Soil & Tillage Research 155 (2016) 19–26

Contents lists available at ScienceDirect

Soil & Tillage Research

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

A comprehensive support vector machine-based classification model

for soil quality assessment
Yong Liua,* , Huifeng Wanga , Hong Zhanga , Karsten Libera,b
a
Institute of Loess Plateau, Shanxi University, Taiyuan, Shanxi, China
b
Toxicology Centre, University of Saskatchewan, Saskatoon, Saskatchewan, Canada

A R T I C L E I N F O A B S T R A C T

Article history: Soil quality is defined here as the capacity of soil to have biological function, to sustain plant and animal
Received 27 April 2015 production, to maintain or enhance water and air quality and to support human health and habitation.
Received in revised form 11 July 2015 There are different soil quality assessment models based on diverse methods and data, but none of the
Accepted 14 July 2015
models can fully meet all purposes. The selection of an appropriate soil classification model therefore
becomes an important aspect in soil quality assessment. This paper presents a new comprehensive
Keywords: support vector machine-based classification model for classification of urban soil quality and then uses
Soil quality assessment
that model to assess the soil quality of Taiyuan relative to Chinese environmental quality standards and
Support vector machine
Comprehensive classification model
regional background values. The results indicated that the support vector machine-based soil quality
Heavy metal contamination model combined soil heavy metal contamination and soil fertility data satisfactorily, with an accuracy of
Soil fertility 98.3333%. The soil quality of Taiyuan was subsequently divided into five classes (IA, IB, IC, IIA and IIB).
Fifty percent of all samples were classified as class IB, indicating that soil quality within the study area
was good. This paper shows that a comprehensive support vector machine-based classification model is
feasible and reliable for soil quality assessment. Furthermore, the assessment presented could provide
references for related ecological problems.
ã 2015 Elsevier B.V. All rights reserved.

1. Introduction and soil sustainability. Urban soil environmental quality has

With the rapid industrialization and urbanization that has
occurred in most cities of the world in recent decades, there have
been increasing concern for and attention on soil quality. The main
threats to soils are identified as a decline in organic matter,
increased soil erosion, compaction and salinization and an
increased probability of floods, landslides, contamination, acidifi-
cation and sealing (Montanarella and Rusco, 2008). Urban soil is an
important component of urban ecosystems, and it provides a
medium and nutrients to urban landscape plants and crops. It is
also a sink and source of urban contamination and thus can affect
urban eco-environmental quality and human health (Cachada
et al., 2012; Praveena et al., 2014). Urban soil quality is influenced
by both natural and anthropogenic factors, but humans play the
most influential role in altering the performance characteristics of
soils by mining, industry, agriculture, waste treatment and
transportation (Karim et al., 2014; Yu et al., 2014).
The assessment of soil quality is generally based on a
combination of soil environmental quality, soil productivity
* Corresponding author.
E-mail address: Liuyong@sxu.edu.cn (Y. Liu).
often been impacted by a substantial amount of contaminations,
including heavy metals, originating from different sources. Heavy
metals do not degrade by biological or biochemical processes,
although their speciation may change, become bioaccumulated
and in a few cases, biomagnified along the food chain (Liao et al.,
2006). Hence, heavy metals in soils are widely considered to
potentially have unfavorable environmental impacts, as well as
affect human health and prosperity (e.g., economic losses, safety
of agricultural products and endangerment of human health,
especially children) (Guo et al., 2012; Praveena et al., 2014). In
many cases, urban soil productivity has been impacted by
anthropogenic factors such as fertilization and wastewater
irrigation, which increase soil fertility (Bhaduri and Purakayastha,
2014). Soil fertility in turn can affect vegetation composition,
plant community biomass and plant functional group biomass.
Many researchers and practicing farmers have also observed that
fertile soil with high soil organic matter content and diversity of
soil microbiota enhances vegetation health through various
processes (Zheng et al., 2015). Urban soil also provides nutrients
for both crops and ornamental plants. Thus, soil fertility is
another important parameter to assess to determine urban soil
productivity and sustainability.

http://dx.doi.org/10.1016/j.still.2015.07.006
0167-1987/ ã 2015 Elsevier B.V. All rights reserved.
20 Y. Liu et al. / Soil & Tillage Research 155 (2016) 19–26
The risk assessment methods used to evaluate soil heavy metal
contamination and soil fertility assessment are diverse (Bhuiyan
et al., 2010; Li et al., 2014a; Saby et al., 2009). Traditional methods
of soil heavy metal contamination assessment include the single-
factor index and the integrated pollution index (Xu and Hu, 2014),
the analytic hierarchy process (AHP) (Chen et al., 2012), the fuzzy
comprehensive evaluation method (Fu et al., 2014; Hui, 2013), and
the geo-accumulation index method (Akbulut et al., 2013). The
geo-accumulation index method is recommended by the U.S.
Environmental Protection Agency and widely used by scholars. The
traditional methods of soil fertility assessment include the analytic
hierarchy process (AHP) (Zhou et al., 2009), the fuzzy comprehen-
sive evaluation method (Fang, 2012) and the gray correlation
analysis method (Zhang et al., 2009). While many researchers are
concerned with the assessment of soil heavy metal contamination
or soil fertility (Taylor et al., 2010; Teng et al., 2014), it is rare that
research combines two approaches to assess soil quality. This is
possibly because there is a complex nonlinear relationship
between soil heavy metal content and soil fertility and because
traditional methods do not perform well in addressing with the
complex nonlinearity. Additionally, because the weights in
assessment indices are artificially set in soil assessments, the
results obtained by traditional methods often lack reliability,
objectivity and currency (Jiang et al., 2014).
Machine learning algorithms, such as artificial neural networks
(ANNs), k-means algorithms, genetic algorithms, decision trees,
support vector machines (SVMs) and multiple linear regression
methods have been used to advance classification and forecasting
in recent years. Among these methods, ANNs have the most
complex mathematical structure and can simulate human learning
and pattern recognition (Were et al., 2015). However, due to the
lack of theoretical results from a statistical perspective, as well as
the low interpretability of this class of black-box models, some
alternative strategies have been considered (Poggi and Portier,
2011).
As a more robust statistical method, SVMs are now widely used
for environmental assessment with satisfactory performance
(Aryafar et al., 2012; Jiang et al., 2014). The SVM method is an
artificial intelligence machine learning theory introduced in the
early 1990s as a non-linear solution for regression and classifica-
tion tasks (Behzad et al., 2009; Vapnik, 1995). It relies on the
statistical learning theory or VC theory, which enables learning
machines to find important support vector information using a
very small number of parameters (Bishop, 2006; Kova9 cevic et al.,
Fig. 1. Location of research area.
2010; Wang, 2005). Many classification algorithms are based on an
independence assumption and are thus greatly influenced by the
correlation among characteristics, but SVMs are not sensitive to
this. SVMs obey the structural risk minimization principle (SRMP),
which has been shown to be superior to many other modeling
techniques obeying the traditional Empirical Risk Minimization
Principle (ERMP) (Araghinejad, 2014). This technique has been
proven to have superior performances in addressing various
problems due to its generalization abilities, robustness against
noise and interferences (Steinwart and Christmann, 2008) and its
computational efficiency compared with several other methods,
such as neural networks and fuzzy networks (Wang, 2005; Were
et al., 2015). The published literature has shown that although the
SVM method has been used to address many environmental
problems (Jiang et al., 2014; Aryafar et al., 2012), it has rarely been
used in research on soil quality assessment.
The aims of this study were (1) to develop a comprehensive
method to assess soil quality; (2) to explore the application of the
SVM method in the classification of heavy metal levels in soils; (3)
to assess soil quality by using the SVM method in combination with
the levels of soil heavy metal contamination and soil fertility; (4) to
identify the distribution of heavy metals and the pattern in soil
quality in Taiyuan city; and (5) to describe correlation among
metals and other parameters in soils.
Therefore, in this paper, we present a comprehensive soil
quality assessment model based on SVM methodology for Taiyuan
city, which could be applied in other areas when evaluating soil
heavy metal contamination and soil fertility, etc. Furthermore, this
paper provides useful information regarding soil quality and soil
management in Taiyuan.
2. Materials and methods
2.1. Area description
Taiyuan city is located on the east edge of the Loess plateau and
at the center of Shanxi Province with heavy industries (Fig. 1). It is
heavily industrialized with structurally complex factories concen-
trated in specific areas. Large volumes of waste products and
emissions are discharged to the local environment and have
resulted in the accumulation of heavy metals in soils, especially
soils in the suburbs. Previous studies have shown that the heavy
metals Hg, Cd and Pb have been enriched in the surface soil of
Taiyuan (Li et al., 2004). In addition, soil contamination in the
 Y. Liu et al. / Soil & Tillage Research 155 (2016) 19–26
southern region of the city is more serious than in the northern
region (Wang et al., 2008).
The topography of Taiyuan resembles like a dustpan. The central
and south regions of the area are comprised of the alluvial plain of
the Fen River. The south side is higher than the north and the Fen
River runs through the whole city from north to south. The total
area of the city is 6988 km2. The highest elevation is 2670 m and
the lowest elevation is 760 m. Annual rainfall in the area varies
from 317 to 395 mm per year.
2.2. Sampling procedures
The study was conducted in the suburban area of Taiyuan city,
except in the west and east of Taiyuan. The west, east and parts of
north Taiyuan are highly mountainous; therefore, these areas
could not be sampled. Based on the above consideration, the study
area was divided into 1
1 km grids and sampled following a
stochastic strategy. Soil samples of approximately 1 kg were
obtained using a stainless-steel spade from the upper 15 cm of the
soil profile depth in April 2013, and they included 140 soil samples
distributed throughout the city (Fig. 2). Soil samples were stored in
polyethylene plastic bags for transportation and storage subse-
quently air-dried at room temperature and then sieved through a
2-mm mesh sieve. Then, each soil sample was further sieved
through a 0.150-mm plastic mesh sieve to focus on metals that are
highly reactive with other molecules (Kelepertzis, 2014).
2.3. Chemical determinations
Analytical procedures were adapted from the standard operat-
ing procedures specified in the State Environmental Protection
Administration of China methodology guidelines. For analysis of As
and Hg, soil digestion was completed using a mixture of HNO3 and
HCL (aqua regia). For analysis of Cr, Cd, Cu, Zn, Ni and Pb, soil
samples were digested by adding HNO3, HCL, HF and HCLO4 in that
order. Atomic Fluorescence Spectrometry was used to quantify Hg
(GB/T 22105.1-2008) and As (GB/T 22105.2-2008) of the soil, with
method detection limits of 0.002 mg/kg and 0.01 mg/kg, respec-
tively. Determination of Cd was accomplished using graphite
Fig. 2. Sampling setting.
21
furnace atomic absorption spectroscopy (GB/T 17141-1997), with a
method detection limit of 0.09 mg/kg. Flame atomic absorption
spectroscopy was used to determine Pb, Cr, Cu, Zn, Ni content of the
soil (GB/T 17140-1997, HJ 491-2009); their method detection
limits were 0.2 mg/kg, 5 mg/kg, 1 mg/kg, 0.5 mg/kg and 5 mg/kg,
respectively.
Several additional soil parameters were also quantified in all
soil samples. Total nitrogen (TN) was determined by the Kjeldahl
method (NY/T 53-1987), available nitrogen (AN) was determined
by the alkaline hydrolysis diffusion method, available phosphorus
(AP) was determined by the Olsen method, available potassium
(AK) was determined by the ammonium acetate method (NY/T
889-2004), and organic matter (OM) content was determined by
heating potassium dichromate in an oil bath (NY/T 1121.6-2006).
Soil pH was measured using a pH meter (NY/T 1121.2-2006).
To ensure the accuracy and precision of the analysis, quality
assurance (QA) protocol was followed through the use of reagent
blanks, analytical duplicates and standard reference materials
(GBW07424, GBW07429) in the digestion and determination. The
recovery rates for metals in the standard reference materials were
approximately 96–102%.
2.4. Assessment methods
2.4.1. Assessment standard
Following the Chinese environmental quality standards for soil
(GB15618-1995), the heavy metal contents were divided into three
grades: I, II and III (Table 1). Grade I represents the average natural
background metal levels for uncontaminated soils in China, and
Grade II represents the metal levels that are hazardous to
agricultural production and human health. Grade III represents
the maximum allowable concentrations for agricultural produc-
tion and plant growth.
According to the classification of type regions and fertility of
cultivated land in China (NY/T 309-1996), the parameters
accounting for soil nutrition were divided into three levels: A, B
and C (Table 1). The conditions of soil fertility in level A are
considered to be the most fertile soils. The conditions in level B are
22 Y. Liu et al. / Soil & Tillage Research 155 (2016) 19–26

Table 1 overfitting problems. The basic idea of SVM can be summarized

Chinese environmental quality standard for soils.
as the follows.
Heavy metal items Grade I Grade II Grade III Suppose that the experimental dataset {xi,yi}, i = 1,...,l is
Cu (mg/kg) (0,35] (35,100] (100,400] composed of an instance space x 2 Rn and the label set y = {1,1}
Zn (mg/kg) (0100] (100,300] (300,500] (Lapin et al., 2014). SVM solves the binary classification problem by
Ni (mg/kg) (0,40] (40,100] (100,200]
finding a hyperplane wT ’ðxÞ þ b ¼ 0, which implements the idea of
Cr (mg/kg) (0,90] (90,350] (350,400]
Pb (mg/kg) (0,35] (35,80] (80,500]
simultaneously minimizing the empirical classification error and
Cd (mg/kg) (0,0.200] (0.200,0.800] (0.800,1] maximizing the geometric margin (Li et al., 2014b). The hyperplane
Hg (mg/kg) (0,0.150] (0.150,1] (1,1.500] is determined by solving the following convex quadratic
As (mg/kg) (0,15] (15,25] (25,40] programming:
Fertility items Level A Level B Level C
1 X l

TN (%) (0.150,+1) (0.075,0.150] (0,0.075] min k w k2 þ C ji s:t:yi ðwT wðxi Þ þ bÞ  1  ji ; ji  0; i ¼ 1; :::; l

2 i¼1
AN (mg/kg) (120,+1) (60,120] (0,60]
AP (mg/kg) (20,+1) (5,20] (0,5] where ’ðÞ is a map from the input space to a feature space, and w, b
AK (mg/kg) (150,+1) (50,150] (0,50]
and ji are the parameters that should be optimized in the training
OM (%) (3,+1) (1,3] (0,1]
phase.
The previous optimizing problem can be solved in the
Lagrangian form:
considered to be moderately fertile soils, and the conditions in
1X l X l X l
level C are considered to be the most barren soils. max  yi yj ai aj Kðxi ; xj Þ þ aj
Because the enrichment and mobility of heavy metals in soils 2 i¼1 j¼1 j¼1
may cause serious harm to humans (Franco-Uría et al., 2009), we
X
l
assign priority to the heavy metal content of soil when assessing s:t: yi ai ¼ 0; 0  ai  C; i ¼ 1; :::; l
overall soil quality. In addition, we consider that the higher i¼1
nutrient content in soil is, the better the soil quality is. Based on where ai are the Lagrange multipliers and K(xi, xj) = <’(xi), ’(xj)> is
these assumptions, the quality of Taiyuan city soil were recorded as the kernel function.
either IA, IB, IC, IIA, IIB, IIC, IIIA, IIIB and IIIC. IA indicates the best The optimal boundary is then determined by the support vector
soil quality, whereas IIIC refers to the worst soil quality. expansion:
!
2.4.2. SVM method Xl

As one of the most best binary classifier methods, SVM is f ðxÞ ¼ sgn ai yi Kðxi ; xÞ þ b


i¼1
applied in a wide variety of fields (Hastie et al., 2004; Li et al., 2014
b). It was first introduced by Vapnik (1995); it is based on the where ai is the support vector and b* is the bias, which can be
structural risk minimization principle and can overcome calculated from the Karush–Kuhn–Tucker conditions.

Soil Heavy Metal Chinese standard: I, II, III

Interpolation

Soil Heavy Metal experiment samples (600)

Training samples Test samples

(480) (120) Soil Fertility Chinese standard: A, B, C
Optimization Verification Interpolation
SVM model Soil Fertility experiment samples (600)

Soil Samples Input A trained SVM for SHM

of Taiyuann classification (SHM-SVM) Training samples Test samples
(140) (480) (120)
Output
Optimization Verification
Soil grade Soil grade Soil grade
SVM model
I II III

Input Input Input

A trained SVM for SF classification (SF-SVM)

Output

Soil class Soil class Soil class Soil class Soil class Soil class Soil class Soil class Soil class
IA IB IC IIA IIB IIC IIIA IIIB IIIC

Fig. 3. The technical flowchart of soil quality comprehensive assessment model.

The number in parentheses () refers to the number of samples.
 Y. Liu et al. / Soil & Tillage Research 155 (2016) 19–26 23

2.4.3. Algorithm and technical flowchart
The algorithm of soil quality comprehensive assessment model:
The comprehensive soil quality classification methodology using
heavy metal and fertility based on SVM (SQ-SVM).
Input: Chinese standard of soil heavy metals (SHM–Chinese-
standard), Chinese standard of soil fertility (SF-Chinese-standard)
and the detected concentrations of heavy metals and soil fertilities
of 140 soil-samples in Taiyuan.
Output: Soil quality comprehensive assessment/classification,
nine classes (IA, IB, IC, IIA, IIB, IIC, IIIA, IIIB and IIIC).
Step 1: Construct the SHM (soil heavy metal) experimental
dataset fðxi ; yi Þgpi¼1 , where xi 2 Rn ; yi 2 ½1; 2; 3; p ¼ 600 is the
number of the experiment data, by interpolating the SHM–
Chinese-standard.
Step 2: Select 80% SHM data as training sets randomly and the
remainder are considered as the test sets.
Step 3: Train the SVM using the training dataset and verify the
validity of the model using the test dataset. The result is a trained
SVM model for SHM classification (SHM–SVM).
Step 4: Input the measured heavy metal concentrations in the
140 samples of Taiyuan soil into the previously trained SHM–SVM
model and obtain the classification results for 140 soil samples
(grade I, grade II and grade III).
Step 5: Construct the SF (soil fertility) experimental dataset
fðxi ; yi Þgqi¼1 , where xi 2 Rm ; yi 2 ½A; B; C; q ¼ 600 is the number of
the experimental data, by interpolating the SF-Chinese-standard.
Step 6: Randomly select 80% of the SF data as the training sets
and the remainder are considered as the test data sets.
Step 7: Train and verify the SVM using the SF experimental data.
The outcome is a trained SVM model for SF classification (SF-SVM).
Step 8: Input the measured soil fertility data for the soil samples
respectively into the trained SF-SVM model, and it will produce the
classification results (i.e., IA, IB, IC, IIA, IIB, IIC, IIIA, IIIB and IIIC).
The technical flowchart of the algorithm is presented in Fig. 3.
The realization of the algorithm is performed with Matlab 2011b.
3. Results
3.1. Descriptive statistical analyses
Descriptive statistical analyses were performed with R version
3.0.2. The descriptive statistics for the 140 soil samples are listed in
Table 2. The coefficient of variation (C.V) value for soil pH was
0.022, which indicated very low variation. Soil pH has often been
found to be less variable than other soil properties (Bai and Wang,
2011; Liu et al., 2013). The C.V for soil Hg and AP were 1.085 and
1.124, respectively, indicating high variation (Nielsen and Bouma,
1985). High variation in AP may result from various planting
patterns and the amount of fertilizer used in different areas. The
high variation in Hg concentrations may be the result of different
sources of contamination. Comparing the heavy metal concen-
trations of the Taiyuan soils with the Chinese standards (GB15618-
1995) and the background values for Taiyuan city, the mean
concentration of each of the heavy metals in each sample was
below CS I (Chinese standard I), except for Cd, which showed only a
light level of metal pollution in the soil environment within the
research area.
3.2. Correlation analyses
Correlation analysis was performed using R 3.0.2. The correla-
tion matrix of soil variables based on Pearson correlation
coefficients is presented in Fig. 4. The significance levels were
0.01 (**) and 0.05(*). The correlations clearly showed the relevance
of all of the measured soil parameters. The pie charts with blue and
positive rotation expressed positive correlations, whereas the pie
charts with red and reverse rotation expressed negative correla-
tions. Furthermore, the deeper the color is, the stronger the
correlation. Correlation coefficients can be found in the block
diagram in the lower left corner of the figure.
From Fig. 4, it can be seen that soil Ph was significantly
negatively correlated with most elements in the soils. On the other
hand, soil organic matter (OM) was significantly positively
correlated with all nutrients and all heavy metals, except for Ni
and As. Organic matter not only determines soil productivity but
also influences heavy metal adsorption and mobility in soil (Li,
2004). The concentrations of all metals were positively correlated
except for As, which was positively or negatively correlated with
the other metals. In addition, As was significantly negatively
correlated with all nutrients and Hg. Ni was significantly negatively
correlated with OM and TN. Cd and Pb were significantly positive
correlated with all of the nutrients. Because there are important
correlations between soil heavy metal contaminations and soil
nutrient status, comprehensive assessments of soil quality should
consider both of these assessment categories.
3.3. SVM-based classification model for heavy metal contamination
assessment
According to the SHM–SVM algorithm and technical flowchart,
the analysis of soil heavy metal content was performed in Matlab.
The kernel function chosen for the SHM–SVM was the Gaussian
kernel function, based on the research of Dibike et al. (2001) and

Table 2
Descriptive statistics of soil elements.

Items Total Mean SD C.V Maximum Minimum CS I CS II BGV

pH 140 7.754 0.169 0.022 8.180 7.080 – – –
Total N(%) 140 0.121 0.058 0.479 0.445 0.026 – – –
O.M.(%) 140 2.199 1.959 0.891 13.420 0.140 – – –
AN(mg/kg) 140 97.749 35.686 0.365 238.480 32.310 – – –
AP(mg/kg) 140 19.236 21.630 1.124 166.710 2.530 – – –
AK(mg/kg) 140 172.314 69.751 0.405 430.000 55.000 – – –
Cu(mg/kg) 140 29.512 9.703 0.329 74.820 15.280 35 100 22.900
Zn(mg/kg) 140 89.187 24.163 0.271 179.060 52.870 100 300 63.500
Ni(mg/kg) 140 29.864 4.635 0.155 47.420 16.840 40 100 29.900
Cr(mg/kg) 140 74.457 20.528 0.276 202.000 39.980 90 250 55.300
Pb(mg/kg) 140 27.310 8.469 0.310 68.430 16.560 35 80 14.700
Cd(mg/kg) 140 0.244 0.138 0.566 0.924 0.040 0.2 0.8 0.102
Hg(mg/kg) 140 0.094 0.102 1.085 0.801 0.012 0.15 1.5 0.023
As(mg/kg) 140 10.725 2.332 0.217 18.080 4.760 15 25 9.100

SD = Standard deviation; C.V = Coefficient of variation; CS I = The average background value of soil heavy metals in China (GB15618-1995); CS II = The value of soil heavy metals
in China for protecting agricultural production and human health (GB15618-1995); BGV = Background value of soil heavy metals in Taiyuan.
24 Y. Liu et al. / Soil & Tillage Research 155 (2016) 19–26

Fig. 4. The correlation matrix of soil variables. (For interpretation of the references to color in the text, the reader is referred to the web version of this article.)
*P < 0.05, **P < 0.01

Fig. 5. Parameters option contour map of SHM–SVM.

Han and Cluckie (2004) who indicated the Gaussian radial basis
function has superior efficiency compared to other Kernel
functions. A cross validation method was used to optimize the
model by searching for better parameters. Fig. 5 shows the
parameter option contour map from the SHM–SVM. The best cost Fig. 7. The spatial distribution of heavy metal grades in Taiyuan.
parameter (C) was 0.00097656, and the optimal parameter of the
kernel function (g) was 2. The SHM–SVM had an accuracy of grade I, and 17.86% were classified as grade II. This suggests that soil
98.5417% in classifying the degree of soil heavy metal contamina- heavy metal contamination was not serious in Taiyuan.
tion in Taiyuan. The results also showed that the samples were
divided unevenly into the two grades, with 115 in grade I and 25 in
grade II. Fig. 6 shows 82.14% of the soil samples were classified as

Fig. 6. Percent of soil samples in each soil heavy metal grade. Fig. 8. Parameter option contour map from SQ-SVM.
 Y. Liu et al. / Soil & Tillage Research 155 (2016) 19–26
Fig. 9. Percent of soil samples in each soil quality classification.
The spatial distribution of soil heavy metal grades in Taiyuan is
shown in Fig. 7. As shown, the Grade I soils were distributed in all
Taiyuan districts. Grade II soils, which indicate a relatively high
degree of contamination were mainly distributed in the southern
part of Taiyuan, which corresponds to the sewage irrigation zone.
3.4. SVM-based classification model for soil quality assessment
Based on the classification of soil heavy metals content, the
assessment of soil quality was performed using the SQ-SVM model.
Fig. 8 shows the parameter option contour map from SQ-SVM. The
best cost parameter (C) was 0.70711, and the optimal parameter of
the kernel function (g) was 2. The accuracy of the SQ-SVM was
98.3333%. The results of the soil quality classification for Taiyuan
based on SQ-SVM are shown in Fig. 9.
From the assessment of soil heavy metal content described
above, 115 samples were classified as grade I. Combined with the
soil fertility levels, these 115 samples were further divided into
three classes, IA, IB and IC, representing 35, 70 and 10 sites,
respectively. For the 25 samples classified as grade II, the SQ-SVM
assessment showed that the samples were divided into two
classes, IIA and IIB, with 14 and 11 sites, respectively. Overall, 25%
of soil samples were classified as class IA. No sample was classified
as grade III. Most soil samples were classified as class IB,
accounting for 50% of the total samples.
Fig. 10. The spatial distribution of soil quality classes in Taiyuan.
25
The spatial distribution of soil quality classes in Taiyuan is
shown in Fig. 10. Generally the soil quality in the north of Taiyuan is
better than that in the south. Most soil samples were classified as
class IB, except in the middle east of Taiyuan. The worst soil quality
in Taiyuan (IIB) occurs in the southeastern part of the city.
4. Discussion and conclusions
This paper presented an innovative SVM-based classification
model that linked the soil characteristics and our understanding of
their meaning to quantify soil quality in a relatively scientific way.
The SHM–SVM model, when applied to 140 soil samples collected
from across Taiyuan city, with an accuracy of 98.5417%, showed
that 115 samples belonging to grade I, accounting for 82.14% of the
total number of samples; the remaining 25 samples were classified
as grade II. Overall, these results suggest that heavy metal
contamination in Taiyuan soils is not serious, but it is higher than
the background values found in many locations, mainly in the
southern part of Taiyuan, which correspond to the sewage
irrigation zone. The sewage irrigation system may be the main
reason leading to the relatively serious soil contamination of
farmland. Therefore, the authorities should immediately pay more
attention to the sewage treatment technology and the improve-
ment of the irrigation system.
The SQ-SVM model, with an accuracy of 98.3333% showed that
35 samples belonged to soil standard IA, accounting for 25% of the
total number of samples. No sample corresponded with standard
IIIC. Most soil samples were classified as class IB, accounting for
50% of all samples. In general, the soil quality in the north of
Taiyuan is better than that in the south. Numerous heavy industrial
factories are located in the southwestern section of the city, which
may account for the increase in soil heavy metal contamination
and decrease in soil fertility, thus making the soil quality relatively
poorer in this area.
The overall goal of this paper was to develop a SVM-based
classification model that combined soil heavy metal contamina-
tion and soil fertility data together in one model to better assess
urban soil quality. The general classification is based on the
Chinese soil quality standards for heavy metals and the published
soil background concentrations of the research area. Undoubtedly,
there are various factors that influence soil quality, including heavy
metal and organic chemical contamination, soil fertility, soil
texture and soil biodiversity. Thus, a SVM-based classification
model could be improved by considering the various soil influence
factors mentioned above.
This study has highlighted several advantages of using
comprehensive SVM-based classification models, including the
following three examples. (1) The SVM-based model structure is
based on statistical learning theory and VC theory. The VC theory
can generalize unseen data; thus, it can learn well with only a very
small number of parameters. (2) Many classification algorithms are
based on the independence assumption and are thus greatly
influenced by the characteristics of the correlation, but they are not
sensitive for the feature correlation. (3) Its computational
efficiency is higher than several other methods.
Of course, the SVM-based model also has some aspects to
carefully consider, as follows. (1) There is no universal solution for
nonlinear problems; therefore, the Kernel function must be
carefully chosen, which may affect the true assessment of soil
quality. (2) Different assessment models group soils based on
diverse national standards and native background values. It is
therefore necessary to select the soil classification standards
cautiously because this operation establishes a link between soil
characteristics and our understanding of their meaning in a wider
context (Montanarella and Rusco, 2008). (3) The experimental
datasets are interpolated by soil standards, and the datasets do not
26 Y. Liu et al. / Soil & Tillage Research 155 (2016) 19–26
include those extreme outliers. One of the main drawbacks of
SVMs is their sensitivity to outliers or noise in the training sample
due to overfitting. Thus, further studies regarding the application
of fuzzy SVMs are proposed to address this issue.
The presented SVM-based classification model is not only valid
for Taiyuan but can also be applied to other areas in China. A key
factor in the application of the model is to understand the
relationship among soil environmental quality, soil productivity
and soil sustainability. Furthermore, the comprehensive SVM-
based classification framework and algorithm is expected to have
wide utilization potential, not only in soil quality assessment but
also in other environmental fields when well designed.
Acknowledgements
This work was financially supported by the National Natural
Science Foundation of China under Grant 41271513 and by the
Important Specialized Science and Technology Item of Shanxi
Province, China under Grant 20121101011.
References
Akbulut, S., Grieken, R., Kılıc, M.A., Cevik, U., Rotondo, G.G., 2013. Identification of
heavy metal origins related to chemical and morphological soil properties using
several non-destructive X-ray analytical methods. Environ. Monit. Assess. 185
(3), 2377–2394.
Araghinejad, S., 2014. Support Vector Machines. Data-Driven Modeling: Using
MATLAB03 in Water Resources and Environmental Engineering. Springer,
Netherlands.
Aryafar, A., Gholami, R., Rooki, R., Ardejani, F.D., 2012. Heavy metal pollution
assessment using support vector machine in the Shur River, Sarcheshmeh
copper mine, Iran. Environ. Earth Sci. 67 (4), 1191–1199.
Bai, Y., Wang, Y., 2011. Spatial variability of soil chemical properties in a jujube slope
on the Loess Plateau of China. Soil Sci. 176 (10), 550–558.
Behzad, M., Asghari, K., Eazi, M., Palhang, M., 2009. Generalization performance of
support vector machines and neural networks in runoff modeling. Expert Syst.
Appl. 36 (4), 7624–7629.
Bhaduri, D., Purakayastha, T.J., 2014. Long-term tillage, water and nutrient
management in rice-wheat cropping system: assessment and response of soil
quality. Soil Tillage Res. 144, 83–95.
Bhuiyan, M.A., Parvez, L., Islam, M.A., Dampare, S.B., Suzuki, S., 2010. Heavy metal
pollution of coal mine-affected agricultural soils in the northern part of
Bangladesh. J. Hazard. Mater. 173 (1), 384–392.
Bishop, C.M., 2006. Pattern Recognition and Machine Learning, vol. 4. Springer, New
York No. 4, p. 12.
Cachada, A., Pereira, M.E., Ferreira, d.S.E., Duarte, A.C., 2012. Sources of potentially
toxic elements and organic pollutants in an urban area subjected to an
industrial impact. Environ. Monit. Assess. 184 (1), 15–32.
Chen, F., Jiang, X., Tang, F., Bian, Y., 2012. Application of AHP and GIS in evaluation of
agricultural soil heavy metals pollution. Environ. Pollut. Control 7 004.
Dibike, Y.B., Velickov, S., Solomatine, D., Abbott, M.B., 2001. Model induction with
support vector machines: introduction and applications. J. Comput. Civil Eng. 15
(3), 208–216.
Fang, R.H., 2012. Fuzzy synthetical evaluation of soil fertility in the area of
Guanzhong Plain and the Loess Plateau—a case study in Chang’an District of
Xi’an. Agric. Res. Arid Areas 1, 006 (in Chinese).
Franco-Uría, A., López-Mateo, C., Roca, E., Fernández-Marcos, M.L., 2009. Source
identification of heavy metals in pastureland by multivariate analysis in NW
Spain. J. Hazard. Mater. 165 (1), 1008–1015.
Fu, S.M., Xiao, F., Su, W.J., Qiu, J.Q., Wang, D.F., Chang, X.Y., 2014. The evaluation of
heavy metals pollution in soils of the lower reaches of the Hengshi River within
the Dabaoshan mining area based on fuzzy mathematics. Geol. Bull. China 33
(8), 1140–1146.
Guo, G., Fengchang, Wu, Xie, F., Zhang, R., 2012. Spatial distribution and pollution
assessment of heavy metals in urban soils from southwest china. J. Environ. Sci.
(China) 24 (3), 410–418.
Han, D.A.W.E.I., Cluckie, I., 2004. Support vector machines identification for runoff
modeling. In Proceedings of the sixth international conference on
hydroinformatics 21–24.
Hastie, T., Rosset, S., Tibshirani, R., Zhu, J., 2004. The entire regularization path for
the support vector machine. J. Mach. Learn. Res. 5, 1391–1415.
Hui, X., 2013. Analysis of heavy metal pollution in soil by fuzzy comprehensive
evaluation method. Comput. Eng. Appl..
Jiang, X., Lu, W.X., Yang, Q.C., Zhao, H.Q., 2014. Application of support vector
machine in soil environmental quality assessment. China Environ. Sci. 34 (5),
1229–1235 (in Chinese).
Karim, Z., Qureshi, B.A., Mumtaz, M., Qureshi, S., 2014. Heavy metal content in urban
soils as an indicator of anthropogenic and natural influences on landscape of
Karachi—a multivariate spatio-temporal analysis. Ecol. Indic. 42, 20–31.
Kelepertzis, E., 2014. Accumulation of heavy metals in agricultural soils of
Mediterranean: insights from Argolida basin, Peloponnese, Greece. Geoderma
221, 82–90.
Kova9 cevi
c, M., Bajat, B., Gaji
c, B., 2010. Soil type classification and estimation of soil
properties using support vector machines. Geoderma 154 (3), 340–347.
Lapin, M., Hein, M., Schiele, B., 2014. Learning using privileged information: SVM+
and weighted SVM. Neural Netw. 53, 95–108.
Li, D.S., Yang, Z.F., Jin, Z.B., 2004. Geochemical characters of trace element of soil
from the Taiyuan basin. Geol. Prospect. 40 (3), 86–89 (in Chinese).
Li, Z., Ma, Z., van der Kuijp, T.J., Yuan, Z., Huang, L., 2014a. A review of soil heavy metal
pollution from mines in China: pollution and health risk assessment. Sci. Total
Environ. 468, 843–853.
Li, Z., Zhou, M., Xu, L., Lin, H., Pu, H., 2014b. Training sparse SVM on the core sets of
fitting-planes. Neurocomputing 130, 20–27.
Li, Y., 2004. Effect of pH and organic matter on the bioavailability Cd and Zn in soil. J.
Yunnan Agric. Univ. 20 (4), 539–543 (in Chinese).
Liao, Y.C., Chien, S.C., Wang, M.C., Shen, Y., Hung, P.L., Das, B., 2006. Effect of
transpiration on Pb uptake by lettuce and on water soluble low molecular
weight organic acids in rhizosphere. Chemosphere 65 (2), 343–351.
Liu, Z.P., Shao, M.A., Wang, Y.Q., 2013. Large-scale spatial interpolation of soil pH
across the Loess Plateau, China. Environ. Earth Sci. 69 (8), 2731–2741.
Montanarella, L., Rusco, E., 2008. Threats to soil quality in Europe.
Nielsen, D.R., Bouma, J., 1985. Soil spatial variability.
Poggi, J.M., Portier, B., 2011. PM10 forecasting using clusterwise regression. Atmos.
Environ. 45 (38), 7005–7014.
Praveena, S.M., Yuswir, N.S., Aris, A.Z., Hashim, Z., 2014. Contamination assessment
and potential human health risks of heavy metals in klang urban soils: a
preliminary study. Environ. Earth Sci..
Saby, N.P.A., Thioulouse, J., Jolivet, C.C., Ratié, C., Boulonne, L., Bispo, A., Arrouays, D.,
2009. Multivariate analysis of the spatial patterns of 8 trace elements using the
French soil monitoring network data. Sci. Total Environ. 407 (21), 5644–5652.
Steinwart, I., Christmann, A., 2008. Support Vector Machines. Springer Science &
Business Media.
Taylor, M.D., Kim, N.D., Hill, R.B., Chapman, R., 2010. A review of soil quality
indicators and five key issues after 12 yr soil quality monitoring in the Waikato
region. Soil Use Manage. 26 (3), 212–224.
Teng, Y., Wu, J., Lu, S., Wang, Y., Jiao, X., Song, L., 2014. Soil and soil environmental
quality monitoring in China: a review. Environ. Int. 69, 177–199.
Vapnik, V.N., 1995. The Nature of Statistical Learning Theory Berlin.
Wang, X.J., Lai, J.Q., Lu, Y.H., Li, D.S., Zhou, J.H., Wang, J.W., 2008. Main source of soil
heavy metal pollution based on factor analysis in Taiyuan. Ecol. Environ. 17 (2),
671–676 (in Chinese).
Wang, L., 2005. Support Vector Machines: Theory and Applications. Nanyang
Technological University, School of Electrical & Electronic Engineering.
Were, K., Bui, D.T., Dick, Øystein B., Singh, B.R., 2015. A comparative assessment of
support vector regression, artificial neural networks, and random forests for
predicting and mapping soil organic carbon stocks across an afromontane
landscape. Ecol. Indic. 52, 394–403.
Xu, F.Y., Hu, Y.Y., 2014. Distribution and pollution assessment on heavy metals in
urban soils of different functional areas in Chongqing, Chinese. J. Soil Sci. 45 (1),
227–231 (in Chinese).
Yu, H., Ni, S.J., He, Z.W., Zhang, C.J., Nan, X., Kong, B., et al., 2014. Analysis of the
spatial relationship between heavy metals in soil and human activities based on
landscape geochemical interpretation. J. Geochem. Explor. 146, 136–148.
Zhang, L., Fang-Li, S.U., Guo, C.J., Hong, Y.L., Song, Y.L., 2009. Application of grey
correlation analysis in different models of ecological restoration in soil quality
evaluation. J. Shenyang Agric. Univ. (in Chinese).
Zheng, L., Wu, W., Wei, Y., Hu, K., 2015. Effects of straw return and regional factors on
spatio-temporal variability of soil organic matter in a high-yielding area of
northern china. Soil Tillage Res. 145, 78–86.
Zhou, X., An, Y.L., Xu, W.C., Deng, M.L., 2009. Fuzzy evaluation on soil fertility of
cultivated land based on GIS and improved AHP—a Case of Puan County in
Guizhou Province [J]. Chin. J. Soil Sci. 1, 020.