1
from the surrounding environment (matrix). Corridors are linear and
MSIEI ¼ ð12Þ banded landscape elements, as a consequence of their form and context,
lnm
structural corridors may function as channels, dispersal conduits, or bar-
where aij is the area (m2) of patch ij, A is the total landscape area (m2), Pi⁎j riers, which can be considered as the important bridge and link for dif-
is the perimeter of patch ij in terms of number of cell surfaces, ai⁎j is the ferent patches influencing the connectivity between them to a great
area of patch ij in terms of number of cells, Z is the total number of cells extent. A landscape is composed typically of several types of landscape
in the landscape, gii is the number of like adjacencies between pixels of elements (usually patches). Of course, the matrix is the most extensive
patch type (class) i based on the single-count method, max → gii is the and most connected landscape element type, and plays the dominant
maximum number of like adjacencies between pixels of patch type i, role in the functioning of the landscape. For example, at a particular
Pi is the proportion of the landscape comprised of patch type (class) i, scale, mature forest may be the matrix with disturbance patches em-
m is the number of patch types (classes) present in the landscape, ex- bedded within; whereas, at a coarser scale, agricultural land may be
cluding the landscape border if present. the matrix with mature forest patches embedded within. Next, the
The expected cumulative frequency and the observed cumulative four statistically significant landscape pattern indices are explained
frequency (P-P diagram) of the standardized residuals and the normal within the landscape ecology order.
probability distribution histogram were used to assess whether the re-
siduals εi were normally distributed. The P-P diagram of the standard-
ized residuals showed that there was a relation between the Table 4
cumulative probability of the expected and observed values (Fig. 7). Four regression models for change in erosion modulus determined from SPSS.
When combined with the normalized residual histogram, the residual

Model and variable R R2 Adjustment Standard error of the

error could be explained by the normal distribution (Fig. 8), and the of R2 estimate
multiple linear regression model was found to be feasible.
1st ΔLPI 0.775 0.601 0.579 0.546
45 S. Zhang
S. Zhang
et al.
et /al.
Science
/ Science
of the
of the
Total
Total
Environment
Environment
574574
(2017)
(2017)
34–45
34–45 45

2nd ΔLPI, ΔCOHESION 0.911 0.830 0.810 0.367

5. Discussion 3rd ΔLPI, ΔCOHESION, Δ 0.924 0.854 0.826 0.351
MSIEI
4th ΔLPI, ΔCOHESION, 0.940 0.883 0.852 0.324
The multiple linear regression equation showed that out of the huge
ΔMSIEI, ΔAI

number of landscape pattern indices, the variation in the △LPI,

46 S. Zhang
S. Zhang
et al.
et /al.
Science
/ Science
of the
of the
Total
Total
Environment
Environment
574574
(2017)
(2017)
34–45
34–45 46

Table 5
T test and collinearity analysis for variables excluded from model 4.

Model Variable Beta In T Sig. T Partial Correlation Collinearity Statistics

Tolerance VIF

4 ΔPD 0.009 0.078 0.939 0.021 0.638 1.568

ΔED −2.072 −1.190 0.254 −0.303 0.003 398.963
ΔSHAPE_MN −0.009 0.091 0.929 0.024 0.824 1.214
ΔFRAC_MN −0.043 −0.363 0.722 −0.096 0.583 1.715
ΔCONTAG −0.075 −0.327 0.748 −0.087 0.156 6.409
ΔSHDI 0.126 0.553 0.589 0.146 0.157 6.364
ΔSHEI 0.073 0.333 0.744 0.089 0.174 6.364
ΔPAFRAC 0.043 0.298 0.770 0.079 0.405 2.472

The largest patch area in the landscape (LPI) has values that range
from 0 to 100, indicating the proportion of the total landscape area.
The regression equation shows that there is a strong negative correla-
tion between △S and △LPI, and sub watershed soil erosion was the
most sensitive index to the LPI index, and the amount of the soil erosion
decreased with an increase of the index. The value of this metric reflects
the dominant species in the landscape and, to a certain degree, indicates
the direction and strength of the effects of human activities. Fig. 2 shows
that the dominant land uses in the watershed are Forest-Mixed, Range-
Brush, and Range-Grasses. Each of these land uses is prone to lower sed-
iment yields. Thus, the largest patch type in most sub watersheds
should be one of these types of land use. Therefore, an increase in LPI in-
dicates an increase in Forest-Mixed, Range-Brush, or Range-Grasses
land use which results in a decrease in sediment yield.
The Patch Cohesion Index (COHESION) can be used to measure the
landscape physical connectivity between the corresponding patch
types. As previously mentioned, corridors can be considered as the
bridge or link for various patches, and different corridor types will affect
the connectivity between patches. Corridor types are divided into natu-
ral corridors (e.g., river, dam and shelterbelt) and artificial corridors
(e.g., green way, road network especially in urban landscapes), which
make the patch boundary more obvious. Corridors have the functions
of filtration; obstruction; interception; and barrier for energy, material
and biological flow in the landscape, which will have a positive effect
for the work of soil and water conservation in the basin. The multiple
linear regression equations showed that there was a positive correlation
between △S and △COHESION, meaning that the amount of soil erosion
increased as COHESION increased. This may reflect enhanced connectiv-
ity between patches that promotes material transport as a result of ag-
ricultural measures for soil and water conservation, like furrow and
ridge tillage and contour cultivation, or engineering measures, such as
terraced fields and drainage ditches on slopes. These measures mainly
induce small-scale effects; for example, slope connectivity may be
blocked to achieve the overall purpose of soil and water conservation
and prevent waterlogging.
The modified Simpson's evenness index (MSIEI) is a landscape di-
versity index that reflects the number of patches in the landscape. In
this study, this index was negatively correlated with △S, meaning that
decreases in the evenness of the patch type resulted in an increase in
the amount of soil erosion. This study's results showed that because of
its rich diversity, the landscape was conducive to soil and water
conservation, which reminds us that water and soil resources need to
be managed to suit local conditions. The regression analysis also indicat-
ed that increases in MSIEI over time resulted in decreases in erosion
modulus over time. As the proportion of land in the dominant land
use types—Forest-Mixed, Range-Brush, and Range-Grasses—increases
MSIEI also would increase because of higher values of Pi and possibly re-
duction in m. Thus, a decrease in sediment yield would be expected for
the study watershed as MSIEI increases. The AI indicates the degree of
fragmentation of patch type within a landscape; when AI approaches
its maximum value, the landscape patch types are aggregated tightly to-
gether and the degree of aggregation is at a maximum. Variation in the
erosion modulus (△S) was negatively correlated with △AI, which indi-
cates that increases in the degree of landscape patch aggregation results
in reduced watershed erosion. The regression analysis also indicated
that increases in AI over time resulted in decreases in erosion modulus
over time. Increases in AI would result from increases in the amount
of land in the Forest-Mixed, Range-Brush, or Range-Grasses land use be-
cause the adjacencies between pixels of these patch types also would in-
crease. Hence, increases in the dominant land use types, which have
lower sediment yields, would increase the value of AI and decrease
the erosion modulus. On the one hand, afforestation, forage grass, and
planning for reafforestation on hillsides have positive implications for
water and soil conservation, wind-breaking, and sand-fixing, further-
more, they increase vegetation cover and strengthen the degree of ag-
gregation by changing the micro-terrain. In arable areas, intercropping
and mixing of crops could effectively enhance the degree of internal ag-
gregation of patches and prevent soil loss at the same time, and could
also reap economic benefits.
Because some research indicates that the soil erosion modulus could
be obtained directly from landscape metrics, the regression relation be-
tween the soil erosion modulus and landscape metrics of the sub-basins
was also examined for the three periods of land use. The same regres-
sion method was applied, except that the erosion modulus and land-
scape metrics were regarded as the dependent and independent

Table 6
Model coefficients and t-test results for the variables.

Model Variable Standardized T Sig. T Collinearity

coefficients Statistics

Value Standard Tolerance VIF

error

4 Constant 0 0.081 −2.484 0.02500

ΔLPI −1.272 0.049 −8.710 0.00000 0.366 2.732
ΔCOHESION 0.597 1.210 5.352 0.00000 0.628 1.592
ΔAI −0.189 0.706 −1.937 0.07200 0.464 2.156
ΔMSIEI −0.285 4.768 −2.199 0.04400 0.817 1.224
Fig. 7. Normalized residual P-P diagram.
47 S. Zhang
S. Zhang
et al.
et /al.
Science
/ Science
of the
of the
Total
Fig. 8. Normal probability distribution histogram.
variables, respectively, instead of their variations. Three regression
models were obtained with determinate coefficients of 0.19, 0.24, and
0.33, respectively. These models were not sufficiently accurate to meet
the requirements of natural science research, which indicates that the
landscape metrics are only one of the factors which influence the soil
erosion; it is not feasible to obtain the soil erosion amount from land-
scape metrics alone for the Liusha River Watershed.
6. Conclusions
From the perspective of landscape ecology, land cover patterns with-
in a watershed have a considerable influence on hydrological connectiv-
ity, and the generation of runoff and sediment transport. The study
indicates that the relation between soil erosion and landscape patterns
is important for watershed management. Variations in the landscape
pattern metrics was used to describe the impacts of landscape pattern
on soil erosion and to explain the ecological significance of landscape
metrics, and the following conclusions were obtained.
Many factors influence watershed soil erosion and sediment yield,
but landscape pattern indices cannot be used on their own to estimate
the amount of soil erosion. The relationship between soil erosion and
landscape patterns in terms of the variation in landscape metrics was
evaluated and the regression equation between variations in the soil
erosion modulus and landscape metrics was obtained. The linear regres-
sion coefficient of each index indicated that changes in △LPI, followed
by △COHESION, △MSIEI, and △AI, had most influence on the erosion
modulus.
The importance of each landscape index in the regression equation
and the positive or negative correlations indicate that erosion in water-
sheds can be reduced by strengthening the control function of the dom-
inant patch, thoroughly improving the evenness of the landscape patch
types, enriching the landscape types, reducing the physical connectivity
between patches, and enhancing the degree of aggregation in landscape
patches. Along with robust planning of land use patterns, these mea-
sures will contribute to improved and optimized soil and water conser-
vation in river basins.
Acknowledgments
This study was supported by the National Natural Science Founda-
tion of China (51379076) and the 13th Five-Year National Key Research
and Development Program of China (2016YFC0401401).
Total
Environment
Environment
574574
(2017)
(2017)
34–45
34–45 47
References
Alatorre, L.C., Beguería, S., Lana-Renault, N., Navas, A., García-Ruiz, J.M., 2012. Soil erosion
and sediment delivery in a mountain catchment under scenarios of land use change
using a spatially distributed numerical model. Hydrol. Earth Syst. Sci. 16, 1321–1334.
Arnold, J.G., Srinivasan, R., Muttiah, R.S., Williams, J.R., 1998. Large area hydrologic model-
ing and assessment part I: model development. J. Am. Water Resour. Assoc. 34,
73–89.
Ciampalini, R., Billi, P., Ferrari, G., 2012. Soil erosion induced by land use changes as deter-
mined by plough marks and field evidence in the Aksum area (Ethiopia). Agric.
Ecosyst. Environ. 146, 197–208.
Durán-Zuazoa, V.H., Francia-Martínez, J.R., García-Tejero, I., Tavira, S.C., 2013. Implications
of land-cover types for soil erosion on semiarid mountain slopes: towards sustainable
land use in problematic landscapes. Acta Ecol. Sin. 33, 272–281.
Feng, X., Wang, Y., Chen, L., Fu, B., Bai, G., 2010. Modeling soil erosion and its response to
land-use change in hilly catchments of the Chinese loess plateau. Geomorphology
118, 239–248.
Forman, R.T.T., 1995. Land Mosaics: The Ecology of Landscape and Regions. Cambridge
University Press, New York.
Fritz, M., Berger, P.D., 2015. Chapter 10–can you relate in multiple ways? Multiple linear
regression and stepwise regression. Improving the User Experience Through Practical
Data Analytics, pp. 239–269.
Gupta, H.V., Sorooshian, S., Yapo, P.O., 1999. Status of automatic calibration for hydrologic
models: comparison with multilevel expert calibration. J. Hydrol. Eng. 4, 135–143.
Hao, F.H., Cheng, L.Q., Liu, C.M., Dai, D., 2004. Impact of land use change on runoff and sed-
iment yield. J. Soil Water Conserv. 18, 5–8.
Hargis, C.D., Bissonette, J.A., David, J.L., 1998. The behavior of landscape metrics commonly
used in the study of habitat fragmentation. Landsc. Ecol. 13, 167–186.
Huang, J., Li, Q., Pontius, R.G., Klemas, V., Hong, H., 2013. Detecting the dynamic linkage
between landscape characteristics and water quality in a subtropical coastal water-
shed, southeast China. Environ. Manag. 51, 32–44.
Liu, Y., Lu, Y.H., 2011. Implication and limitation of landscape metrics in delineating rela-
tionship between landscape pattern and soil erosion. Acta Ecol. Sin. 31, 267–275.
Liu, Q.J., Yu, X.X., 2010. Vertical landscape pattern on soil erosion intensity in the rocky
area of northern China: a case study in the Yimeng mountainous area, Shandong.
Geogr. Res. 8, 1471–1483.
McGarigal, K., Cushman, S.A., Ene, E., 2012. FRAGSTATS v4: Spatial Pattern Analysis Pro-
gram for Categorical Maps. Computer Software Program Produced by the Authors
at the University of Massachusetts, Amherst. bhttp://www.umass.edu/landeco/
research/fragstats/fragstats.htmlN.
Moriasi, D.N., Arnold, J.G., Liew, M.W.V., Bingner, R.L., Harmel, R.D., Veith, T.L., 2007. Model
evaluation guidelines for systematic quantification of accuracy in watershed simula-
tions. Trans. ASABE 50, 885–900.
Munro, R.N., Deckers, J., Haile, M., Grove, A.T., Poesen, J., Nyssen, J., 2008. Soil landscapes,
land cover change and erosion features of the Central Plateau region of Tigrai, Ethio-
pia: photo-monitoring with an interval of 30 years. Catena 75, 55–64.
Nash, J., Sutcliffe, J., 1970. River flow forecasting through conceptual models. Part I –A dis-
cussion of principles. J. Hydrol. 10, 282–290.
Neitsch, S.L., Arnold, J.G., Kiniry, J.R., Williams, J.R., 2011. SWAT User Manual, Version
2009. Texas Water Resources Institute Technical Report. Texas A and M University,
Texas, USA.
Neter, J., Wasserman, M., Kutner, M.H., 1996. Applied Linear Statistical Models. Irwin, VOL,
Chicago, p. 4.
Ni, J.R., Li, X., Borthwick, A.G.L., 2008. Soil erosion assessment based on minimum poly-
gons in the Yellow River basin, China. Geomorphology 93, 233–252.
Nunes, A.N., Almeida, A.C.D., Coelho, C.O.A., 2011. Impact of land use and cover type on
runoff and soil erosion in a marginal area of Portugal. Appl. Geogr. 2, 687–699.
Ochoa, P.A., Fries, A., Mejía, D., Burneo, J.I., Ruíz-Sinoga, J.D., Cerdà, A., 2016. Effects of cli-
mate, land cover and topography on soil erosion risk in a semiarid basin of the Andes.
Catena 140, 31–42.
Ouyang, W., Skidmore, A.K., Hao, F.H., Wang, T.J., 2009. Soil erosion dynamics response to
landscape pattern. Sci. Total Environ. 408, 1358–1366.
Ouyang, W., Hao, F.H., Skidmor, A.K., Toxopeus, A.G., 2010. Soil erosion and sediment yield
and their relationships with vegetation cover in upper stream of the Yellow River. Sci.
Total Environ. 409, 396–403.
Qi, S.S., Hao, F.H., Ouyang, W., Cheng, H.G., 2012. Characterizing landscape and soil erosion
dynamics under pipeline interventions in southwest China. Procedia Environ. Sci. 13,
1863–1871.
Saedi, T., Shorafa, M., Gorji, M., Moghadam, B.K., 2016. Indirect and direct effects of soil
properties on soil splash erosion rate in calcareous soils of the central Zagross, Iran:
a laboratory study. Geoderma 271, 1–9.
Shi, Z.H., Ai, L., Li, X., Huang, X.D., Wu, G.L., Liao, W., 2013. Partial least-squares regression
for linking land-cover patterns to soil erosion and sediment yield in watersheds.
J. Hydrol. 498, 165–176.
Silva, A.M.D., Huang, C.H., Francesconi, W., Saintil, T., Villegas, J., 2015. Using landscape
metrics to analyze micro-scale soil erosion processes. Ecol. Indic. 56, 184–193.
Wang, K., Wang, H.J., Shi, X.Z., Weindorf, D.C., Yu, D.S., Liang, Y., 2009a. Landscape analysis
of dynamic soil erosion in subtropical China: a case study in Xingguo county, Jiangxi
province. Soil Tillage Res. 105, 313–321.
Wang, W., Zhang, S., Li, Y., Deng, R., Yan, Y., 2009b. Research on the Relationship Between
Soil Erosion and Landscape Pattern in the Wuyuer River Basin Based on GIS. Interna-
tional Conference on Photonics and Image in Agriculture Engineering (PIAGENG
2009). Society for Optics and Photonics, International.
Wei, J.B., Xiao, D.N., Li, X.Z., Ren, C.B., Zhang, C.S., 2006. Relationship between landscape
pattern and soil erosion of an agricultural watershed in the mollisols region of north-
eastern China. Acta Ecol. Sin. 26, 2608–2615.
48 S. Zhang
S. Zhang
et al.
et /al.
Science
/ Science
of the
of the
Total
Total
Environment
Environment
574574
(2017)
(2017)
34–45
34–45 48

Winchell, M., Srinivasan, R., Di Luzio, M., Arnold, J., 2013. ArcSWAT Interface for
SWAT2012: User's Guide. Blackland Research and Extension Center. Texas Agrilife
Research, Grassland, Soil and Water Research Laboratory, USDA Agricultural Research
Service, Texas.
Xiang, J., Tang, Y., 2006. The influence of the adjustment of agricultural industrial struc-
ture on environment and economy in the dry valley region- select Hanyuan area as
an example. Sichuan Environ. 25, 31–36.
Xiao, H., Ji, W., 2007. Relating landscape characteristics to non-point source pollution in
mine waste-located watersheds using geospatial techniques. J. Environ. Manag. 82,
111–119.
Yang, Y,J., 2012. Analysis of hydrological characteristics in Liusha Basin. Sichuan Water
Conservancy. 2, 27–29.
Zhang, S.H., Liu, Y., Wang, T.W., 2014. How land use change contributes to reducing soil
erosion in the Jialing River Basin, China. Agric. Water Manag. 133, 65–73.