Science of the Total Environment 592 (2017) 704713

Contents lists available at ScienceDirect

Science of the Total Environment

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

Prediction of soil organic carbon in an intensively managed reclamation

zone of eastern China: A comparison of multiple linear regressions and
the random forest model
Huan Zhang a, Pengbao Wu a, Aijing Yin b, Xiaohui Yang a, Ming Zhang c, Chao Gao a,
a
School of Geographic and Oceanographic Sciences, Nanjing University, Nanjing 210023, China
b
Institute of Agricultural Resources and Environment, Jiangsu Academy of Agricultural Sciences, Nanjing 210014, China
c
Nanjing Center, China Geological Survey, Nanjing 210000, China

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

SOC was predicted using multiple linear

regressions (MLR) and random forest
model (RF).
Soil/sediment pH, Cl and silt contents
were identied as the most important
factors impacting SOC.
RF preformed much better than MLR in
predicting SOC.
Paddy management was recommended
as an environment-friendly approach to
managing newly reclaimed lands.

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

Article history: Organic carbon is a key component of soils and plays a fundamental role in soil fertility and climate change. De-
Received 12 October 2016 termining the importance of potential drivers of soil organic carbon (SOC) and thus predicting the distribution of
Received in revised form 17 February 2017 SOC are important for measuring carbon sequestration or emissions. Coastal wetlands are precious land resources
Accepted 17 February 2017
that are currently undergoing rapid reclamation in China. The alternations in soil physicochemical conditions
Available online 22 March 2017
caused by reclamation can strongly impact the cycle of organic carbon. However, identication of the important
Editor: D. Barcelo drivers of SOC dynamics and prediction of SOC using the potential drivers remain largely unclear. In this study,
we used classication and regression tree (CART) to identify the importance of the potential drivers of SOC at
Keywords: 241 sites from an intensively managed reclamation zone of eastern China. Multiple linear regressions (MLR)
Soil organic carbon and random forest (RF) models were applied to predict the distribution of SOC using continuous variables,
Random Forest such as the contents of Cl, CaO, Fe2O3, Al2O3, SiO2, clay, silt, and sand as well as the soil pH, along with categorical
Factor importance variables, such as land use and reclamation duration. The results indicate that the soil/sediment pH was the most
Reclaimed soils important variable impacting SOC, followed by the Cl and silt contents. The RF and MLR involving all predictor
variables produced much higher R2 and lower error indices than the RF and MLR models involving independent
variables (pH and CaO). RF performed much better than MLR as it revealed much lower error indices (ME, MSE,
and RMSE) and a higher R2 than MLR. The superiority of RF in predicting SOC is related to its capability to deal
with non-linear and hierarchical relationships between SOC and predictors. Analyses of land use effects on SOC
dynamics indicated that paddy soils were superior in sequestering SOC than other land use types, which is likely

Corresponding author.
E-mail address: chgao@nju.edu.cn (C. Gao).

http://dx.doi.org/10.1016/j.scitotenv.2017.02.146
0048-9697/ 2017 Elsevier B.V. All rights reserved.
 H. Zhang et al. / Science of the Total Environment 592 (2017) 704713
ascribed to the rapid desalination and dealkalization of paddy eld management. Therefore, paddy eld manage-
ment is recommended as an environment-friendly approach for managing newly reclaimed lands.
1. Introduction
Soil is the largest organic carbon pool in terrestrial ecosystems
and stores two or threefold more organic carbon than either the at-
mosphere or terrestrial vegetation (Lal, 2004; Schmidt et al., 2011).
Soil organic carbon (SOC) thus constitutes a major global change
driver due to its high sensitivity to human disturbance (Smith et
al., 2015). To mitigate increasing global warming, soils should be
managed to sequester more organic carbon and discharge less green-
house gases (Wiesmeier et al., 2015; Zhang et al., 2016). SOC is also
closely related to soil fertility and productivity due to its signicant
role in supplying nutrients and in the formation of improved soil
structure (Tiessen et al., 1994). A further understanding of SOC dy-
namics and related driving factors would be essential for predicting
greenhouse emissions and for pursuing sustainable land use strate-
gies (Hu et al., 2007; Wiesmeier et al., 2012).
Coastal wetlands are precious land resources and play a signi-
cant role in carbon sequestration (Chmura et al., 2003). Coastal wet-
lands reclamation has been a common practice for gaining new
arable land in China and in other countries for thousands of years
(Cheng et al., 2009; Fernndez et al., 2010; Wolff, 1992). In China,
N50% of the coastal wetlands have been reclaimed since the 1950s
due to increasing population pressures and related socio-economic
demands (An et al., 2007), resulting in 13,380 km2 of land reclama-
tion between 1950 and 2008 (Wang et al., 2014b). Reclamation of
coastal wetlands for agricultural uses may have a notable effect on
SOC because of the alternation in soil conditions and OC inputs and
outputs (Deng et al., 2016; Jin et al., 2013; Yin et al., 2016). Numer-
ous studies have revealed signicant differences in SOC between
reclaimed lands and natural coastal wetlands (Cui et al., 2012;
Deng et al., 2016; Klbl et al., 2014).
Climate conditions, soil properties, topography, and anthropogenic
activities are general factors driving SOC dynamics, as stated in previous
studies (Doetterl et al., 2015; Hobley et al., 2015; Six et al., 2000). How-
ever, SOC dynamics in coastal reclaimed lands is not only affected by the
above factors but also by certain specic factors e.g., salinity, pH and rec-
lamation duration (Klbl et al., 2014; Wang et al., 2014a; Wissing et al.,
2011). Generally, SOC tends to accumulate in areas with lower temper-
atures and higher precipitation (Chris et al., 2004; Hobley et al., 2015).
Soils with ner particles are favorable for SOC accumulation because
of the stronger adsorption ability of soil clay and silt to SOC (Hassink,
1997). Anthropogenic factors, such as land use change and agricultural
practices, also exert signicant impacts on SOC dynamics by altering
OC inputs, the OC decomposition rate and dissolved organic carbon
leaching (Cambardella et al., 1994; DeGryze et al., 2004; Richter et al.,
1999). On the salt-affected reclaimed lands, salinity plays an opposite
role in SOC dynamics, by limiting plant growth, which would result in
low OC input, reduced OC decomposition rates and consequently SOC
accumulation when the OC input is constant (Setia et al., 2013). The rel-
ative alkalinity of newly reclaimed soil is also a primary limiting factor
for plant growth, thus impacting SOC dynamics (Deng et al., 2016).
After reclamation, SOC increased signicantly with increased reclama-
tion duration, which can probably be attributed to the result of fertiliza-
tion and the accumulation of organic residues (Fu et al., 2014; Li et al.,
2014). Some studies have also highlighted that reclaimed lands used
as paddy elds have much higher SOC content than those used as up-
lands (Cheng et al., 2009; Cui et al., 2012; Fu et al., 2014).
Numerous models have been developed to predict the spatial distri-
bution of SOC using the potential factors related to SOC dynamics.
Among those models, multiple linear regression has been used as a
705
2017 Elsevier B.V. All rights reserved.
most common approach for SOC prediction (Florinsky et al., 2002;
Guo et al., 2015; Thompson and Kolka, 2005). The preference for this
method can be linked to its simplicity in application, efciency in com-
putation, and ease of interpretation (Thompson et al., 2006). This meth-
od can only detect liner relationships between the response variable
and environmental predictors. However, these relationships are gener-
ally complex and non-linear (Lark, 1999). Therefore, some studies have
used classication and regression tree (CART) model (Hobley et al.,
2015; Ru et al., 2016), generalized linear model (Mckenzie and Ryan,
1999), and articial neural networks (Alvarez et al., 2011; Guo et al.,
2013) to detect the non-linear relationships between target variables
and environmental predictors. The CART method is a widely used ma-
chine learning method that has proven its superiority in handling
non-linear and non-additive relationships (Myles et al., 2004). Random
forest (RF), developed as an extension of CART, is a relatively new ap-
proach to improve prediction accuracy (Breiman, 2001). This method
has some advantages over other prediction models, such as resistance
to overtting, insensitivity to noise features and an implemented unbi-
ased measure of error rate (Breiman, 2001). To date, some studies have
used RF to link SOC to environmental variables (Grimm et al., 2008; Guo
et al., 2015; Rossel and Behrens, 2010; Wiesmeier et al., 2011). Howev-
er, on reclaimed lands, RF has only been applied once to the prediction
of SOC (Deng et al., 2016). Considering the rapid dynamics of soil prop-
erties that might affect SOC, accurate prediction of SOC would be essen-
tial for improving understanding of SOC dynamics and its related
controlling factors on reclaimed lands.
The objectives of this study were as follows: 1) to predict SOC using
MLR and RF models and to compare the performances of these two
models; and 2) to identify the importance of the factors impacting
SOC in an intensively managed reclamation zone.
2. Materials and methods
2.1. Study area
The study area is located in Dafeng County, Jiangsu Province,
China (3300N3330N, 12030E12100E). This study area
covers approximately 1700 km2, with an elevation of several meters
above sea level. The area has a subtropical monsoon climate, with a
mean annual temperature and rainfall of 14 C and 1042 mm, respec-
tively. The area is a marine deposit plain, with large portions of the
deposited sediments derived from nearby rivers. The Yangtze River
is one of the transporters that carries abundant terrestrial material
into the Yellow Sea, and the material re-accumulates in the coastal
area of Dafeng. Material from the old Yellow River also contributed
considerable amounts of sediment to the coastal area of Dafeng be-
tween 1495 and 1851, until its channel was altered northward
(Zhang et al., 2013). In the past century, the tidal ats in the study
area experienced intensive reclamation due to a high sediment ac-
cretion rate and rapid expansion of coastal wetlands. According to
IUSS Working Group WRB (2007), the soil type of the lands
reclaimed within 50 years was classied as Stagnic Gleyic Cambisol
(Galcaric Siltic); whereas the lands with N50 years of reclamation
history was identied as Gleyic Cambisol (Eutric, Siltic).
Dike building has been a common method to separate coastal
wetlands and reclaimed lands in this area. The land inside a certain
dike was reclaimed at approximately the identical time. In this
study area, a total of four dikes have been built during the past cen-
tury, thereby resulting in a chronosequence with reclamation dura-
tions of approximately 10, 30, 65, and 90 years. Dike-building
706 H. Zhang et al. / Science of the Total Environment 592 (2017) 704713
dating was obtained from the interpretation of satellite images and
reviewing local chronicles. The coastal wetlands (unreclaimed
land) were 23 km in width and were sparsely covered with Spartina
alterniora, Suaeda salsa and Phragmites australis. After the build-up
of dikes, the enclosed land (10-year-old land) was immediately
used for shponds, with a primary purpose of rapid desalination.
Next, the lands with N 10-years of reclamation duration were inten-
sively used for farming. Major crop rotation on 30- and 65-year-old
reclaimed land involved rice (Oryza sativa)-wheat (Triticum
aestivum) rotation, with a small proportion of other rotations such
as wheat-maize (Zea mays) and maize-garlic (Allium sativum).
Major crop rotations on 90-year-old land involved wheat-maize,
wheat-cotton (Gossypium hirsutum) and maize-rape (Brassica
campestris).
2.2. Sampling methods and chemical analyses
A total of 241 topsoil and sediment (020 cm) samples were collect-
ed in Dafeng in the summer of 2012 (Fig. 1). At each site, ve subsam-
ples within a 10 m radius were collected and merged into one
composite sample. All of the samples were air-dried and gently crushed
to pass through a 2-mm sieve.
The soil/sediment pH was determined with a glass electrode at a
ratio of 1:2.5 soil:water. The clay, silt and sand contents were deter-
mined using a sedimentation method. The elemental contents of Ca,
Cl, Fe, Al and Si were determined using X-ray uorescence spectrom-
etry (XFS) after 5 g of each air-dried sample was ground to b200
mesh in an agate mortar and composed into a disc. Before determi-
nation of SOC, the stored soil/sediment samples were ground to
pass through a 0.154-mm sieve. The SOC was determined with a
wet-combustion method. Detailed procedures for these analyses of
soil/sediment properties are described by Lu (2000), Bao (2000)
and Klbl et al. (2014).
Fig. 1. Location of the study area in Dafeng County and the spatial distribution of the sampling sites.
2.3. Calculation of soil bulk density, SOC density and topographic variables
Bulk density (BD) of soil/sediment was calculated as a function of
soil organic matter (OM) and soil texture according to Rawls (1983):
100
BD   1
OM 100OM
BDOM BDMM
where OM is the content of organic matter (%, SOC content 1.724),
BDOM is the BD of the OM (0.244 g cm3) and BDMM is the BD of the
mineral matter (g cm3). BDMM was calculated with Eq. (2) by estimat-
ing BDMM as a function of sand content and depth according to Tranter
et al. (2007):
BDMM 1:35 0:0045  S 6  105 44:7S2 0:06  logH 2
where S is the sand content (%) and H is the depth of soil/sediment
(20 cm in this study).
Then, soil/sediment organic carbon density (SOCD) of the upper
20 cm layer was calculated using Eq. (3):
SOCD SOC  BD  H  1C  101 3
where SOCD and SOC are the density (t ha1) and content (g kg1) of
soil organic carbon, respectively. BD is the soil bulk density (g cm3),
H is the thickness (cm), and C is the fraction (%) of N2-mm in soil.
Three primary and secondary terrain attributes were derived from a
90 m resolution Digital Elevation Model (SRTM; Jarvis et al., 2008). Pri-
mary terrain attributes include elevation and slope. Topographic wet-
ness index (TWI) is a secondary terrain attribute calculated from
primary attributes (Wilson and Gallant, 2000).
 H. Zhang et al. / Science of the Total Environment 592 (2017) 704713 707

2.4. Predication models
In this study, MLR and RF were employed to predict SOC contents,
whereas CART was applied to identify the main factors controlling
SOC and to gain insight into the tree-splitting process.
MLR is a classical approach that has been largely applied for
predicting the values of a dependent variable from predictor variables.
It explores how the dependent and independent variables are correlat-
ed. In this study, MLR was implemented using SPSS 18.0 for windows.
CART is a non-parametric data mining technique that can reect
non-linear and non-additive relationships between the response
variable and predictor variables because it uses recursive partitioning
of the dataset to explore the relationships (Myles et al., 2004). Details
of CART are published elsewhere (De'Ath and Fabricius, 2008;
Speybroeck, 1998). CART can indicate the importance of the predictor
variables and associated threshold values for the prediction of the re-
sponse variable and can help gain insight into the splitting process
(Wiesmeier et al., 2011). CART was implemented in R using the rpart
package (R Development Core Team, 2007).
RF was developed as an extension of CART to improve the prediction
accuracy of the model (Breiman, 2001). The RF model-building process
is similar to that of CART, with the difference being that it consists of a
combination of many trees. The details of RF can be found elsewhere
(Chagas et al., 2016; Wiesmeier et al., 2011). The RF model was imple-
mented using the package randomforest in R (R Development Core
Team, 2007).
In the RF model, the errors could be calculated using so-called out-
of-bag (OBB) data, which is a random subset of data that is not involved
in the tree-building process. The mean square error (MSEOBB) was esti-
mated using Eq. (4) (Liaw and Wiener, 2001):
 2
n ^ i OOB
MSEOBB N1 i1 Zi Z
^i OOB is the average of
where Zi is the measured value of the variable and Z
all OBB predictions. The MSEOBB is normalized as it depends on the unit
of the response variable, and the percentage of explained variance
(Varex) is calculated using Eq. (5):
Varex 1MSEOBB =Varz
where the Varz is the total variance of the variable.
2.5. Validation of predication models
To evaluate the performance of the predication models, 60 sampling
points were randomly selected as the validation samples. The perfor-
mances of the models were accessed using the error indices e.g., mean
error (ME), mean squared error (MSE), root mean squared error
(RMSE), and coefcient of determination (R2).
3. Results
3.1. Descriptive statistics
Table 1 gives the descriptive statistics of SOC, SOC density (SOCD),
and other soil/sediment properties for the calibration dataset and the
validation dataset. Descriptive statistics revealed high similarity in
mean values, ranges, SDs, and CVs for SOC, SOCD, and other soil proper-
ties between the calibration dataset and the validation dataset. The
mean values and CVs for SOC of the calibration and validation datasets
were 7.17 g kg1 and 7.42 g kg1, and 44.2% and 42.0%, respectively.
The mean values and CVs for SOCD of the calibration and validation
datasets were 21.9 t ha1 and 22.3 t ha1, and 40.1% and 38.1%, respec-
tively. The values of CV for the other soil/sediment properties varied
from 2.94% to 151.6% (Table 1). According to the CV values, SOC,
SOCD, clay, silt, sand, CaO, and Fe2O3 of both datasets showed moderate
variation, whereas pH, Al2O3 and SiO2 exhibited weak variation. The Cl
content had the highest CV value of 151.6% for the calibration dataset
and 138.8% for the validation dataset, which indicated that Cl had strong
variation in coastal reclaimed land.
3.2. Variable importance
Fig. 2 presents the measured importance of the investigated predic-
tor variables derived from the RF model. Soil/sediment pH was the most
important variable for explaining the variation in SOC. The Cl content
was the second most important variable controlling SOC dynamics.
Soil/sediment silt, sand and CaO contents and reclamation duration
were less important but were still major factors impacting the spatial
4 pattern of SOC. The remaining variables, e.g., Fe2O3, Al2O3, clay and
SiO2 contents, land use, elevation, slope and topographic wetness
index (TWI) were the least important factors for SOC dynamics.
The results of CART also revealed that soil/sediment pH, Cl and silt
contents were the most important factors inuencing the spatial pat-
tern of SOC (Fig. 3). In this tree, the rst split divided the sites into
two groups based on soil/sediment pH: the left branch had a soil/sedi-
ment pH 8.56 and the right branch had a soil/sediment pH b 8.56.
5 The left branch with high soil/sediment pH was subdivided based on
silt content. Higher SOC contents were associated with higher silt con-
tent. The subgroup with higher silt content was further divided on the
basis of Cl content. Sites with a Cl content 1062 mg kg1 had lower
SOC content than the sites with Cl content b1062 mg kg1.
For the right branch divided by the rst split, this group was again
divided based on soil/sediment pH (Fig. 3). The subgroup with higher
pH (pH 8.445) was associated with lower SOC content and was further
divided by Cl content. Sites with higher Cl content (Cl 579.8 mg kg-1)

Table 1
Descriptive statistics of SOC and other soil/sediment properties in the training and the validation dataset.

Calibration set (n = 181) Validation set (n = 60)

Max Min Mean SD CV Max Min Mean SD CV

1
SOC (g kg ) 15.5 0.94 7.17 3.17 44.2 13.9 1.45 7.42 3.12 42.0
SOCD (t ha1) 47.9 3.79 21.9 8.78 40.1 39.8 5.39 22.3 8.50 38.1
Clay (%) 50.2 4.30 15.2 6.73 44.3 41.2 5.50 16.4 6.63 40.4
Silt (%) 48.0 0 32.6 13.4 41.1 50.0 2.40 34.6 11.6 33.5
Sand (%) 94.1 8.21 52.2 17.3 33.1 92.1 17.2 49.0 15.9 32.4
Cl (mg kg1) 15,151 56.4 1607.8 2438 151.6 7305 68.6 1386 1924 138.8
CaO (%) 7.49 3.31 5.48 0.69 12.6 6.99 2.46 5.53 0.77 13.9
pH 9.1 7.8 8.5 0.25 2.94 9.5 7.8 8.5 0.30 3.53
Al2O3 (%) 14.7 9.04 11.2 0.95 8.48 14.1 9.58 11.4 0.89 7.81
Fe2O3 (%) 6.49 3.06 4.17 0.50 12.0 5.94 2.97 4.23 0.54 12.8
SiO2 (%) 71.1 50.6 64.2 3.54 5.51 70.2 52.8 63.5 3.40 5.35

SD, standard deviation; CV, coefcient of variation; SOC, soil/sediment organic carbon content; SOCD, soil/sediment organic carbon density.
708 H. Zhang et al. / Science of the Total Environment 592 (2017) 704713

Table 2
Varimax-rotated principal components matrix.

Variables Factor 1 Factor 2

Clay 0.93 0.11

Silt 0.42 0.84
Sand 0.69 0.69
RD 0.26 0.86
Cl 0.50 0.66
Al2O3 0.94 0.25
SiO2 0.99 0.03
CaO 0.76 0.33
Fe2O3 0.95 0.06
pH 0.19 0.76
Eigenvalue 5.36 2.98
% of total variance 52.1 31.3
Cumulative % of variance 52.1 83.4

RD, reclamation duration.

strong intercorrelations among soil/sediment clay, Al2O3, Fe2O3, SiO2,

and CaO. Factor 2 revealed the close relationships among soil/sediment
silt, Cl, pH, and reclamation duration.

3.4. Multiple linear regression model

Table 3 provides the parameters obtained from the MLR models. In

Fig. 2. Variable importance derived from the random forest model for soil organic carbon
(RD and LU are reclamation duration and land use, respectively).
had lower SOC than the sites with lower Cl content (Cl b 579.8 mg kg1).
The subgroup with lower pH (pH b 8.445) was further split based on
reclamation duration. Higher SOC contents were observed in the areas
with longer reclamation durations. Those with longer reclamation dura-
tions were again divided by pH. Sites with higher pH (pH 8.395) were
associated with lower SOC than those with lower pH (pH b 8.395).
Fig. 4 provides the SOC content and density of the lands with differ-
ent reclamation durations. Signicant differences in SOC content and
density were observed among the lands with different reclamation du-
rations. The tidal at (0 year) had low SOC content and density of
3.94 g kg1 and 12.5 t ha1, respectively. After tidal at reclamation,
SOC content and density presented increased trends with reclamation
durations and reached peak values (9.35 g kg1 and 28.7 t ha1) after
approximately 65 years of land reclamation. Then, SOC content and
density reached stable levels in the conversion from the 65-year old
land to the 90-year old land.
The SOC content and density for the primary land uses in the lands
with different reclamation durations are shown in Fig. 5. In the tidal
ats, grassland revealed signicantly higher SOC content and density
than the bare tidal at at a level of p b 0.05. The reclaimed shpond
and cropland had signicantly higher SOC content and density levels
than the tidal at. On lands with the same reclamation duration, the
SOC content and density of the paddy land were signicantly higher
(p b 0.05) than those of the upland. The highest SOC content
(10.3 g kg1) and density (30.5 t ha 1) were observed in the 65-
year-old paddy land.
3.3. Principal component analysis
To explore the pattern of correlations among the predictor variables,
principal component analysis (PCA) was performed with varimax rota-
tion (Table 2). PCA extracted two uncorrelated components, which ex-
plained 83.4% of the total variance of the database. Factor 1 indicated
the MLR model involving all predictor variables (MLRAll), the sand
content did not enter into the model. The MLRAll model revealed that
the estimated parameters, e.g., Cl, CaO, pH, Al2O3, and SiO2 were
signicant at a level of p b 0.001. The parameters for the constant,
clay, silt, Cl, CaO, pH, Al2O3, SiO2, and Fe2O3 were 338.62, 0.049,
0.043, 0.529, 6.877, 4.014, 5.776, 2.953, and 1.258, respec-
tively. The adjusted coefcient of determination (R2adj), obtained from
the MLRAll model, was 0.797, which indicated that the MLRAll model
could explain 79.7% of the variance in SOC in the study area.
Based on the result of PCA and variable importance (Table 2 and Fig.
2), the important independent variables (pH and CaO) were used to
predict SOC. Table 3 presents the estimated parameters of the MLR
model involving independent predictor variables (pH and CaO)
(MLRInd). The parameters for constant, pH and CaO were 91.131,
0.623 and 9.471, respectively. The estimated parameters e.g. pH
and CaO were signicant at a level of p b 0.001. The adjusted coefcient
of determination (R2adj) of the MLRInd model was 0.607.
3.5. Random forest model
The parameters and results of the RF models are shown in Table 4. In
this study, the number of ntree, mtree, and the nodesize were set to 500,
3 and 5, respectively. Accordingly, the MSEOBB and varex obtained from
the RFAll model were 2.88 g kg1 and 71.3%, respectively. The MSEOBB
(3.16 g kg 1) of the RFInd model was higher than that of the RFAll
model, while the varex (65.1%) of the RFInd model was lower than that
of the RFAll model.
3.6. Validation of prediction models
The performance of the MLR and RF were tested using the validation
dataset (Fig. 6). The ME, MSE, RMSE, and R2 were the indicators used to
determine the performance of the two models.
As shown in Fig. 6, the RFAll model revealed low predication errors
(ME = 0.0248 g kg1, MSE = 0.438 g kg1, RMSE = 0.662 g kg1)
and high R2 (R2 = 0.97). In the MLRAll model, the values of the three
error indices were much higher (ME = 0.186 g kg 1, MSE =
2.063 g kg 1, RMSE = 1.436 g kg 1) while the R2 (R2 = 0.79) was
much lower compared with the RFAll model.
The RFInd and MLRInd models revealed higher MSE (3.793 and
4.185 g kg1) and RMSE (1.948 and 2.046 g kg1) and lower R2 (0.62
and 0.56) than the RFAll and MLRAll models (Fig. 6). Similarly, the RFInd
 H. Zhang et al. / Science of the Total Environment 592 (2017) 704713 709

Table 3
Estimated parameters of the multiple linear regression model for SOC.

Variable MLRAll MLRInd

Parameter Std. error t-Value p-Value Parameter Std. error t-Value p-Value

Constant 338.62 31.455 10.765 0.000 91.131 5.023 18.141 0.000

Clay 0.049 0.042 1.142 0.255
Silt 0.043 0.016 2.608 0.010
Sanda
Cl 0.529 0.075 7.058 0.000
CaO 6.877 0.695 9.898 0.000 0.623 0.216 2.882 0.004
pH 4.014 0.613 6.547 0.000 9.471 0.583 16.233 0.000
Al2O3 5.776 0.814 7.097 0.000
SiO2 2.953 0.319 9.250 0.000
Fe2O3 1.258 0.716 0.198 0.081

MLRAll, the MLR model involving all predictor variables. MLRInd, the MLR model involving independent variables (pH and CaO).
a
The independent variable that was not involved in the MLRAll model.

model produced lower error indices and higher R2 than the MLRInd
model (Fig. 6).
4. Discussion
4.1. Effects of soil/sediment pH, Cl and silt on SOC dynamics on coastal
reclaimed lands
The results of RF and CART indicated that the soil/sediment pH was
identied as the most important factor inuencing SOC dynamics in a
coastal reclaimed region (Figs. 2 and 3). On coastal reclaimed lands,
SOC can be affected by the soil pH in direct and indirect ways. On the
one hand, coastal reclaimed soils were characterized by a relatively
high pH due to the affection of seawater (Table 1). Under the inuence
of high soil/sediment pH, plant growth may be limited by the decreased
solubility of iron, zinc, and manganese (Grybos et al., 2009). The input of
crop residues into the soil is thus limited, resulting in lower SOC content
in the topsoil. Moreover, soil dissolved organic carbon is easier to re-
lease under high pH conditions than under low pH condition
(Andersson et al., 2000; Grybos et al., 2009), which consequently led
to short SOC turnover rate in relatively acidic conditions. On the other
hand, soil/sediment pH is a sensitive indicator of land management
such as fertilizer application (Guo et al., 2010; Hgberg et al., 2006). In
the nutrition-poor reclaimed lands, the increase in intensity and fre-
quency of chemical fertilizer promotes crop biomass production and
thus improves the SOC level (Kalbitz et al., 2013). A study conducted
on coastal reclaimed lands also revealed that the soil pH was the most
important factor controlling the SOC dynamics (Deng et al., 2016).
The Cl content was the second most important variable related to
the SOC dynamics (Figs. 2 and 3). It has been widely reported that soil
salinity is a major factor affecting organic carbon cycling in salt-affected
soils (Morrissey et al., 2014; Rath and Rousk, 2015; Setia et al., 2012). In
general, soils with relatively high salinity are low in SOC, probably due
to the limitation of plant growth and thus the reduction of crop residues
into soils (Setia et al., 2013). Moreover, increasing soil salt levels causes
an ion imbalance and nutrient deciency that may decrease plant pro-
ductivity and impact nutrient cycling (Setia et al., 2012). Other studies
Table 4
Parameters and results of the RF model.
Model parameters Model results
Ntree Mtry Nodesize MSEOBB (g kg1)
RFAll 500 3 5 2.88
RFInd 500 3 5 3.16
RFAll, the RF model involving all predictor variables. RFInd, the RF model involving indepen-
dent variables (pH and CaO). Ntree, the number of trees in the forest. Mtry, the number of
variables tried at each split. Nodesize, the minimum size of the leaf. Varex indicates the per-
centage of the explained variance.
have also revealed signicant and negative correlations between soil sa-
linity and SOC content in tidal wetlands (Morrissey et al., 2014).
The soil/sediment silt content was identied as a less but still impor-
tant factor affecting the spatial pattern of SOC in coastal reclaimed lands
(Figs. 2 and 3). Soil ne particles have been widely proven to be one of
the main factors impacting organic carbon in soils (Kong et al., 2009;
Mao et al., 2015; Zinn et al., 2005). Generally, ne-textured soils with
higher silt content can store more organic carbon than coarse-tex-
tured soils when supplied with a similar input of organic material
(Hassink, 1997; Wiesmeier et al., 2015). This can be linked to the im-
portant role of soil silt not only in OM inputs indirectly, by impacting
plant productivity via water availability and soil fertility (Schimel
and Parton, 1986), but also in OM outputs through the role of silt in
the protection of OM from decomposition (Anderson et al., 1981;
Hassink, 1997).
4.2. SOC dynamics in relation to reclamation duration and land use in
coastal reclaimed lands
Reclamation duration has been typically used to study SOC dynamics
in coastal reclaimed lands. Some studies have revealed remarkable im-
pacts of reclamation duration on SOC dynamics (Cheng et al., 2009; Cui
et al., 2012; Fu et al., 2014). In this study, we also observed signicant
increases in SOC content and density with increased reclamation dura-
tion (Fig. 4). After coastal wetlands reclamation, the reclaimed lands
were normally used as farmland. Thus, the increase in SOC content
and density along with reclamation duration may be caused by the in-
crease in agricultural intensity because the intensity of agricultural ac-
tivities was positively correlated with reclamation duration (Wang et
al., 2014a). Organic manure application can directly improve the SOC
level, while chemical fertilization can promote biomass production
and thus enhance crop residue inputs (Meng et al., 2005; Steiner et al.,
2007). Rapid desalination and dealkalization along with reclamation
duration caused by intensive fresh water irrigation may also be major
factors in SOC accumulation (Iost et al., 2007; Xu et al., 2013; Yin et
al., 2016) because soil/sediment pH and Cl content were identied as
the two most important variables impacting SOC in reclaimed soils
(Figs. 2 and 3).
According to the RF model, land use was a relatively weak variable
for predicting SOC in comparison with these dominating predictors
(Fig. 2). This was inconsistent with other studies that highlighted the
Varex (%) decided importance of land use in predicting SOC spatial distribution
(Deng et al., 2016; Wiesmeier et al., 2011). Although land use is not
71.3
65.1 an important predictor for predicting SOC dynamics in this study
(Fig. 2), the differences in SOC between different land uses reached sig-
nicant levels (Fig. 5). The signicantly higher SOC levels of shponds
and croplands compared to tidal ats reected the distinct impact of
human activity on SOC accumulation in reclaimed soils.
710 H. Zhang et al. / Science of the Total Environment 592 (2017) 704713
Fig. 3. Regression tree structure for soil organic carbon (RD is the reclamation duration).
On lands with identical reclamation duration, the remarkable differ-
ences in SOC between different land uses was a reection of how land
use affects SOC dynamic on coastal reclaimed lands. On tidal ats, the
grasslands were characterized by higher SOC content and density than
those of the adjacent bare tidal ats because of the larger SOM inputs
of grasslands (Kagawa et al., 2014) (Fig. 5). Paddy soils were superior
to uplands in sequestering SOC (Fig. 5), which is a common result not
only on general cropland (Huang et al., 2014; Pan et al., 2003; Wu,
2011) but also on reclaimed cropland (Fu et al., 2014; Kalbitz et al.,
2013; Wissing et al., 2013). Paddy management is thus widely recog-
nized as a promising way for SOC sequestration and global warming
mitigation (Deng et al., 2016; Wissing et al., 2014). The larger OM inputs
and lower OM decomposition rate of paddy soils relative to uplands are
primary mechanisms for the higher SOC level of paddy soil (Cui et al.,
2012). However, the following several mechanisms might also explain
the higher topsoil OC level of paddy soil relative to uplands. First, com-
pared with uplands, soils under ooding conditions enhance the forma-
tion of amorphous Fe oxyhydrate, which has a large specic surface area
for absorbing OM (Cheng et al., 2009; Kgel-Knabner et al., 2010; Klbl
et al., 2014). Second, the formation of a dense pan under paddy elds
under management leads to accumulation of SOC in topsoil (Klbl et
al., 2014) because the dense pan decreases the DOC leaching from top-
soil into subsoil. Third, the relatively low SOC level on uplands is likely
linked to agricultural tillage practices (Hernanz et al., 2002). For soils
with low water content, the seasonal tillage tends to destroy the topsoil
structure, destabilize soil macroaggregates, and thereby enhance the
loss of SOC (Six et al., 2000).
Fig. 4. Soil organic carbon content and density on the lands with different reclamation durations (years). Different letters above each box indicate a signicant difference at a level of
p b 0.05.
4.3. Evaluation and comparison of RF and MLR
Since the training and validation datasets had similar descriptive sta-
tistics, e.g., maximums, minimums, mean values, SDs and CVs (Table 1),
analyses of the performance of RF and MLR by independent validation
datasets are reasonable. The high R2 for both RFAll (R2 = 0.97) and
MLRAll (R2 = 0.79) indicated they had satisfactory predictive abilities
for OC on coastal reclaimed lands. The low MEs (0.0248 for RFAll and
0.186 for MLRAll) were relatively close to zero for RF and MLR and re-
vealed unbiased predictions for SOC, although the negative values for
the models point towards slight overestimations. The R2 of the RFInd
and MLRInd models were much lower than those of the RFAll and MLRAll
models, indicating the importance of other predictor variables in SOC
prediction. Remarkably, the higher R2 (Fig. 6) with the lower error indi-
ces (ME, MSE and RMSE) (Fig. 6) signied that the RF performed much
better than the MLR in predicating SOC on coastal reclaimed lands. The
superior results obtained from the RF compared with MLR for the pre-
diction of SOC and other soil properties have already been reported in
other studies (Chagas et al., 2016; Guo et al., 2015).
In general, the relationships between the response variable and pre-
dictor variables were nonlinear (Lark, 1999). The use of MLR can only
explain part of the variance of the response variable. However, the RF
model required no assumptions about relationships between the re-
sponse variable and predictor variables, it could handle the nonlinear
and hierarchical relationships that cannot be handled by MLR (Guo et
al., 2015). In fact, the relationship between SOC and the predictor vari-
ables proved to be nonlinear and hierarchical in CART for the coastal
 H. Zhang et al. / Science of the Total Environment 592 (2017) 704713 711

Fig. 5. Soil organic carbon content and density for different land uses in Dafeng. Different letters above each column indicate signicant differences at a level of p b 0.05.

reclaimed lands (Fig. 3). Thus, in this study, the RF indicates better per-
formance than the MLR in predicting soil/sediment OC.
4.4. Implication for coastal reclaimed soil management
Low fertility and high salinity and alkalinity are general characteris-
tics for coastal lands (Iost et al., 2007; Li et al., 2013). With the acceler-
ated reclamation of coastal wetlands, efcient land management is
required to improve the original poor soil conditions. Given the key
roles of SOC in soil fertility and climate change (Lal, 2004; Tiessen et
al., 1994), soil management strategies aimed at enlarging the SOC pool
are particularly needed. The RF and CART models revealed that desali-
nation and dealkalization are the most efcient methods to improving
SOC level in coastal reclaimed lands (Figs. 2 and 3). Thus, after tidal
ats reclamation, the reclaimed soils were immediately used for fresh
water aquaculture, followed by cereal and cash crop production. These
land management strategies have proven their ability to decrease soil
pH and salt content (Xu et al., 2013; Yin et al., 2016) and enhance the

Fig. 6. Performance of the RF and MLR models in predicting soil organic carbon. RFAll, the RF model involving all predictor variables; MLRAll, the MLR model involving all predictor variables;
RFInd, the RF model involving independent variables (pH and CaO); MLRInd, the MLR model involving independent variables (pH and CaO).
712 H. Zhang et al. / Science of the Total Environment 592 (2017) 704713
SOC pool (Kalbitz et al., 2013; Zhao et al., 2013). Generally, SOC in-
creased signicantly with an increase in reclamation duration, as re-
vealed by the results of this study (Fig. 4) and other studies (Cheng et
al., 2009; Fu et al., 2014). However, the highest level of SOC was ob-
served on the 65-year-old paddy land, rather than that of the 90-year-
old upland (Fig. 5). This was a reection of the superiority of paddy
management in sequestering SOC.
As stated above, the relatively higher SOC level of paddy soils com-
pared to that of upland soils can be linked to the higher OM input and
the lower OM output, the formation of amorphous Fe oxyhydrates, the
dense pan, and the tillage-induced loss of SOC on uplands (Cui et al.,
2012; Klbl et al., 2014; Pan et al., 2003; Six et al., 2000). Nevertheless,
in the salt-affected reclaimed lands, paddy soils irrigated with freshwa-
ter may be characterized by a higher rate of desalinization and
dealkalization, which were recognized as the two most important
drivers of SOC dynamics (Figs. 2 and 3). Moreover, in eastern China,
rice is a major cereal crop with a signicantly higher yield than up-
land-based crops (Zhao et al., 2012). Considering the signicant role of
paddy management in SOC sequestration and cereal production (Pan
et al., 2003; Kalbitz et al., 2013), paddy eld management can thus be
recommended as an efcient approach for mitigating global warming
and ensuring food security. In fact, paddy elds have normally been rec-
ognized as an environmentally friendly and sustainable land use (Cui et
al., 2012; Kgel-Knabner et al., 2010; Zou et al., 2011).
5. Conclusion
In coastal reclaimed lands, the applied MLRAll and RFAll approaches to
the prediction of SOC yielded promising results as both models revealed
low error indices and high R2. The RFInd and MLRInd produced lower R2
and higher error indices than the MLRAll and RFAll models. However,
RF performed much better than MLR, which is likely due to the superi-
ority of RF in handling non-linear and hierarchical relationships be-
tween SOC and the predictors. The soil/sediment pH, Cl and silt
contents were identied as the most important factors impacting the
spatial pattern of SOC on coastal reclaimed lands. The analyses of land
use effects on SOC dynamics conrmed that paddy soils were superior
to other land uses in sequestering SOC. Considering the potential rapid
desalinization and dealkalization under submerged conditions, paddy
management is thus recommended as an environmentally sustainable
approach to managing newly reclaimed lands.
Acknowledgements
This study was funded by the Chinese Geological Survey
(1212011220053).
References
Alvarez, R., Steinbach, H., Bono, A., 2011. An articial neural network approach for
predicting soil carbon budget in agroecosystems. Soil Sci. Soc. Am. J. 75, 965975.
An, S., Li, H., Guan, B., Zhou, C., Wang, Z., Deng, Z., Zhi, Y., Liu, Y., Xu, C., Fang, S., 2007.
China's natural wetlands: past problems, current status, and future challenges.
AMBIO J. Hum. Environ. 36, 335342.
Anderson, D., Saggar, S., Bettany, J., Stewart, J., 1981. Particle size fractions and their use in
studies of soil organic matter: I. The nature and distribution of forms of carbon, nitro-
gen, and sulfur. Soil Sci. Soc. Am. J. 45, 767772.
Andersson, S., Nilsson, S.I., Saetre, P., 2000. Leaching of dissolved organic carbon (DOC)
and dissolved organic nitrogen (DON) in mor humus as affected by temperature
and pH. Soil Biol. Biochem. 32, 110.
Bao, S.D., 2000. Soil Agricultural Chemistry Analysis. China Agricultural Press, Beijing (In
Chinese).
Breiman, L., 2001. Random forests. Mach. Learn. 45, 532.
Cambardella, C., Moorman, T., Parkin, T., Karlen, D., Novak, J., Turco, R., Konopka, A., 1994.
Field-scale variability of soil properties in central Iowa soils. Soil Sci. Soc. Am. J. 58,
15011511.
Chagas, C.D.S., Junior, W.D.C., Bhering, S.B., Filho, B.C., 2016. Spatial prediction of soil sur-
face texture in a semiarid region using random forest and multiple linear regressions.
Catena 139, 232240.
Cheng, Y.Q., Yang, L.-Z., Cao, Z.-H., Ci, E., Yin, S., 2009. Chronosequential changes of select-
ed pedogenic properties in paddy soils as compared with non-paddy soils. Geoderma
151, 3141.
Chmura, G.L., Anisfeld, S.C., Cahoon, D.R., Lynch, J.C., 2003. Global carbon sequestration in
tidal, saline wetland soils. Glob. Biogeochem. Cycles 17:1111. http://dx.doi.org/10.
1029/2002GB001917.
Chris, J., Claire, M.C., Kevin, C., Peter, C., Peter, F., David, J., David, P., 2004. Global climate
change and soil carbon stocks; predictions from two contrasting models for turnover
of organic carbon in soil. Glob. Chang. Biol. 11, 154166.
Cui, J., Liu, C., Li, Z.L., Wang, L., Chen, X.F., Ye, Z.Z., Fang, C.M., 2012. Long-term changes in
topsoil chemical properties under centuries of cultivation after reclamation of coastal
wetlands in the Yangtze Estuary, China. Soil Tillage Res. 123, 5060.
De'Ath, G., Fabricius, K.E., 2008. Classication and regression trees: a poweful yet simple
technique for ecological data analysis. Ecology 81, 31783192.
DeGryze, S., Six, J., Paustian, K., Morris, S.J., Paul, E.A., Merckx, R., 2004. Soil organic carbon
pool changes following land-use conversions. Glob. Chang. Biol. 10, 11201132.
Deng, X., Zhan, Y., Wang, F., Ma, W., Ren, Z., Chen, X., Qin, F., Long, W., Zhu, Z., Lv, X., 2016.
Soil organic carbon of an intensively reclaimed region in China: current status and
carbon sequestration potential. Sci. Total Environ. 565, 539546.
Doetterl, S., Stevens, A., Six, J., Merckx, R., Van Oost, K., Pinto, M.C., Casanova-Katny, A.,
Muoz, C., Boudin, M., Venegas, E.Z., 2015. Soil carbon storage controlled by interac-
tions between geochemistry and climate. Nat. Geosci. 8, 780783.
Fernndez, S., Santn, C., Marqunez, J., lvarez, M.A., 2010. Saltmarsh soil evolution after
land reclamation in Atlantic estuaries (Bay of Biscay, North coast of Spain). Geomor-
phology 114, 497507.
Florinsky, I.V., Eilers, R.G., Manning, G.R., Fuller, L.G., 2002. Prediction of soil properties by
digital terrain modelling. Environ. Model. Softw. 17, 295311.
Fu, Q., Ding, N., Liu, C., Lin, Y., Guo, B., 2014. Soil development under different cropping
systems in a reclaimed coastal soil chronosequence. Geoderma 230231, 5057.
Grimm, R., Behrens, T., Mrker, M., Elsenbeer, H., 2008. Soil organic carbon concentrations
and stocks on Barro Colorado Island digital soil mapping using random forests
analysis. Geoderma 146, 102113.
Grybos, M., Davranche, M., Gruau, G., Petitjean, P., Pdrot, M., 2009. Increasing pH drives
organic matter solubilization from wetland soils under reducing conditions.
Geoderma 154, 1319.
Guo, J., Liu, X., Zhang, Y., Shen, J., Han, W., Zhang, W., Christie, P., Goulding, K., Vitousek, P.,
Zhang, F., 2010. Signicant acidication in major Chinese croplands. Science 327
(5968), 10081010.
Guo, P.T., Wu, W., Sheng, Q.K., Li, M.F., Liu, H.B., Wang, Z.Y., 2013. Prediction of soil organic
matter using articial neural network and topographic indicators in hilly areas. Nutr.
Cycl. Agroecosyst. 95, 333344.
Guo, P.T., Li, M.F., Luo, W., Tang, Q.F., Liu, Z.W., Lin, Z.M., 2015. Digital mapping of soil or-
ganic matter for rubber plantation at regional scale: an application of random forest
plus residuals kriging approach. Geoderma 237238, 4959.
Hassink, J., 1997. The capacity of soils to preserve organic C and N by their association
with clay and silt particles. Plant Soil 191, 7787.
Hernanz, J., Lpez, R., Navarrete, L., Sanchez-Giron, V., 2002. Long-term effects of tillage
systems and rotations on soil structural stability and organic carbon stratication in
semiarid central Spain. Soil Tillage Res. 66, 129141.
Hobley, E., Wilson, B., Wilkie, A., Gray, J., Koen, T., 2015. Drivers of soil organic carbon stor-
age and vertical distribution in Eastern Australia. Plant Soil 390, 111127.
Hgberg, P., Fan, H., Quist, M., Dan, B., Tamm, C.O., 2006. Tree growth and soil acidication
in response to 30 years of experimental nitrogen loading on boreal forest. Glob.
Chang. Biol. 12, 489499.
Hu, K., Li, H., Li, B., Huang, Y., 2007. Spatial and temporal patterns of soil organic matter in
the urbanrural transition zone of Beijing. Geoderma 141 (3), 302310.
Huang, S., Pan, X., Guo, J., Qian, C., Zhang, W., 2014. Differences in soil organic carbon
stocks and fraction distributions between rice paddies and upland cropping systems
in China. J. Soils Sediments 14, 8998.
Iost, S., Landgraf, D., Makeschin, F., 2007. Chemical soil properties of reclaimed marsh soil
from Zhejiang Province P.R. China. Geoderma 142, 245250.
Jarvis, A., Reuter, H.I., Nelson, A., Guevara, E., 2008. Hole-lled Seamless SRTM Data V4. In-
ternational Centre for Tropical Agriculture (CIAT) http://srtm.csi.cgiar.org.
Jin, W., Yang, J., Wang, X., 2013. Spatial distribution of organic carbon in coastal saline soil
and its correlation with reclamation age. Trans. Chin. Soc. Agric. Eng. 2013, 8994 (in
Chinese with English abstract).
Kagawa, M., Matsubara, K., Kimura, K., Shiono, H., Fukui, Y., 2014. Impacts of organic mat-
ter amendments on carbon and nitrogen dynamics in grassland soils. Soil Biol.
Biochem. 68, 5261.
Kalbitz, K., Kaiser, K., Fiedler, S., Klbl, A., Amelung, W., Bruer, T., Cao, Z., Don, A., Grootes,
P., Jahn, R., 2013. The carbon count of 2000 years of rice cultivation. Glob. Chang. Biol.
19, 11071113.
Kgel-Knabner, I., Amelung, W., Cao, Z., Fiedler, S., Frenzel, P., Jahn, R., Kalbitz, K., Klbl, A.,
Schloter, M., 2010. Biogeochemistry of paddy soils. Geoderma 157, 114.
Klbl, A., Schad, P., Jahn, R., Amelung, W., Bannert, A., Cao, Z., Fiedler, S., Kalbitz, K.,
Lehndorff, E., Mller-Niggemann, C., 2014. Accelerated soil formation due to paddy
management on marshlands (Zhejiang Province, China). Geoderma 228, 6789.
Kong, X., Dao, T.H., Qin, J., Qin, H., Li, C., Zhang, F., 2009. Effects of soil texture and land use
interactions on organic carbon in soils in North China cities' urban fringe. Geoderma
154, 8692.
Lal, R., 2004. Soil carbon sequestration impacts on global climate change and food securi-
ty. Science 304 (5677), 16231627.
Lark, R.M., 1999. Soillandform relationships at within-eld scales: an investigation using
continuous classication. Geoderma 92, 141165.
Li, X., Sun, Y., Mander, ., He, Y., 2013. Effects of land use intensity on soil nutrient distri-
bution after reclamation in an estuary landscape. Landsc. Ecol. 28, 699707.
 H. Zhang et al. / Science of the Total Environment 592 (2017) 704713
Li, J., Pu, L., Zhu, M., Zhang, J., Li, P., Dai, X., Xu, Y., Liu, L., 2014. Evolution of soil properties
following reclamation in coastal areas: a review. Geoderma 226227, 130139.
Liaw, A., Wiener, M., 2001. Classication and Regression by Random Forest. R News, p. 23.
Lu, R., 2000. Methods of Soil Agrochemistry Anlysis. Agricultural Science and Technology
Press, Beijing (in Chinese).
Mao, D., Wang, Z., Li, L., Miao, Z., Ma, W., Song, C., Ren, C., Jia, M., 2015. Soil organic carbon
in the Sanjiang Plain of China: storage, distribution and controlling factors. Biogeosci-
ences 12, 16351645.
Mckenzie, N.J., Ryan, P.J., 1999. Spatial prediction of soil properties using environmental
correlation. Geoderma 89, 6794.
Meng, L., Ding, W., Cai, Z., 2005. Long-term application of organic manure and nitrogen
fertilizer on N2O emissions, soil quality and crop production in a sandy loam soil.
Soil Biol. Biochem. 37, 20372045.
Morrissey, E.M., Gillespie, J.L., Morina, J.C., Franklin, R.B., 2014. Salinity affects microbial
activity and soil organic matter content in tidal wetlands. Glob. Chang. Biol. 20,
13511362.
Myles, A.J., Feudale, R.N., Liu, Y., Woody, N.A., Brown, S.D., 2004. An introduction to deci-
sion tree modeling. J. Chemom. 18, 275285.
Pan, G., Li, L., Wu, L., Zhang, X., 2003. Storage and sequestration potential of topsoil organ-
ic carbon in China's paddy soils. Glob. Chang. Biol. 10, 7992.
Rath, K.M., Rousk, J., 2015. Salt effects on the soil microbial decomposer community and
their role in organic carbon cycling: a review. Soil Biol. Biochem. 81, 108123.
Rawls, W.J., 1983. Estimating soil bulk density from particle size analysis and organic mat-
ter content. Soil Sci. 135, 123125.
Richter, D.D., Markewitz, D., Trumbore, S.E., Wells, C.G., 1999. Rapid accumulation and
turnover of soil carbon in a re-establishing forest. Nature 400 (6739), 5658.
Rossel, R.A.V., Behrens, T., 2010. Using data mining to model and interpret soil diffuse re-
ectance spectra. Geoderma 158, 4654.
Ru, F., Yin, A., Jin, J., Zhang, X., Yang, X., Zhang, M., Gao, C., 2016. Prediction of cadmium
enrichment in reclaimed coastal soils by classication and regression tree. Estuar.
Coast. Shelf Sci. 177, 17.
Schimel, D., Parton, W., 1986. Microclimatic controls of nitrogen mineralization and nitri-
cation in shortgrass steppe soils. Plant Soil 93, 347357.
Schmidt, M.W., Torn, M.S., Abiven, S., Dittmar, T., Guggenberger, G., Janssens, I.A., Kleber,
M., Kgel-Knabner, I., Lehmann, J., Manning, D.A., 2011. Persistence of soil organic
matter as an ecosystem property. Nature 478 (7367), 4956.
Setia, R., Smith, P., Marschner, P., Gottschalk, P., Baldock, J., Verma, V., Setia, D., Smith, J.,
2012. Simulation of salinity effects on past, present, and future soil organic carbon
stocks. Environ. Sci. Technol. 46, 16241631.
Setia, R., Gottschalk, P., Smith, P., Marschner, P., Baldock, J., Setia, D., Smith, J., 2013. Soil
salinity decreases global soil organic carbon stocks. Sci. Total Environ. 465, 267272.
Six, J., Elliott, E., Paustian, K., 2000. Soil macroaggregate turnover and microaggregate for-
mation: a mechanism for C sequestration under no-tillage agriculture. Soil Biol.
Biochem. 32, 20992103.
Smith, P., House, J.I., Bustamante, M., Sobock, J., Harper, R., Pan, G., West, P., Clark, J.,
Adhya, T., Rumpel, C., 2015. Global change pressures on soils from land use and man-
agement. Glob. Chang. Biol. 22, 10081028.
Speybroeck, N., 1998. Classication and Regression Trees. Chapman & Hall/CRC.
Steiner, C., Teixeira, W.G., Lehmann, J., Nehls, T., de Macdo, J.L.V., Blum, W.E., Zech, W.,
2007. Long term effects of manure, charcoal and mineral fertilization on crop produc-
tion and fertility on a highly weathered Central Amazonian upland soil. Plant Soil
291, 275290.
Thompson, J.A., Kolka, R.K., 2005. Soil carbon storage estimation in a forested watershed
using quantitative soil-landscape modeling. Soil Sci. Soc. Am. J. 69, 10861093.
Thompson, J.A., Pena-Yewtukhiw, E.M., Grove, J.H., 2006. Soillandscape modeling across
a physiographic region: topographic patterns and model transportability. Geoderma
133, 5770.
713
Tiessen, H., Cuevas, E., Chacon, P., 1994. The role of soil organic matter in sustaining soil
fertility. Nature 371 (6500), 783785.
Tranter, G., Minasny, B., McBratney, A., Murphy, B., McKenzie, N., Grundy, M., Brough, D.,
2007. Building and testing conceptual and empirical models for predicting soil bulk
density. Soil Use Manag. 23, 437443.
Wang, L., Coles, N., Wu, C., Wu, J., 2014a. Effect of long-term reclamation on soil properties
on a coastal plain, southeast China. J. Coast. Res. 30, 661669.
Wang, W., Liu, H., Li, Y., Su, J., 2014b. Development and management of land reclamation
in China. Ocean Coast. Manag. 102, 415425.
Wiesmeier, M., Barthold, F., Blank, B., Kgelknabner, I., 2011. Digital mapping of soil or-
ganic matter stocks using random forest modeling in a semi-arid steppe ecosystem.
Plant Soil 340, 724.
Wiesmeier, M., Sprlein, P., Geu, U., Hangen, E., Haug, S., Reischl, A., Schilling, B., Ltzow,
M., Kgel-Knabner, I., 2012. Soil organic carbon stocks in southeast Germany (Bavar-
ia) as affected by land use, soil type and sampling depth. Glob. Chang. Biol. 18,
22332245.
Wiesmeier, M., Munro, S., Barthold, F., Steffens, M., Schad, P., Kgel-Knabner, I., 2015. Car-
bon storage capacity of semi-arid grassland soils and sequestration potentials in
northern China. Glob. Chang. Biol. 21, 38363845.
Wilson, J.P., Gallant, J.C., 2000. Terrain Analysis: Principles and Applications. Wiley, New
York.
Wissing, L., Klbl, A., Vogelsang, V., Fu, J.-R., Cao, Z.-H., Kgel-Knabner, I., 2011. Organic
carbon accumulation in a 2000-year chronosequence of paddy soil evolution. Catena
87, 376385.
Wissing, L., Klbl, A., Husler, W., Schad, P., Cao, Z.H., Kgel-Knabner, I., 2013. Manage-
ment-induced organic carbon accumulation in paddy soils: the role of organo-miner-
al associations. Soil Tillage Res. 126, 6071.
Wissing, L., Klbl, A., Schad, P., Bruer, T., Cao, Z.-H., Kgel-Knabner, I., 2014. Organic car-
bon accumulation on soil mineral surfaces in paddy soils derived from tidal wetlands.
Geoderma 228229, 90103.
Wolff, W.J., 1992. The end of a tradition: 1000 years of embankment and reclamation of
wetlands in the Netherlands. Ambio 287291.
Wu, J., 2011. Carbon accumulation in paddy ecosystems in subtropical China: evidence
from landscape studies. Eur. J. Soil Sci. 62, 2934.
Xu, Y., Pu, L., Zhu, M., Li, J., Zhang, M., Li, P., Zhang, J., 2013. Spatial variation of soil salinity
in the coastal reclamation area, Eastern China. J. Coast. Res. 30, 411417.
Yin, A., Zhang, M., Gao, C., Yang, X., Xu, Y., Wu, P., Zhang, H., 2016. Salinity evolution of
coastal soils following reclamation and intensive usage, Eastern China. Environ.
Earth Sci. 75, 111.
Zhang, X., Yan, C., Xu, P., Dai, Y., Yan, W., Ding, X., Zhu, C., Mei, D., 2013. Historical evolu-
tion of tidal at reclamation in the Jiangsu coastal areas. Acta Geograph. Sin. 68,
15491558 (in Chinese with English abstract).
Zhang, X., Xu, X., Liu, Y., Wang, J., Xiong, Z., 2016. Global warming potential and green-
house gas intensity in rice agriculture driven by high yields and nitrogen use efcien-
cy. Biogeosciences 13, 27012714.
Zhao, X., Zhou, Y., Wang, S., Xing, G., Shi, W., Xu, R., Zhu, Z., 2012. Nitrogen balance in a
highly fertilized ricewheat double-cropping system in Southern China. Soil Sci.
Soc. Am. J. 76, 10681078.
Zhao, G., Xing, S., Liu, L., Liu, Z., 2013. Innovative articial agro-ecosystems enhance soil
carbon sequestration in coastal zones of Southeast China. Clean: Soil, Air, Water 41,
581586.
Zinn, Y.L., Lal, R., Resck, D.V.S., 2005. Texture and organic carbon relations described by a
prole pedotransfer function for Brazilian Cerrado soils. Geoderma 127, 168173.
Zou, P., Fu, J., Cao, Z., 2011. Chronosequence of paddy soils and phosphorus sorptionde-
sorption properties. J. Soils Sediments 11, 249259.