1 Introduction

Many negative effects on the environment are caused by human activities when those activities eliminate existing vegetation, destroy the genetic soil profile, disturb wildlife and habitat, alter current land uses, and permanently change the regional topography and the surface ecological system. Therefore, land cover change is a significant indicator of environmental evolution. As the most important information sources, remote sensing techniques and remotely sensed data play a key role in monitoring and analyzing this process. Extracting the amount of soil erosion using remotely sensed data is an important task in both remote sensing technology and agricultural application fields. In fact, different approaches for soil erosion extraction using the Universal Soil Loss Equation/Revised Universal Soil Loss Equation(USLE/RUSLE) and other algorithms have been developed, some of them using only ground surveying methods(Hammad et al.,2004; Yuan, 2008; Wang and Bing, 2009; Qin et al.,2010).

Remote sensing(RS)has been viewed as the one of the most effective tools for environmental monitoring, urban resources and environment investigation, change detection, and urban growth analysis. RS has been used to monitor land use/cover change, quantify subsidence land , analyze the dynamic change and simulate the trends in the land scape, and assess the feasibility and performance of land reclamation and ecological reconstruction. Supported by RS data, many land cover classification methods(Read and Lam, 2002; Latifovic et al.,2005; Townsend et al.,2009; Wu et al.,2009)as well as prediction and simulation algorithms(Lin and Wang, 2006; Yang et al.,2007; Wang and Bing, 2009)have been proposed and are widely used to monitor land use/cover and environment changes. However, there has been insufficient research on soil erosion, and the application extent is also somewhat superficial(Lin and Wang, 2006; Drzewiecki and Mularz, 2008; Yuan, 2008; Zhang et al.,2008; Xie and Lin, 2010), even though it is of great significance to predict and have a thorough underst and ing of soil erosion.

This research was conducted in the exurb area of Xuzhou City, located in eastern Jiawang and western Tongshan counties, east-central China. In order to underst and the scope of land cover change and corresponding environmental impacts in that coal mining area, we integrated a Support Vector Machine(SVM)classifier, the CA_Markov model, and the RUSLE to visualize, analyze, assess, simulate, and predict the spatial pattern and evolution dynamics of land use/cover change(LUCC) and soil erosion in the area.

This study was structured as follows. First, land cover types of the exurb area were obtained using an SVM classifier. Next, the CA_Markov model was explored to predict the land cover types and land scape situations according to multi-temporal land cover maps derived by the previous phase. Then soil erosion amounts and distribution maps were obtained using the RUSLE method. Finally, relationships between land cover types and soil erosion amounts and distributions, and the effects of the various parameters in RUSLE, were analyzed.

Considering the real situation of the exurb and the need to monitor LUCC, land sat TM/ETM+ scenes were categorized into three classes: vegetation, built-up area, and water body. An SVM supervised classifier was adopted to obtain land cover maps, in terms of efficiency, accuracy, and generalization ability(Giacco et al.,2010). With the CA_Markov model, future land cover/use can be modeled on the basis of preceding state, and a matrix of observed transition probabilities between different periods can be used for prediction(Peterson et al.,2009). The area soil erosion was calculated based on RUSLE, which could provide a scientific basis for soil erosion monitoring. A flow chart of this work is shown in Figure 1.

SVM is a classification system derived from statistical learning theory. An SVM classifier with RBF(radial basis function)kernels was selected in this work because, according to the results from previous researches, it works well in most cases. The mathematical representation of RBF is as follows:

where γ is the bias term in the kernel function for the polynomialand sigmoid kernels(Wu et al.,2004).

The CA_Markov model uses cellular automata in combination with Markov chain analysis, and allows the transition probabilities of one pixel to be a function of neighboring pixels. The process of LUCC and soil erosion simulation based on CA_Markov model can be described as follows(Sang et al.,2010; Liu et al.,2012):(1)Original data obtain and processing. Regional LUCC maps were derived using SVM classifier; Soil erosion maps were calculated using RUSLE.(2)Determining the transition rules. LUCC transition probability matrix and the transfer area of the matrix are achieved with GIS spatial overlay analysis. For soil erosion maps, the quantitative data is firstly converted to qualitative data, the spatial overlay analysis is applied on the qualitative soil erosion maps. The calculated transition probability matrix was served as CA_Markov transition rule.(3)Simulation based on CA_Markov model. It is import to determine start status, cell loop times and CA filters to run CA_Markov model. Taking a stage for instance, in order to simulate LUCC/ soil erosion map in 2013, the LUCC/ soil erosion maps in 2011 and 2009 were selected to calculate transition rule. The results of 2008 were served as starting point and the st and ard 5×5 contiguity filer is used as the neighborhood definition in this case.

The RUSLE(Yuan, 2008)is the most well-known soil erosion modeling tool; relatively few parameters are needed to evaluate soil erosion. In the RUSLE model, six major factors(rainfall pattern, soil type, slope length, slope steepness, cover system, and management practices)are utilized to compute the average and total erosion. The formula is:

where the result A is the annual soil erosion amount. The main factors in this function are rainfall-runoff erosivity(R), soil erodibility(K), cover management(C), support practice(P), and topography(L, slope length; S, slope steepness). Thus, A can be calculated based on Wischmeier and Smith's algorithm(Hammad et al.,2004; Liu et al.,2009):

where P_i is monthly total rainfall(mm) and P is annual rainfall. The values of parameter R are identical for all fields, in cases small areas like a coal mining area. K is the soil factor, which can be obtained through reference tables and digital elevation model(DEM)information. L is the slope-length factor, and can be acquired by Equation (4)(Qin et al.,2010):

where λis slope length and m is the slope-length index; m can be obtained by:

where β is the ratio of rill erosion and gully erosion. Parameter βcan be calculated as:

where θis the slope; S is the slope-steepness factor, which can be calculated as follows:

where s is the percentage of slope and θ is the slope.

C is the cover and management factor, which can be obtained through a regression equation of the vegetation fraction(VF)as follows(Yang et al.,2008; Qi et al.,2009; Fensholt et al.,2010; Xie and Lin, 2010):

P is the support practice factor, which can be obtained through look-up tables based on the classification results above. The area soil erosion was calculated based on RUSLE, which can provide a scientific basis for soil erosion monitoring.

Figure 1 Flow chart of analysis method adopted in this paper

Figure options

The exurb of Xuzhou City was selected as the test area, which is located in eastern Jiawang and western Tongshan counties(256.581 km²). Multi-temporal land sat TM/ETM+ data captured on April 3 in 2001, May 11 in 2003, April 14 in 2005, May 4 in 2007, May 3 in 2009, and March 30 in 2011, were obtained and a 1:150, 000 topographic map and the vector map of this study area were also employed in the study.

The land sat TM7 data obtained on April 14, 2005 and March 30, 2011 were scan-line corrected based on imagery captured on April 14, 2005 and March 5, 2009, respectively, using local linear histogram matching methodology. After radiometric correction, we used the polynomial method to register by image-image mode, and the accuracy of the pixel root mean square(RMS)was 0.5 pixel. The parameters of the SVM classifier were assigned as 0.143 gamma genes, 100-penalty parameter, and 0 pyramid levels. The results of that overall accuracy and the kappa coefficients are shown in Table 1.

Based on SVM classification results, a matrix of transition probabilities was calculated by the CA_Markov model, and then the land cover maps in 2001, 2003, 2005, and 2007 were taken as the basis land cover image. The cellular automata was assigned a st and ard 5×5 contiguity filter, and the number of cellular automata iterations was set to four. The model run was set at six-month increments. The realization process was as follows:

1)The initial data acquisition and processing. The SVM classifier was selected to obtain the land cover types in the test area.

2)Determining the conversion rules of cellular automata. In this step, GIS analysis methods were chosen to calculate the matrix of transition probability and conditional probability; this matrix was set as a conversion rule.

3)Prediction model based on CA_Markov. It is essential to determine the initial state, number, and filter type of the cellular automata in order to use the CA_Markov model. In this experience, the land cover maps in 2001, 2003, 2005, and 2007 were selected as initial state, the number of cellular automata was set as four, and the land scape maps of this research area in 2005, 2007, 2009, 2011, and 2013 were simulated respectively(Figure 2, Table 2).

Figure 2 Simulation results in 2005, 2007, 2009, 2011, and 2013r

Figure options

DEM information downloaded from SRTM with 90-m resolution was used to estimate soil erosion, combined with land cover maps obtained from the SVM classifier and the CA_Markov simulation result in RUSLE. The same value of parameters R, L, S, and K factor were set for all the land cover types in this area based on similar small-study areas in other researches(Hammad et al.,2004; Fu et al.,2005). Based on the classifications in 2001, 2003, 2005, 2007, 2009 and 2011, the P factors were obtained. Thus, the total soil erosion in the mining area from 2001 to 2013 was calculated(Figure 3, Table 3).

Figure 3 Soil erosion in 2001, 2003, 2005, 2007, 2009, and 2011

Figure options

The results(described as Table 1)demonstrate that the SVM classifier had a good ability to obtain land cover maps in this test area, the overall accuracy of classification is higher than 95%. Such land cover maps acquired based on an SVM classifier wereconducive to land scape analysis and the follow-up status of land scape change simulation. Based on the classification results by SVM, the CA_Markov model could simulate and predict the LUCC trend efficiently and accurately. On the basis of supervised classification results, the CA_Markov simulation result, and other additional information, the RUSLE model could get a better outcome of soil erosion. The amount of soil erosion in the existing conditions, and the future soil erosion condition combined with the CA_Markov model, were both estimated.

As can be seen from the outcome of map of soil erosion(shown as Figure 3), our analysis of the effects of the R, C, and P factors in the RUSLE demonstrated that the C factor had an important impact on soil erosion; the lower the value of parameter C was, the less the ground cover soil erosion there was. We also found that soil erosion was closely correlated with the land cover types, and the land cover types in this study area were mainly farmland and slope. Our analysis of the trend of soil erosion suggested that soil erosion grade in this study area was fairly minimal(level)during these research periods, and the trend is one of continued decline.

Figure 3 and Table 3 show that the largest erosion grades were distributed along the river on steep upstream slopes with little vegetation, due to loose surface soiland lack of vegetation protection, but there was considerable variation in the amounts of soil erosion all over this test area. From these land cover and soil erosion maps, we can also conclude that soil erosion was closely correlated with land cover types. The soil erosion intensity of low-vegetation fractions(building areas)was much greater than in land cover types with ample vegetation.

Table 3 illustrates that the soil erosion grade in the study area was mainly minimal(level)from 2001 to 2013. There was a declining soil erosion trend from 2001 to 2013: the average soil erosion was about 3.05 t/(km²·a)in 2001, 2.55 t/(km²·a)in 2003, 1.54 t/(km²·a)in 2005, 2.72 t/(km²·a)in 2007, 2.30 t/(km²·a)in 2009, 0.85 t/(km²·a)in 2011, and 0.65 t/(km²·a)in 2013. Thus, in general, this mining area is at low risk of soil water erosion.

This research also demonstrates that our proposed framework for soil erosion change monitoring, from spatialand temporal aspects based on multi-temporal remotely sensed data, is better adapted and more efficient than traditional methods of soil erosion monitoring. It also has the potential to analyze relationships between soil erosion amount and different ground land scape types, and can simulate soil erosion change trends which other methods cannot achieve.

Soil erosion is a complex and slow process, so in order to obtain better calculation and prediction results it is necessary to be improve in the following areas. There must be access to comprehensive, detailed social attribute data, such as the rainfall amount, and information on support practices such as land contouring versus strip cropping; accurate topographic data, e.g., DEM, DSM; and excellent classifiers to obtain accurate land cover types, such as a classifier ensemble. Also, the simulation accuracy should be improved by using multi-temporal, higher-spatial-resolution land cover data, with smaller cell sizes.

This paper was supported by the Fundamental Research Funds for the Universities of Henan Province(NSFRF140113), the Jiangsu Provincial Natural Science Foundation(No. BK2012018), the Natural Science Foundation of China(No. 41171323), the Special Funding Projects of Mapping and Geographic Information Nonprofit research(No. 201412020), a joint project of the National Natural Science Foundation of China and the Shenhua Coal Industry Group Co., Ltd.(No. U1261206), and the Ph.D. Fund of Henan Polytechnic University(No. B2015-20) and the youth fund of Henan Polytechnic University(No. Q2015-3). The authors also give sincere thanks to Prof. Paolo Gamba from the University of Pavia, Italy, for his suggestions for this research.