See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/284437824

New insight into the correlations between land

use and water quality in a coastal watershed of
China: Does point source pollution weaken it?

ARTICLE in SCIENCE OF THE TOTAL ENVIRONMENT · NOVEMBER 2015

Impact Factor: 4.1 · DOI: 10.1016/j.scitotenv.2015.11.063

CITATION READS

1 137

4 AUTHORS:

Pei Zhou Jinliang Huang

Xiamen University Xiamen University
3 PUBLICATIONS 7 CITATIONS 54 PUBLICATIONS 332 CITATIONS

SEE PROFILE SEE PROFILE

Robert Gilmore Pontius Huasheng Hong

Clark University Xiamen University
93 PUBLICATIONS 3,692 CITATIONS 217 PUBLICATIONS 3,871 CITATIONS

SEE PROFILE SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate, Available from: Jinliang Huang
letting you access and read them immediately. Retrieved on: 27 January 2016
 Science of the Total Environment 543 (2016) 591–600

Contents lists available at ScienceDirect

Science of the Total Environment

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

New insight into the correlations between land use and water quality in a
coastal watershed of China: Does point source pollution weaken it?
Pei Zhou a,b, Jinliang Huang a,b,⁎, Robert Gilmore Pontius Jr c, Huasheng Hong a,b
a
Coastal and Ocean Management Institute, Xiamen University, Xiamen 361102, China
b
Fujian Provincial Key Laboratory of Coastal Ecology and Environmental Studies, Xiamen University, Xiamen 361102, China
c
Graduate School of Geography, Clark University, Worcester, MA 01610, USA

H I G H L I G H T S G R A P H I C A L A B S T R A C T

• We develop a self-organizing map-

based approach.
• Different patterns of water quality are
delineated and visualized across water-
shed.
• Spatial variation of the land use-water
quality correlation is analyzed.
• Point source pollution can weaken or
hide the land use-water quality correla-
tion.

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

Article history: Uncovering the associations between land use and river water quality is useful for managing land-based pollu-
Received 1 July 2015 tion in the catchment-coast continuum. However, it is not clear how land use affects water quality in the context
Received in revised form 21 October 2015 of simultaneous point source (PS) pollution. In this study, we develop a self-organizing map (SOM)-based ap-
Accepted 12 November 2015
proach to explore the relationship between land use and water quality in the Minjiang River Watershed, South-
Available online xxxx
east China. Water samples from 139 headwater sub-watersheds were associated with six land use categories,
Editor: D. Barcelo namely, Woodland, Agriculture, Orchard, Built-up, Unused land and Water. Sampling sites are delineated into
six clusters based on six water quality parameters: ammonium-N, nitrate-N, total nitrogen, soluble reactive phos-
Keywords: phate, total phosphate and potassium permanganate index. Local relationships between land use and water qual-
Land use ity among four clusters that have sufficient sample sizes are further identified. There is no significant land use-
Water quality water quality correlation in one of the four clusters (including 37 sub-watersheds). And the greater the PS pollu-
Point source pollution tion is, the less significant the land use-water quality correlations are in clusters. The results demonstrate how PS
Self-organizing map pollution weakens the land use-water quality correlation. Our method can help to determine whether non-point
Spatial variation
source or PS pollution exerts greater influence on the quality of the water coming from watershed.
© 2015 Published by@ Elsevier B.V.

⁎ Corresponding author.
E-mail address: jlhuang@xmu.edu.cn (J. Huang).

http://dx.doi.org/10.1016/j.scitotenv.2015.11.063
0048-9697/© 2015 Published by@ Elsevier B.V.
592 P. Zhou et al. / Science of the Total Environment 543 (2016) 591–600
1. Introduction
Land use within a watershed has important effects on water quality
of rivers, lakes, estuarine and coastal waters (Huang et al., 2013a, Bu
et al., 2014, Hur et al., 2014). Watershed land use affects water quality
through non-point source (NPS) pollutants, which are major contribu-
tors of contaminants to the catchment-coast continuum (Swaney
et al., 2012). As the Secretary-General of the United Nations reported,
land-based sources of pollution contribute to more than 80% of the pol-
lutants entering the seas (UN General Assembly, 2004). Water Quality
and Sustainable Practices on Land is included in the nine priority objec-
tives for a national ocean policy articulated by the US National Ocean
Council, in order to enhance water quality by promoting and
implementing sustainable practices on land (US National Ocean
Council, 2010). Therefore, exploring the linkage between land use and
surface water quality, particularly in coastal watersheds, is critical for
developing watershed management practices (Uuemaa et al., 2005,
Xiao and Ji, 2007, Wan et al., 2014) and controlling land based pollution
in coastal bays (Huang et al., 2013a).
Many statistical methods have been widely employed to reveal the
relationships between land use and water quality (Céréghino and
Park, 2009, Bierman et al., 2011, Huang et al., 2013b), such as correlation
analysis (Lee et al., 2009, Bu et al., 2014, Li et al., 2015), multiple regres-
sion (Huang et al., 2013b, Bu et al., 2014, Park et al., 2014), and redun-
dancy analysis (Maarten et al., 2008, Shen et al., 2015). Researchers
usually apply these methods to analyze the overall relationship be-
tween land use and water quality for the entire study area (Huang
et al., 2013b, Bu et al., 2014). Nevertheless, the pollution sources
might vary across the study area, especially among watersheds that
are dominated by various uses (Baker, 2003), therefore the global ap-
proach to use the abovementioned statistics might miss the spatial var-
iation of the relationships and hide some local relationships between
land use and water quality (Kang et al., 2010). Recently, geographically
weighted regression (GWR) has been used to capture spatial variations
by incorporating spatial coordinates into the regression model (Tu and
Xia, 2008, Tu, 2011). GWR frequently uses only one land use indicator
as the independent variable to analyze the indicator's association with
each water quality dependent variable. When dealing with multivariate
data, a large number of GWR models are performed to account for the
large number of land use categories and water quality variables. There-
fore, GWR is of limited value to make land use management strategies
because of the excessive detailed spatial variation in the results. How
should we process multivariate environmental data to make appropri-
ate management strategies while considering the spatial variation? A
plausible way is to cluster sites with similar characteristics and then to
analyze the land use-water relationships in each cluster (Martin et al.,
2011). By use of clusters, we can reveal land use-water quality patterns
that global statistics may hide (Löhr et al., 2010, Choi et al., 2014) and
thus simplify land use management. What's more, researchers seldom
consider the influence from point source (PS) pollution on land use-
water quality relationships. PS pollution is not proportional with the
land use area. A small area with intensive industrial or domestic dis-
charge can generate severe PS pollution. Therefore, PS pollution might
cause huge uncertainty concerning the linkage between land use and
water quality.
A self-organizing map (SOM) is a branch algorithm that uses an arti-
ficial neural network. A SOM is a versatile tool for the classification and
association of samples and their variables (Vesanto et al., 2000, Compin
and Céréghino, 2007, Bierman et al., 2011, Li et al., 2012, Kohonen, 2001,
Kohonen, 2013). SOMs have become popular in environmental studies
because SOMs can deal with nonlinearities (Kalteh et al., 2008, Jeong
et al., 2010, Kohonen, 2013). Kalteh et al. (2008) and Céréghino and
Park (2009) reviewed applications of SOMs to environmental science,
especially regarding water resources classification. Chon (2011)
reviewed the applications of the SOM techniques in ecological sciences.
Our previous studies used multiple linear regression and GWR to de-
tect the global and local variations in relationships between land use
and water quality in a coastal watershed (Huang et al., 2013b, Huang
et al., 2015). However, we need more attempts to know how land use
affects water quality in the presence of PS pollution. In this study, we de-
velop a framework that integrates SOM, cluster analysis, backward step-
wise regression and correlation analysis, to explore the river water
quality variation and its relationship with land use at 139 sampling
sites across the Minjiang River Watershed. We hypothesize that severe
PS pollution might weaken the correlation between land use and water
quality. We suspect that exclusion of the water quality data that are in-
fluenced by PS pollution will result in a stronger correlation between
land use and water quality. Specifically, we aim (1) to delineate the pat-
tern of water quality variation among 139 headwater sub-watersheds,
(2) to explain the roles of NPS pollution from land use and PS pollution
in controlling each water quality pattern, and (3) to reveal the spatial
variation of the land use-water quality correlation and the influence of
PS pollution on the correlation.
2. Materials and methods
2.1. Study area
The Minjiang River Watershed covers 60,992 km2 and is the largest
watershed in Fujian Province, China (116°23′–119°43′ E, 25°23′–28°19′
N) (Fig. 1). The watershed is in a subtropical zone with a monsoon cli-
mate: the annual average temperature is 16–20 °C and annual average
precipitation is 1617 mm, of which 70% occurs between April and Sep-
tember (Chen et al., 2011).
The study area mainly consists of three cities: Sanming, Nanping and
Fuzhou (Fig. 1). More than 12 million residents use the Minjiang River
as their source of water for residential, industrial and agricultural activ-
ities. The watershed's gross domestic product accounts for more than
thirty percent of Fujian Province's economic output. Population in the
watershed accounts for about thirty five percent of Fujian Province.
Sanming City is in the western part of the study area, which is the
most important industrial base for raw material in Fujian Province. In-
dustry in Sanming City mainly is in Yong'an and Sanming and Shaxian
Counties. Sanming City is also the main agricultural production base
for Fujian Province. Jian'ou County is famous for its lotus, pear and
rice. Ninghua County is famous for its agricultural production, especially
rice. Taining County is developing ecotourism recently. Nanping City lo-
cated in the central part of the study area, which is one of the main for-
est areas in the south of China. It is the main bread basket and provides
about one third grain for Fujian Province. Recently poultry production is
developing in Nanping City. Fuzhou City is the capital of Fujian Province,
and is located the downstream end of the study area. Pollution from res-
idential land and poultry production accounts for the major contami-
nants in Fuzhou City.
According to Fujian Statistical Yearbook (Fujian Statistical Bureau,
2014), industrial water consumption of Fuzhou, Sanming and Nanping
Cities in 2013 are 523, 1627, 224 million tons (Fig. 2A). Treated indus-
trial waste water accounts for 34%, 36% and 35% of the water consump-
tion. The chemical oxygen demand in treated waste water decreases by
94%, 86% and 85% from the untreated industrial waste water. The am-
monia removal rate is 84%, 64% and 90% in industrial waste water treat-
ment facilities. In terms of the sewage, the volume of treated
wastewater in Fuzhou, Sanming, Yong'an, Nanping, Shaowu, Wuyishan,
Jian'ou and Jianyang Counties are 214.7, 39.4, 24.4, 17.5, 6.3, 5.2, 4.6 and
4.9 million m3. Treated sewage accounts for 86%, 85%, 86%, 89%, 88%,
86%, 67% and 86% of the total sewage discharge (Fig. 2B). However, ac-
cording to Chinese Environment quality standard for surface water (GB
3838-2002) (Ministry of Environmental Protection of the People's
Republic of China, 2002), the treated waste water still does not meet
standards for drinking water. Furthermore, the concentrations of some
pollutants discharged from waste water treatment facilities exceed the
 P. Zhou et al. / Science of the Total Environment 543 (2016) 591–600
Fig. 1. The study area showing (A) the water sampling sites, the river system, and the cities, and (B) land use and counties.
highest concentration limit of the standard (Fujian Provincial
Department of Environmental Protection, 2015). Therefore, the PS pol-
lution is still severe in the study area due to the insufficient treated sew-
age percentages, small amount of industrial and domestic sewage
treatment, and the high concentration of pollutants from waste water
treatment facilities.
2.2. Methods and data collection
Fig. 3 shows the flowchart of the SOM-based method. Firstly, we
gathered water quality data on field survey and obtained land use
data from satellite images classification. Secondly, the sampling sites
were clustered and visualized using SOM and hierarchical cluster anal-
ysis. We evaluated the water quality pattern in each cluster. Thirdly,
we performed backward stepwise regressions by using the whole land
use and water quality data and excluding the data in each cluster indi-
vidually. We analyzed the influences of PS and NPS pollution on the
water quality in each cluster by combining the regression results with
information concerning land use, industrial discharge, and domestic
discharge. Fourthly, we performed correlation analysis between land
Fig. 2. (A) The industrial waste water treatment in main cities and (B) the sewage treat-
ment in main counties.
593
use and water quality in eligible clusters. We verified the influence of
PS pollution on the land use-water quality correlation, based on the re-
sults concerning PS pollution from the backward stepwise regression
analysis and the correlation analysis. Finally, we made some suggestions
for water management in the study area.
2.2.1. Sampling sites and water quality parameters
We carried out an overall investigation concerning the land use and
water quality of the study area from 27th July to 2nd August, 2014. We
collected surface water samples from 139 sampling sites, which are at
the outlet of tributaries (Fig. 1). The samples were kept at 4 °C and
transported to the laboratory for analysis. Six water quality parameters
were analyzed following standard methods (State Environmental
Protection Administration of China, 2002). These parameters were am-
−
monium N (NH+ 4 -N), nitrate N (NO3 -N), total nitrogen (TN), soluble re-
active phosphate (SRP), total phosphate (TP) and the potassium
permanganate index (CODMn). Chinese Environmental quality standards
for surface water (GB 3838-2002) (Ministry of Environmental Protection
of the People's Republic of China, 2002) regulates five classes (from I to
V) for each specific water quality parameter. The SRP and NO− 3 -N in this
study are not included in the standard. Water pollution increases from
class I to V. The water that is classified as IV or V is not suitable for serv-
ing as drinking water source. In this study, we define water quality pa-
rameter as “substandard” if the parameter's concentration beyond the
upper level of class III.
The daily mean precipitation in the study area during our sampling
period ranged from 0 to 0.39 mm. The daily mean precipitation for
7 days before sampling ranged from 0.33 to 1.60 mm. We took samples
during wet season because the water quality in wet season represents
the joint influence of NPS and PS pollution. PS pollution tends to be in-
dependent of flow while NPS pollution is principally dependent of
flow (Bowes et al., 2008). Therefore, water quality in the dry season rep-
resents mainly PS pollutant (Mouri et al., 2011; Li et al., 2013) and is not
the focus of our research.
2.2.2. Land use data
To obtain a complete and cloud free coverage of the study area, we
chose six Landsat 8 images in 2014. These Operational Land Imager im-
ages were acquired from Institute of Remote Sensing and Digital Earth,
Chinese Academy of Sciences (http://ids.ceode.ac.cn/). Table 1 shows
our classification scheme. The classification method is an unsupervised
classification method using the Iterative Self-Organizing Data Analysis
clustering algorithm. We selected 256 pixels in a stratified random sam-
ple and checked each pixel by visual interpretation of the Landsat
594 P. Zhou et al. / Science of the Total Environment 543 (2016) 591–600
Fig. 3. The flowchart of the SOM-based method.
images and fine resolution images from Google Earth. The overall agree-
ment between the classification and our visual interpretation for the
256 pixels is 87%. Land use in each sub-watershed was extracted by clip-
ping the land use map of the study area with the boundary of each sub-
watershed.
2.2.3. Self-organizing map (SOM)
We used a SOM to characterize the spatial distribution patterns of
water quality at the 139 sampling sites. SOM is an unsupervised learn-
ing procedure to visualize and to interpret multi-dimensional data in a
two-dimensional lattice (Kohonen, 2001). A typical SOM consists of an
input layer and an output layer. All of the neurons in the input layer
are connected to all of the neurons in the output layer through a weight
vector. The neurons in the output layer are connected to adjacent neu-
rons by a neighborhood relation dictating the topography of the map.
Through a learning process, SOMs classify input data into categories.
The optimum arrangement for the output layer is a hexagonal lattice,
because this configuration does not favor the horizontal and vertical di-
rections as much as a rectangular array (Kohonen, 2001). The number of
Table 1
The land use description in this study.
Land use Description
category
Agriculture Land used for cultivation, including newly cultivated land, fallow
land, swidden land, rotation plow land, land mainly used for planting
and beach cultivated more than three years
Built-up Residential area, including the surrounded enterprise area,
entertainment area, all kinds of road and airport. Industrial area
Water Inland water area and water conservancy facilities
Orchard Areas for planting perennial woody plants and perennial herb which
were used for collecting their fruit, leaves and rhizome. Cover degree
N50%, including Fruit nursery
Woodland Including arbor, bamboo, shrub, Mangrove and pastureland.
Residential land use for greening, plants used along railways,
highways and rivers are not included
Unused Barren land, alkaline land, sand, waste land
land
output neurons in an SOM is important to detect the deviation of the
data. We combined two methods to select the number of neurons. The
pffiffiffi
first method is that the number of map neuron is close to 5 n
(Vesanto et al., 2000), where n is the number of training samples, mean-
ing the 139 sample sites. The second method is to select the number of
neurons based on minimum values for Quantity Error (QE) and Topol-
ogy Error (TE) (Kohonen, 2001).QE is the average distance between
each input vector and its best match neuron, while TE is a measure for
topological preservation of all input vectors.
In order to present correlations between water quality and land use
in an intuitive way, we set the mask value to zero for all the land use cat-
egories during the learning process. This step can remove the effect of
land use on SOM organization (Vesanto et al., 2000, Compin and
Céréghino, 2007). The values for each land use category were then visu-
alized on the SOM previously trained with the six water quality
parameters.
2.2.4. Hierarchical cluster analysis
We applied Ward linkage method using Euclidean distance to
weight the vectors of SOM. This method is a hierarchical cluster analysis
that defines the cluster boundaries in the SOM neurons. It provides a
better understanding about the group formations and reduces noise
without loss of information (Vesanto and Alhoniemi, 2000). The num-
ber of clusters is formed according to the Davies–Bouldin index. The
lower value the Davies–Bouldin index is, the better those clusters are
differentiated. The Kruskal–Wallis test is then conducted to assess the
differences between water quality parameters for clusters (Olkowska
et al., 2014).
2.2.5. Backward stepwise regression analysis
We conducted backward stepwise regressions to detect the influ-
ence of PS pollution on water quality in each cluster. PS pollution is
not proportional to the land use area. A small area with intensive indus-
trial and domestic discharge can generate severe PS pollution. The water
quality data that are influenced by PS pollution are outliers when
performing the land use-water quality regressions. Excluding the data
 P. Zhou et al. / Science of the Total Environment 543 (2016) 591–600
that are influenced by PS pollution will generate a better land use-water
quality regression model fit, as indicated by a larger R-squared. Firstly,
we log transformed the water quality data because the nonlinearity of
the land use and water quality relationship (Huang et al., 2013b). Sec-
ondly, we performed the regression with the log transformed water
quality parameters as dependent variables and the percentage of land
use categories as the independent variables for all the data. Thirdly,
we excluded the data in each cluster individually and performed the
backward stepwise regressions. Finally we compared the fits of the
cluster-excluded regressions with the all data-based regression individ-
ually. An increase to regression model fit indicates that PS pollution
weakens the correlation between land use and water quality. No in-
crease to regression model fit indicates the PS pollution does not affect
the correlation between land use and water quality in the excluded
cluster.
2.2.6. Correlation analysis
The correlation analysis reveals the variation of land use-water qual-
ity correlation. Under our hypothesis, the larger the PS pollution is, the
smaller the correlations are in clusters. So the correlation analysis
helps us to verify the disturbance of PS pollution to the land use-water
quality correlation. Correlations between land use categories and
water quality parameters in different clusters were tested using
Pearson's correlation coefficients with statistical significances at p b
0.01and p b 0.05 levels (2-tailed). The one-sample Kolmogorov–
Smirnov test was used in testing the normal distribution for all variables
(Mirabella, 2006). The results showed all 2-tailed p-values are greater
than 0.05, which indicates the suitability of the data for correlation
analysis.
Fig. 4. Visualization of each variable in the trained SOM. Note: “d” under legend means denormalization of the SOM results.
595
3. Results
3.1. Classification and visualization of sampling sites
Before the classification, we trained the SOM with various map sizes
pffiffiffi
around the number determined by the heuristic equation 5 n, where
n = 139. We selected 54 (6 × 9) as the map size (Fig. 4) because the
QE (0.092) and TE (0.029) are minimum.
Fig. 4 shows the SOM of the water quality dataset and the land use
categories. Some patterns of similarity among water quality parameters
could be distinguished. For example, the sampling sites with the highest
values for both nitrogen and phosphorus are located in the lower left
−
part of the SOM plane. NH+ 4 -N, NO3 -N and TN show similar patterns,
meaning the value decreases from the sites located in the lower left
part to the sites located in the upper right part of the SOM plane. SRP
and TP show high values in the corner on the left side of the plane.
CODMn concentration of the sampling sites at the lower part of the
plane is high and decreases towards the upper part of the plane.
The correlation between land use categories and water quality pa-
rameters is also clearly discerned. For example, water samples at the
left corner of the bottom on the map have high nitrogen and phospho-
rous concentrations. This generally corresponds with the trend of Agri-
culture and Built-up. The sampling sites with high proportions of
Woodland show low values of water quality parameters. The sampling
sites on the lower left part of the plane show high values of water qual-
ity parameters, where the Woodland percentages are low.
Fig. 5 shows the grouping of all water sampling sites. Seven clusters
were formed according to the Davies–Bouldin index, where minimal
596 P. Zhou et al. / Science of the Total Environment 543 (2016) 591–600

Fig. 5. (A) dendrogram of the hierarchical cluster analysis according to the Ward linkage method using Euclidean distance, defining seven clusters. (B) Davies–Bouldin index plot.
(C) Distribution of sampling sites on the SOM according to the six water quality parameters, and clustering of the trained SOM. Small font codes are sampling site identifiers. The sampling
sites are grouped into six clusters because cluster 7 is empty.

value is at 7 clusters. The sampling sites are grouped into six clusters be-
cause cluster 7 is empty.
3.2. Various patterns of water quality and land use
The differences between water quality parameters for clusters are
significant, as all Kruskal–Wallis tests of significance have p-values
less than 0.01. This confirmed the reasonability of our classification.
Different clusters show different water quality patterns (Fig. 6A). Ac-
cording to Chinese Environmental Quality Standard for Surface Water (GB
3838-2002), we evaluated the classes for NH+ 4 -N, TN, CODMn and TP in
each cluster (Table 2). Among the six clusters, cluster 1 shows the
least water pollution because the concentration of each water quality
parameter is the lowest (Fig. 6A). Clusters 2 and 3 have the same
water quality classes (Table 2), though the water quality parameters
vary in the concentration (Fig. 6A). Water in clusters 1, 2 and 3 satisfies
the requirements for a drinking water source, while water in clusters 4,
5 and 6 does not. This is because TN in clusters 4 and 5 are class IV and
NH+ 4 -N, TN and TP in cluster 6 exceed the upper limit of class V. In clus-
ters 5 and 6, the CODMn values are much higher than other clusters (Fig.
6A, Table 2). The concentrations of water quality parameters, except the
CODMn, are higher in cluster 6 than in other clusters (Fig. 6A). Therefore,
water pollution generally increases with the cluster number from 1 to 6,
while the clusters 2 and 3 are the same in terms of water quality classes.
Fig. 6B shows the land use pattern in each cluster and Fig. 7 shows
the location of the sampling sites. Among the six clusters, cluster 1 is
mainly in Yongtai, Dehua and Pucheng Counties, whose forest coverages
rank the top among the counties. The sub-watersheds in cluster 1 have
high proportions of Woodland and Unused land, medium proportions of
Orchard, and low proportions of Agriculture, Built up and Water. Cluster
2 is mainly in Pucheng, Guangze, Youxi, Taining Counties, which are fa-
mous for agricultural products and ecotourism. The sub-watersheds in
cluster 2 have high proportions of Water, medium proportions of
Woodland, Agriculture Orchard and Unused land, and low proportions

Fig. 6. Average values and 95% confidence intervals for (A) the water quality parameters and (B) land use categories for each cluster. The Kruskal–Wallis test indicates the differences be-
tween water quality parameters among clusters are significant (p b 0.01). “Site Numbers” in the legend for panel A indicates the number of sites in each cluster.
 P. Zhou et al. / Science of the Total Environment 543 (2016) 591–600 597

Table 2
The classes of water quality parameters in each cluster.⁎

Cluster 1 Cluster 2 Cluster 3 Cluster 4 Cluster 5 Cluster 6

NH+4 -N II II II II II Exceed V
TN III III III IV IV Exceed V
CODMn II II II II III III
TP II III III II III Exceed V
⁎ According to Chinese Environmental Quality Standard for Surface Water (GB 3838-
2002). “Exceed V” means the water quality parameter's concentration is beyond the upper
level of class V.

of Built-up. Cluster 3 is mainly in Jianyang and Jian'ou and Shunchang

Counties, which are famous for their forestry and grain. The sub-
watersheds in this cluster have high proportions of Woodland, and
low proportions of Agriculture, Built-up and Orchard. Cluster 4 is mainly
in Shaxian, Yong'an, Jian'ou, Jianning, Nanping and Sanming Counties.
Industrial and agricultural activities are intensive in these regions (Fig.
2B). The sub-watersheds in cluster 4 have high proportions of Agricul-
ture, Built-up and Orchard, medium proportions of Water and low pro-
portions of Woodland. Cluster 5 is mainly in Jian'ou and Nanping Fig. 8. Fits of the backward stepwise regression models. Legend indicates the cluster that is
Counties, where treated sewage is 67% and 89% of total sewage dis- excluded when performing the regression. The dashed horizontal lines indicate the R2
charge (Fig. 2B). The sub-watersheds in this cluster have high propor- values of the models based on all the data.
tions of Built-up and Water, medium proportions of Woodland and
Agriculture, and low proportions of Orchard and Unused land. Cluster
6 is mainly in Shaxi and Gutian Counties, where industry and aquatic excluding cluster 1, R2 increased for TP. When excluding cluster 2, R2 in-
production dominate the human activities respectively. The sub- creased for all the regression models. When excluding cluster 3, R2 in-
−
watersheds in this cluster have high proportions of Agriculture, Built- creased for NH+ 4 -N, NO3 -N, TN and SRP. When excluding cluster 4, R
2
+
up and Unused land, and low proportions of Woodland. increased for NH4 -N, TP, SRP and CODMn. When excluding cluster 5,
R2 increased for NO− 2
3 -N and TN. When excluding cluster 6, R increased
−
3.3. Comparison of the backward stepwise regression model fit for NH+ 2
4 -N, NO3 -N, TN and CODMn. An increase of R in the regression
models indicates the PS pollution affects the water quality in the ex-
We excluded the data in each cluster individually and performed the cluded cluster.
regression analysis. Fig. 8 shows the fits of the regression models. When

Fig 7. Six clusters of the 139 sampling sites.

598 P. Zhou et al. / Science of the Total Environment 543 (2016) 591–600

3.4. Correlation between water quality and land use in different clusters land as a proxy for point source pollution (Ahearn et al., 2005; Shen
et al., 2014).
Correlation analysis shows various patterns of correlation between In this study, we detected various water quality patterns in the six
land use and water quality (Table 3). We did not perform the correlation clusters. Cluster 1 has the lowest concentration for all the water quality
analysis for cluster 5 and 6 because the sample size is small (Fig. 6A). In parameters (Fig. 6A, Table 2). This is because the sub-watersheds in the
cluster 1, NH+4 -N is correlated positively with the proportion of Built-up cluster 1 counties, such as Taining, Dehua and Pucheng (Fig. 7), have for-
(p b 0.01). SRP is correlated positively with the proportion of Unused est coverages that rank among the highest in the study area (Fig. 6B).
land (p b 0.01). CODMn is correlated positively with the proportion of Clusters 2 and 3 have the same water quality class (Table 2), though
Woodland (p b 0.05), and negatively with the proportion of Unused the concentration of each water quality parameter varies (Fig. 6A).
land, Orchard (p b 0.01) and Built-up (p b 0.05). In cluster 2, NH+ 4 -N is The concentrations of water quality parameters in cluster 2 and 3 are
correlated positively with the proportion of Built-up (p b 0.01) and Ag- less than in clusters 4–6. The purification ability of Woodland and rela-
riculture (p b 0.05), and negatively with the proportion of Woodland tively low pollution from the medium intensity of agricultural activities
(p b 0.05). SRP is correlated positively with the proportion of Agricul- might explain this phenomenon (Fig. 6B). Water in clusters 4, 5 and 6 is
ture and negatively with the proportion of Woodland (p b 0.01). In clus- not suitable for providing drinking water according to Chinese Environ-
ter 3, TP is correlated positively with the proportion of Built-up (p b mental quality standards for surface water (GB 3838-2002) (Ministry of
0.05). SRP is correlated positively with the proportion of Water (p b Environmental Protection of the People's Republic of China, 2002). The
0.01). In cluster 4, there are no significant correlations between land sub-watersheds in clusters 4–6 are mainly in urban areas, such Yong'an,
use categories and any water quality parameters. Sanming, Shaxian, Nanping and Jianning (Fig. 6). The Built-up and Agri-
culture proportions are high, while the Woodland proportion is low in
these regions (Fig. 6B). The industrial, domestic and agricultural pollut-
4. Discussion ants could be the cause of the contaminated water. There are waste
water treatment facilities in the counties. But the various treated sew-
4.1. Water quality patterns in each cluster age percentages in the cities (such as Sanming and Yong'an, Fig. 2)
and the high pollutant concentration discharged from the waste water
Watershed land use affects water quality through NPS pollution treatment facilities (Fujian Provincial Department of Environmental
(Swaney et al., 2012). Agricultural activities, such as fertilizer and pesti- Protection, 2015) still contaminate the riverine water.
cide application, are recognized as important factors influencing water
quality (Smith et al., 2013; UNEP, 2014). Agriculture and Orchard are 4.2. Contribution of NPS pollution from land use and PS pollution to the wa-
recognized as the sources of NPS pollution in the study area. Meanwhile, ter quality
PS pollution from domestic and industrial activities in Built-up area is
also recognized as one of the most important factors threatening Lots of research documents that the relationship between land use
water quality (Wu and Chen, 2013). Some research considers urban and water quality indeed exist (Huang et al., 2013b, Bu et al., 2014).
However, researchers seldom consider the influence of PS pollution on
Table 3 the correlation between land use and water quality. In this study, we
Pearson's correlation coefficients between land use categories and water quality specified whether PS or NPS exerts more influence on water pollution
parameters. in the six clusters.
Woodland Water Unused land Orchard Built-up Agriculture From the backward stepwise regression results (Fig. 8), there are in-
creased R2 in every cluster. Under our hypothesis, this increase indicates
Cluster 1
NH+ −.345 −.098 .474⁎⁎
that PS pollution influences the correlation between land use and water
4 -N .167 .051 .312
NO−3 -N −.227 .036 −.128 −.047 .316 .261 quality.
TN −.195 .191 −.335 −.156 .331 .331 In clusters 1, 2 and 3, both the NPS pollution from land use and PS
TP .013 .089 −.035 .132 .178 −.129 pollution are not severe. This is because water quality parameters in
SRP −.128 −.093 .431⁎ −.034 −.065 −.031
these clusters meet the standard for cleanliness (Table 1). The sub-
CODMn .436⁎ .079 −.476⁎⁎ −.505⁎⁎ −.458⁎ −.070
watersheds in clusters 1–3 are mainly in counties whose industrial ac-
Cluster 2 tivities are not intensive (Fig. 7). This accounts for the apparently light
NH+4 -N −.415⁎ −.151 −.263 .074 .478⁎⁎ .413⁎
pollution from PS. The high proportion of Woodland and low propor-
NO−3 -N −.052 .288 .313 .260 .037 −.162
TN −.190 .074 .018 .203 .250 .074 tions of Agriculture, Orchard and Built-up (Fig. 6B) might explain the
TP .081 −.206 −.115 −.233 −.019 .000 light NPS pollution.
SRP −.457⁎⁎ .011 −.126 .245 .338 .454⁎⁎ In cluster 4, both NPS pollution from land use and PS pollution con-
CODMn −.045 .026 −.194 −.141 .030 .147 tribute to the pollution. PS pollution influences all the water quality pa-
Cluster 3 rameters, especially NH+ 4 -N, TP, SRP and CODMn (Fig. 8). This can be
NH+4 -N .095 .224 .118 .042 .114 −.252 attributable the intensive industrial activities and domestic waste
NO−3 -N .018 .011 .143 .107 .028 −.105 water in the counties, such as Yong'an, Sanming, Shaxian, Jianning and
TN .114 .042 −.042 −.020 .092 −.169
TP −.260 .191 .129 −.169 .393⁎ .070
Jian'ou (Fig. 2, Fig. 7). Meanwhile, the sub-watersheds in cluster 4
SRP −.100 .516⁎⁎ .169 −.271 .128 −.062 have high proportions of Agriculture, Orchard and Built-up (Fig. 6B).
CODMn .048 −.302 .071 −.218 −.118 .013 These land use categories are closely related with NPS pollution
(Smith et al., 2013; Hur et al., 2014, Huang et al., 2015). Only TN (class
Cluster 4
NH+4 -N −.148 .148 −.266 .061 .272 .101 IV) is substandard (Table 2). The NPS pollution and PS pollution are
NO−3 -N −.005 −.200 −.043 .301 .048 −.033 not severe enough to make NH+ 4 -N, TP, SRP and CODMn to be substan-
TN .016 .050 −.153 .188 .072 −.051 dard. The substandard TN is likely caused by the joint effects of NPS pol-
TP −.166 −.035 .053 −.054 .085 .179
lution from the high proportions of Agriculture, Orchard and Built-up
SRP .032 −.110 .129 −.225 −.174 .050
CODMn .070 .094 −.123 −.025 −.189 .026 and PS pollution from urban areas.
In cluster 5, PS pollution plays a major role in controlling the water
We did not perform the correlation analysis for clusters 5 and 6 because the sample sizes
are small.
quality. R2 increases for the NO−3 -N and TN regression models, which in-
⁎⁎ p b 0.01 (2-tailed). dicates that PS pollution influences these two water quality parameters.
⁎ p b 0.05 (2-tailed). The sub-watersheds are mainly in urban area such as Jian'ou and
 P. Zhou et al. / Science of the Total Environment 543 (2016) 591–600
Nanping Counties (Fig. 7). The treated sewage percentages are lower
than other places (Fig. 2). We suspect that industrial and domestic dis-
charges lead to the severe PS pollution. The medium proportion of Agri-
culture and low proportion of Orchard (Fig. 6B) in this cluster indicate
that the NPS pollution is not severe. TN is the only substandard param-
eter (Table 2), thus we conclude the PS pollution controls the TN, while
PS pollution's influence is not severe enough to make the NO− 3 -N to be
substandard.
In cluster 6, PS pollution controls NH+
4 -N, TN and CODMn, while NPS
pollution controls the high concentration of NO− 3 -N, TP and SRP. PS pol-
lution influences NH+ 4 -N, TN and CODMn (Fig. 8). The intensive indus-
trial activities in Shaxi and Gutian counties could explain the severe
influence of the PS pollution. Meanwhile, the concentrations of water
quality parameters, except CODMn, are the highest in cluster 6. CODMn
concentration in the discharge from waste water treatment facilities is
standard, while the concentrations of NH+ 4 -N, TN and TP are beyond
the upper limit of class V (Fujian Provincial Department of
Environmental Protection, 2015). This explains why the great influence
of PS pollution does not lead to high concentrations of CODMn. We sus-
pect the high proportion of Agriculture (Fig. 6B) causes the high concen-
trations of NO− 3 -N, TP and SRP.
We recommend that water management strategies in the study area
should attempt to control both NPS pollution from land and PS pollution
in sub-watersheds of cluster 4 and 6. Controlling PS pollution is the key
to improve the water quality in cluster 5. Though water quality in clus-
ter 1, 2 and 3 meets the cleanliness standard, PS pollution still exists and
warrants attention.
4.3. Spatial variation of the land use-water quality correlation and the dis-
turbance from PS pollution
The correlation between land use and water quality varies spatially
in the study area (Table 3). Such spatial variation is documented by
other researchers. The general explanation is that pollution sources
tend to vary across study areas, particularly among watersheds that
have various land uses (Tu and Xia, 2008, Kang et al., 2010, Tu, 2011,
Huang et al., 2015). There is a general understanding that forest can ab-
sorb nutrients, while built-up land and agricultural land are associated
with PS and NPS pollution (Tu, 2011, Bu et al., 2014, Hur et al., 2014,
Wan et al., 2014, Huang et al., 2015). This understanding could help in
explaining the negative correlation between NH+ 4 -N, SRP and the pro-
portion of Woodland in Cluster 1, the positive correlation between
Built-up and NH+ 4 -N in clusters 1 and 2, Built-up and TP in cluster 4,
and Agriculture and NH+ 4 -N, SRP in cluster 1.
SRP is positively correlated with the proportion of Water cover in
cluster 3. This is similar to the observation by Wang et al. (2014) and ex-
plained as the water body accumulates phosphorus migrating from
other land uses, i.e., Agriculture and Orchard. If the accumulated nutri-
ents greatly exceed the retention capacity of the water body, then the
water serves as “transit shed” and transports nutrients to the catchment
outlets. The watercourse and aquatic biology characteristic might con-
tribute to this feature in cluster 3 and make it different from other
clusters.
Surprisingly, CODMn is positively correlated with the proportion of
Woodland and negatively correlated with the proportions of Unused
land, Orchard and Built-up in cluster 1. This result is unexpected and
contradicts other research (Bu et al., 2014, Hur et al., 2014, Lee et al.,
2014, Huang et al., 2015). Most sampling sites in cluster 1 are located
in Yongtai, Dehua and Pucheng Counties (Fig. 7), whose forest cover-
ages rank at the top among counties in the study area (Fig. 6B). There-
fore, a possible reason of this special phenomenon is that the organic
matter derived from decaying plant material is refractory, while the or-
ganic matter from industrial land decomposes easily after being
transported into river via surface runoff (Nguyen and Hur, 2011, Lee
et al., 2014). The increase of the Woodland in cluster 1 would cause ac-
cumulation of refractory organic matter, while the increases of Built-up,
599
Orchard and Unused land would increase the labile organic matter that
has decomposed at the tributary outlet. Another possible reason is that
agricultural practices in cluster 1 might hinder much loss of organic
matter into the runoff water from vegetable fields as compared with
those from forest (Yang et al., 2013). In this case, the increase of Orchard
might hinder the organic matter loss into the runoff water.
Our results show that PS pollution influences the correlation be-
tween land use and water quality. From the stepwise regression analy-
sis, we found PS pollution influences the water quality in every cluster.
(Table 2, Fig. 8). PS pollution is the least severe in cluster 1, and followed
by cluster 2 and 3. The PS pollution in clusters 4, 5 and 6 are more severe
than in the other clusters. In the correlation analysis (Table 3), cluster 1
has the most pairs of significant correlations. Each of cluster 2 and clus-
ter 3 has fewer pairs of significant correlations than cluster 1. There are
no significant land use-water quality correlations in cluster 4. Therefore,
the more severe the PS pollution is, the fewer the pairs of significant
land use-water quality correlations exist. We conclude that PS pollution
hides land use-water quality correlation in cluster 4 and weakens the
correlation in clusters 1, 2 and 3.
4.4. Limitations
The relationship between water quality and land use can be incon-
sistent across space and time. Many researches document this inconsis-
tency in terms of seasonal or inter-annual variations (Lee et al., 2009,
Huang et al., 2013b, Hur et al., 2014). In this study, we obtained the
data from July to August, so the results reveal the relationship between
water quality parameters and land use categories during only this cer-
tain season. The frequent survey of a large river basin is costly in
terms of both time and money. Nevertheless we plan to carry out
more surveys in the future to reveal the temporal variations of the rela-
tionship, so that we can make water management strategies for various
seasons.
In this study, we focused on the influences of PS and NPS pollution
on the water quality. However, many factors affect the land use-water
quality correlation. In terms of land use, the landscape composition,
configuration (Shen et al., 2014) and spatial scale (Shen et al., 2015)
are important when analyzing the land use-water quality relationship.
Weather conditions (Huang et al., 2014; Yu et al., 2016) and topo-
graphic characteristics such as slope (Yu et al., 2016) could also influ-
ence water quality. The influence of abovementioned factors might
explain the small R2 in the backward stepwise regressions (Table 2)
and should be investigated in the future. Also, different quantity of sam-
pling sites might lead to different results from SOM. These are important
issues for future work.
5. Conclusion
This study developed a SOM-based approach to explore the relation-
ship between river water quality and land use in the Minjiang River Wa-
tershed. Six clusters were delineated and visualized for the 139
headwater sub-watersheds. The degrees of pollution, patterns of
water quality and types of land use vary among the six clusters. PS pol-
lution exists in every cluster and exerts various levels of contamination.
In the slightly polluted clusters, both the NPS pollution from land use
and PS pollution generate low pollutants. In the heavily polluted clus-
ters, both NPS and PS pollution influences water quality parameters.
PS pollution weakens or even hides the correlation between land use
and water quality. The more severe the PS pollution is, the fewer the
pairs of significant land use-water quality correlations there are in the
clusters. This study demonstrates how PS pollution can mask the corre-
lation between land use and water quality. The method can help to ex-
pose the roles of NPS and PS pollution in order to explain water quality
in sub-watersheds across a watershed.
600 P. Zhou et al. / Science of the Total Environment 543 (2016) 591–600

Acknowledgments Li, Y., Jiao, J., Wang, Y., Yang, W., Meng, C., Li, B., Li, Y., Wu, J., 2015. Characteristics of nitro-
gen loading and its influencing factors in several typical agricultural watersheds of
subtropical China. Environ. Sci. Pollut. Res. 22, 1831–1840.
This study was supported by the National Natural Science Founda- Maarten, D.J., Bart, V.d.V., Ronny, B., Lieven, B., 2008. Responses of aquatic organisms to
tion of China (Grant No. 41471154). Anonymous reviewers supplied metal pollution in a lowland river in Flanders: a comparison of diatoms and macro-
invertebrates. Sci. Total Environ. 407, 615–629.
constructive feedback that helped to improve this paper. We thank Martin, S., Soranno, P., Bremigan, M., Cheruvelil, K., 2011. Comparing hydrogeomorphic
Huan Gu at Clark University for suggestions on revising this paper. approaches to lake classification. Environ. Manag. 48, 957–974.
Ministry of Environmental Protection of the People's Republic of China, 2002l. Environ-
mental Quality Standards for Surface Water (GB 3838-2002).
References
Mirabella, J., 2006. Hypothesis Testing With SPSS: A Non-statistician's Guide & Tutorial.
Available from: http://www.drjimmirabella.com/ebook/.
Ahearn, D.S., Sheibley, R.W., Dahlgren, R.A., Anderson, M., Johnson, J., Tate, K.W., 2005.
Mouri, G., Takizawa, S., Oki, T., 2011. Spatial and temporal variation in nutrient parame-
Land use and land cover influence on water quality in the last free-flowing river
ters in stream water in a rural–urban catchment, Shikoku, Japan: effects of land
draining the western Sierra Nevada, California. J. Hydrol. 313, 234–247.
cover and human impact. J. Environ. Manag. 92, 1837–1848.
Baker, A., 2003. Land use and water quality. Hydrol. Process. 17, 2499–2501.
Nguyen, H.V.-M., Hur, J., 2011. Tracing the sources of refractory dissolved organic matter
Bierman, P., Lewis, M., Ostendorf, B., and Tanner. J. 2011. A review of methods for
in a large artificial lake using multiple analytical tools. Chemosphere 85, 782–789.
analysing spatial and temporal patterns in coastal water quality. Ecol. Indic., 11:
Olkowska, E., Kudłak, B., Tsakovski, S., Ruman, M., Simeonov, V., Polkowska, Z., 2014. As-
103–114.
sessment of the water quality of Kłodnica River catchment using self-organizing
Bowes, M.J., Smith, J.T., Jarvie, H.P., Neal, C., 2008. Modelling of phosphorus inputs to riv-
maps. Sci. Total Environ. 476–477, 477–484.
ers from diffuse and point sources. Sci. Total Environ. 395, 125–138.
Park, Y.-S., Kwon, Y.-S., Hwang, S.-J., Park, S., 2014. Characterizing effects of landscape and
Bu, H., Meng, W., Zhang, Y., Wan, J., 2014. Relationships between land use patterns and
morphometric factors on water quality of reservoirs using a self-organizing map. En-
water quality in the Taizi River basin, China. Ecol. Indic. 41, 187–197.
viron. Model Softw. 55, 214–221.
Céréghino, R., Park, Y.S., 2009. Review of the self-organizing map (SOM) approach in
Shen, Z., Hou, X., Li, W., Aini, G., 2014. Relating landscape characteristics to non-point
water resources: commentary. Environ. Model Softw. 24, 945–947.
source pollution in a typical urbanized watershed in the municipality of Beijing.
Chen, Y., CHEN, X.-w., YIN, Y.-x., 2011. Characteristics of runoff changes in the Minjiang
Landsc. Urban Plan. 123, 96–107.
River Basin from 1960 to 2006(in Chinese). Journal of Natural Resources 8,
Shen, Z., Hou, X., Li, W., Aini, G., Chen, L., Gong, Y., 2015. Impact of landscape pattern at
1402–1411.
multiple spatial scales on water quality: a case study in a typical urbanised watershed
Choi, B.-Y., Yun, S.-T., Kim, K.-H., Kim, J.-W., Kim, H.M., Koh, Y.-K., 2014. Hydrogeochemi-
in China. Ecol. Indic. 48, 417–427.
cal interpretation of South Korean groundwater monitoring data using self-
Smith, A.P., Western, A.W., Hannah, M.C., 2013. Linking water quality trends with land use
organizing maps. J. Geochem. Explor. 137, 73–84.
intensification in dairy farming catchments. J. Hydrol. 476, 1–12.
Chon, T.-S., 2011. Self-organizing maps applied to ecological sciences. Ecological Informat-
State Environmental Protection Administration of China, 2002. Methods of Monitoring
ics 6, 50–61.
and Analyzing for Water and Wastewater. fourth ed. China Environmental Science
Compin, A., Céréghino, R., 2007. Spatial patterns of macroinvertebrate functional feeding
Press, Beijing.
groups in streams in relation to physical variables and land-cover in Southwestern
Swaney, D.P., Humborg, C., Emeis, K., Kannen, A., Silvert, W., Tett, P., Pastres, R., Solidoro,
France. Landsc. Ecol. 22, 1215–1225.
C., Yamamuro, M., Hénocque, Y., Nicholls, R., 2012. Five critical questions of scale for
Fujian Provincial Department of Environmental Protection, 2015. Information Platform
the coastal zone. Estuar. Coast. Shelf Sci. 96, 9–21.
for Key Pollution Sources in Fujian Province. (http://Wryfb.fjemc.org.cn/index.aspx.
Tu, J., 2011. Spatially varying relationships between land use and water quality across an
access data: Oct.4th.2015).
urbanization gradient explored by geographically weighted regression. Appl. Geogr.
Fujian Statistical Bureau, 2014. “Fujian Statistical Yearbook (2014)”.
31, 376–392.
Huang, J., Huang, Y., Pontius Jr., R.G., Zhang, Z., 2015. Geographically weighted regression
Tu, J., Xia, Z.-G., 2008. Examining spatially varying relationships between land use and
to measure spatial variations in correlations between water pollution versus land use
water quality using geographically weighted regression I: model design and evalua-
in a coastal watershed. Ocean Coast. Manag. 103, 14–24.
tion. Sci. Total Environ. 407, 358–378.
Huang, J., Huang, Y., Zhang, Z., 2014. Coupled effects of natural and anthropogenic con-
UNEP, 2014. In: Bringezu, S., Schütz, H., Pengue, W., O'Brien, M., Garcia, F., Sims, R.,
trols on seasonal and spatial variations of river water quality during baseflow in a
Howarth, R., Kauppi, L., Swilling, M., Herrick, J. (Eds.), Assessing Global Land Use:
coastal watershed of Southeast China. PLoS One 9, e91528. http://dx.doi.org/10.
Balancing Consumption with Sustainable Supply. A Report of the Working Group
1371/journal.pone.0091528.
on Land and Soils of the International Resource Panel.
Huang, J., Li, Q., Huang, L., Zhang, Z., Mu, J., Huang, Y., 2013a. Watershed-scale evaluation
UN General Assembly, 2004. Oceans and the Law of the Sea, Report of the Secretary-
for land-based nonpoint source nutrients management in the Bohai Sea Bay, China.
General of 18 August 2004. (A/59/62/Add.1, 29, para. 9. http://www.un.org/Depts/
Ocean Coast. Manag. 71, 314–325.
los/general_assembly/general_assembly_reports.htm. access data: Oct.17th.2015).
Huang, J., Li, Q., Pontius Jr., R.G., Klemas, V., Hong, H., 2013b. Detecting the dynamic link-
US National Ocean Council, 2010. Strategic Action Plans. (https://www.whitehouse.gov/
age between landscape characteristics and water quality in a subtropical coastal wa-
administration/eop/oceans/sap. access data: Oct 17th, 2015).
tershed, Southeast China. Environ. Manag. 51, 32–44.
Uuemaa, E., Roosaare, J., Mander, Ü., 2005. Scale dependence of landscape metrics and
Hur, J., Nguyen, H.-M., Lee, B.-M., 2014. Influence of upstream land use on dissolved or-
their indicatory value for nutrient and organic matter losses from catchments. Ecol.
ganic matter and trihalomethane formation potential in watersheds for two different
Indic. 5, 350–369.
seasons. Environ. Sci. Pollut. Res. 21, 7489–7500.
Vesanto, J., Alhoniemi, E., 2000. Neural Networks 11. IEEE Transactions on, pp. 586–600.
Jeong, K.-S., Hong, D.-G., Byeon, M.-S., Jeong, J.-C., Kim, H.-G., Kim, D.-K., Joo, G.-J., 2010.
Vesanto, J., Himberg, J., Alhoniemi, E., Parhankangas, J., 2000. SOM toolbox for Matlab 5.
Stream modification patterns in a river basin: field survey and self-organizing map
Technical Report A57. Neural Networks Research Centre, Helsinki University of Tech-
(SOM) application. Ecological Informatics 5, 293–303.
nology, Helsinki, Finland.
Kalteh, A.M., Hjorth, P., Berndtsson, R., 2008. Review of the self-organizing map (SOM)
Wang, Y., Li, Y., Liu, X., Liu, F., Li, Y., Song, L., Li, H., Ma, Q., Wu, J., 2014. Relating land use
approach in water resources: analysis, modelling and application. Environ. Model
patterns to stream nutrient levels in red soil agricultural catchments in subtropical
Softw. 23, 835–845.
central China. Environ. Sci. Pollut. Res. 21, 10481–10492.
Kang, J.-H., Lee, S.W., Cho, K.H., Ki, S.J., Cha, S.M., Kim, J.H., 2010. Linking land-use type and
Wan, R., Cai, S., Li, H., Yang, G., Li, Z., Nie, X., 2014. Inferring land use and land cover impact
stream water quality using spatial data of fecal indicator bacteria and heavy metals in
on stream water quality using a Bayesian hierarchical modeling approach in the
the Yeongsan river basin. Water Res. 44, 4143–4157.
Xitiaoxi River Watershed, China. J. Environ. Manag. 133, 1–11.
Kohonen, T., 2001. Self-Organizing Maps. Springer, Berlin.
Wu, Y., Chen, J., 2013. Investigating the effects of point source and nonpoint source pollu-
Kohonen, T., 2013. Essentials of the self-organizing map. Neural Netw. 37, 52–65.
tion on the water quality of the East River (Dongjiang) in South China. Ecol. Indic. 32,
Löhr, S.C., Grigorescu, M., Hodgkinson, J.H., Cox, M.E., Fraser, S.J., 2010. Iron occurrence in
294–304.
soils and sediments of a coastal catchment: a multivariate approach using self
Xiao, H., Ji, W., 2007. Relating landscape characteristics to non-point source pollution in
organising maps. Geoderma 156, 253–266.
mine waste-located watersheds using geospatial techniques. J. Environ. Manag. 82,
Lee, S.-W., Hwang, S.-J., Lee, S.-B., Hwang, H.-S., Sung, H.-C., 2009. Landscape ecological
111–119.
approach to the relationships of land use patterns in watersheds to water quality
Yang, Y., He, Z., Wang, Y., Fan, J., Liang, Z., Stoffella, P.J., 2013. Dissolved organic matter in
characteristics. Landsc. Urban Plan. 92, 80–89.
relation to nutrients (N and P) and heavy metals in surface runoff water as affected
Lee, Y., Hur, J., Shin, K.-H., 2014. Characterization and source identification of organic mat-
by temporal variation and land uses — a case study from Indian River Area, south
ter in view of land uses and heavy rainfall in the Lake Shihwa, Korea. Mar. Pollut. Bull.
Florida, USA. Agric. Water Manag. 118, 38–49.
84, 322–329.
Yu, S., Xu, Z., Wu, W., Zuo, D., 2016. Effect of land use types on stream water quality under
Li, F., Cai, Q., Qu, X., Tang, T., Wu, N., Fu, X., Duan, S., Jähnig, S.C., 2012. Characterizing mac-
seasonal variation and topographic characteristics in the Wei River basin, China. Ecol.
roinvertebrate communities across China: large-scale implementation of a self-
Indic. 60, 202–212.
organizing map. Ecol. Indic. 23, 394–401.
Li, S., Xia, X., Tan, X., Zhang, Q., 2013. Effects of catchment and riparian landscape setting
on water chemistry and seasonal evolution of water quality in the upper Han River
basin, China. PLoS One 8, e53163. http://dx.doi.org/10.1371/journal.pone.0053163.