*Article* **GIS-Based Evaluation of Landslide Susceptibility Models Using Certainty Factors and Functional Trees-Based Ensemble Techniques**

#### **Xia Zhao 1,2 and Wei Chen 1,2,\***


Received: 15 October 2019; Accepted: 25 November 2019; Published: 18 December 2019

**Abstract:** The main purpose of this paper is to use ensembles techniques of functional tree-based bagging, rotation forest, and dagging (functional trees (FT), bagging-functional trees (BFT), rotation forest-functional trees (RFFT), dagging-functional trees (DFT)) for landslide susceptibility modeling in Zichang County, China. Firstly, 263 landslides were identified, and the landslide inventory map was established, and the landslide locations were randomly divided into 70% (training data) and 30% (validation data). Then, 14 landslide conditioning factors were selected. Furthermore, the correlation analysis between conditioning factors and landslides was applied using the certainty factor method. Hereafter, four models were applied for landslide susceptibility modeling and zoning. Finally, the receiver operating characteristic (ROC) curve and statistical parameters were used to evaluate and compare the overall performance of the four models. The results showed that the area under the curve (AUC) for the four models was larger than 0.74. Among them, the BFT model is better than the other three models. In addition, this study also illustrated that the integrated model is not necessarily more effective than a single model. The ensemble data mining technology used in this study can be used as an effective tool for future land planning and monitoring.

**Keywords:** landslide susceptibility mapping; ensemble techniques; functional trees; bagging; rotation forest; dagging

#### **1. Introduction**

A landslide is a complex natural phenomenon [1]. It is influenced by many geological environmental factors, such as topography, landform, geology, land use, and vegetation [2]. A landslide is one of the most familiar and disastrous geological hazards with great destructiveness, which always poses a serious threat to human life, property, and living environment, and restricts human progress and development, especially when geological environments are increasingly affected by human engineering activities [3]. Therefore, landslide prediction is of great significance for landslide prevention and control [4,5]. One of the greatest tasks of landslide disaster and risk mitigation is to prepare landslide susceptibility maps [6].

With the development and progress of the geographic information system (GIS), its application in spatial analysis of landslides is becoming more and more popular. With proper use of GIS, most of the landslide susceptibility mapping methods can realize the automation of evaluation and standardization of data management technology, and enable us to build more efficient and accurate maps [7,8]. This is because these technologies can obtain, query, store, analyze, manipulate, and display a set of spatial and non-spatial data about landslide conditioning factors [8–10]. Landslide

susceptibility zoning mapping technology includes a variety of statistical techniques and statistical methods, including Dempster–Shafer [11–13], entropy [14–16], logistic regression [17–19], certainty factors [20–22], statistical index [23,24], analytic hierarchy process [25–27], frequency ratio [20,28], weight of evidence [29–32], index of entropy [20,33], multivariate adaptive regression spline [34–36], and evidential belief function [37–39].

Landslide susceptibility mapping is a typical complex nonlinear problem in a large area of a landslide research area [5]. Thus, the results obtained by statistical techniques and statistical methods may not be able to achieve satisfactory accuracy [5,40]. Later, many researchers proposed a large number of machine learning techniques for evaluating the susceptibility of landslides, which usually have high prediction accuracy and better performance in data-driven models, such as naive Bayes [41–43], random forests [2,44–46], artificial neural networks [47–50], kernel logistic regression [51,52], support vector machine [53,54], and decision trees [55,56]. However, the performance of machine learning methods is generally influenced by the quality and quantity of training data, and the dependence on modeling parameters is very high [5,57]. So far, it is not clear which method is most suitable for landslide susceptibility mapping [5].

In recent years, hybrid technology is considered to be more effective than single technology [58]. In order to explore more reasonable and perfect research results, a variety of integrated algorithms have been developed for landslide susceptibility modeling [6], such as adaptive neuro-fuzzy inference system [59,60], artificial neural networks-Bayes analysis [61], and Evidential Belief Function-fuzzy logic [62]. The important capability of the integrated model is that the method is more accurate in identification and greatly improves the prediction ability compared with the single machine learning model [6].

The purpose of this study is to propose and validate the ability and effect of ensemble techniques in landslide susceptibility modeling, and functional trees are selected as the base classifier to ensemble with bagging, rotation forest, and dagging models in Zichang County (China). Receiver operating characteristics (ROCs) and statistical parameters were used to evaluate and compare the overall performance of the four models.

#### **2. Study Area and Data Used**

#### *2.1. The Study Area*

Zichang County is located in the north of Yan'an City, Shaanxi Province, China, between longitudes 109◦11 58" E and 110◦01 22" E and between latitudes 36◦59 30" N and 37◦30 00" N, with a total area of 2405 km<sup>2</sup> (Figure 1). Zichang County is a typical hilly and gully region of the Loess Plateau. The terrain is tilted from northwest to southeast, with an elevation of 933 to 1574 m. Zichang County prevails a warm temperate semi-arid continental monsoon climate, with low temperature and large temperature difference. The annual average temperature within the territory is 9.1 ◦C, the annual average precipitation is 514.7 mm. The rivers in the territory belong to the Yellow River system, which is divided into three tributaries: Qingjian River, Wuding River, and Yanhe River.

#### *2.2. Data Preparation*

The quality of landslide inventory is very significant for landslide susceptibility modeling, and an accurate landslide inventory map is the foundation of landslide susceptibility modeling [63,64]. In this study, three techniques were used to improve the reliability and accuracy of the landslide inventory map: historical report, aerial photo interpretation, and field survey using Global Navigation Satellite Systems (GNSS). According to the landslide inventory map in this area, 263 landslides were identified, and 184 landslide locations (70%) were randomly sampled as the training data and the other 79 landslide locations (30%) were used to validate models.

**Figure 1.** Study area.

After compiling the landslide inventory, it is necessary to choose the landslide conditioning factors to create the landslide susceptibility map [65]. The selection principle is to consider the mechanism and geo-environmental characteristics of landslide occurrence in the study area. Generally, the landslide conditioning factors used to evaluate landslide susceptibility include three categories: topographic factors, geological factors, and environmental factors. In this paper, 14 landslide conditioning factors were selected and transformed into the same resolution (30 × 30 m), including elevation, slope, aspect, plan curvature, profile curvature, sediment transport index (STI), stream power index (SPI), topographic wetness index (TWI), the normalized difference vegetation index (NDVI), land use, lithology, soil, distance to roads, and distance to rivers (Table 1, Figure 2).


**Figure 2.** *Cont.*

**Figure 2.** Thematic maps: (**a**) elevation, (**b**) slope, (**c**) aspect, (**d**) plan curvature, (**e**) profile curvature, (**f**) STI, (**g**) SPI, (**h**) TWI, (**i**) NDVI, (**j**) land use, (**k**) lithology, (**l**) soil, (**m**) distance to roads, (**n**) distance to rivers.

#### **3. Modeling Approach**

The chapter included the illumination of five models, namely certainty factors, functional trees, bagging, rotation forest and dagging. The certainty factors model was used to express the correlation between landslide and conditioning factors, the functional trees model was used as a base classifier, the bagging, rotation forest, and dagging were used as ensemble algorithms.

#### *3.1. Certainty Factors*

The certainty factor (CF) belongs to a probability function, which was first proposed in 1990 [66] and modified subsequently [67]. The certainty factor can be expressed as [68]:

$$CF = \begin{cases} \begin{array}{c} \frac{PPa-PPs}{PPa(1-PPs)} \\ \frac{PPa-PPs}{PPs(1-PPa)} \end{array} & \text{if} \quad PPa \ge PPs \end{cases} \tag{1}$$

where, *PPa* is the conditional probability of landslide in class *a* in study area A, *PPs* is the prior probability of the total number of landslides in study area A.

The range of *CF* is −1 to 1, the positive value indicates that the degree of certainty of landslide occurrence increases, while the negative value indicates that the degree of certainty of landslide occurrence decreases [69–71].

#### *3.2. Functional Trees*

Functional trees (FT) are a combination of a discriminant function and multivariable decision tree through constructive induction [72]. Functional trees use logistic regression functions to calculate the splitting of internal nodes (called oblique splitting) and estimation of leaves [73–75]. FT learns the classification tree based on the attributes of leaf nodes, decision nodes or nodes and leaves [38,76]. The decision nodes are built while the trees are growing, while the functional leaves construct when the trees are pruning [76]. Functional trees have the following three usage types: (1) the full functional tree using a regression model for internal nodes and leaves; (2) function tree internal-only uses the regression model for internal nodes; (3) functional tree leaves only use the regression model for leaves [75,76].

In the leaf logic regression function, the logic enhancement (iteration are weighted) of the least-squares function is determined for each output consisting of two classes [77]. Among them, training datasets of D and *n* samples (*Ai*, *Bi*) with *Ai* <sup>∈</sup> *Rn*, *Bi* <sup>∈</sup> {1, 0} [76]. *Ai* is the input vector containing all landslide condition factors [75,76]; whereas *P*(*A*) is the probability prediction value of landslide occurrence; *Bi* is the coefficient of the *i* component of the input vector *Ai*. The posterior probability *P*(*A*) of the left ventricle is calculated as follows [78]:

$$f\_{B\_i}(A) = \sum\_{i=1}^{14} B\_i A\_i + B\_0 \tag{2}$$

$$P(A) = \frac{\mathbf{e}^{2f\_{\mathbb{R}\_i}(A)}}{1 + \mathbf{e}^{2f\_{\mathbb{R}\_i}(A)}}\tag{3}$$

#### *3.3. Bagging*

Bagging is based on the concepts of bootstrapping and aggregating, which is used to obtain a more robust and accurate landslide model. Bagging is one of the most popular integration algorithms [79]. The process of a bagging algorithm includes:

Firstly, the bootstrap samples *S*(*xi*, *yi*) are randomly resampled from a training set (*xi*, *yi*), forming a set of training subsets, where, *xi* ∈ *R*, *yi* ∈ (landslide, non-landslide) [80]. Then, several models based on a classifier are constructed according to each subset, *Ci*(*x*) is a classifier constructed from each guiding sample. All models based on classifier (*Li*) are aggregated to generate the final model (*L* ), where, *L*1, *L*2, ... , *Ln* generates a combined classifier (*L* ). *L* predicts the class label of a given instance *x* by calculating the votes using the following equation [81]:

$$L'(\mathbf{x}) = \underset{y \in Y}{\text{arg}\max} \sum\_{i=1}^{t} L(\mathbb{C}\_i(\mathbf{x}) = y) \tag{4}$$

#### *3.4. Rotation Forest*

Rotation forest (RF) is a popular aggregation technique proposed by Rodriguez et al. [82]. RF is an effective technique for improving weak classifiers [83]. It uses principal component analysis (PCA), a multivariate technique used for analyzing large multivariate datasets, to reduce its dimensions [84]. In this method, features are extracted from the learning (training) dataset and a base classifier is used to generate learning sub training dataset [82].

For the use of RF: randomly divide the training dataset into D subsets, where D is the parameter of the algorithm, and construct the rotated sparse matrix by performing feature extraction for each subset. The classifier is based on the feature of a repeated matrix projection, and the result is obtained by combining the output of multiple classifiers [84]. RF can be used with any basic classifier, and the feature extraction of each classifier retains all the features that promote variability [84].

In the RF algorithm, *x* = (*x*1, *x*2, ··· , *xn*) is the training sample set, and Y is the corresponding class label, that is used to consider landslides and non-landslides; *D*1, *D*2, ··· , *DL* are the classifier in the set frame; and P is the set of landslide condition factors. The coefficients of the rotation matrix *Ri a* are obtained by transformation and base classifier. Obtain *Ri <sup>a</sup>* by rearranging *Ri* matrix [84]:

$$R\_i = \begin{bmatrix} b\_{i,1}^{(1)} \dotsm \,\_r b\_{i,1}^{(M\_1)} & 0 & \dotsb & 0\\ 0 & b\_{i,2}^{(1)} \dotsm \,\_r b\_{i,2}^{(M\_2)} & \dotsb & 0\\ & \vdots & \vdots & \ddots & \vdots\\ 0 & 0 & \dotsb & b\_{i,K}^{(1)} \dotsm \,\_r b\_{i,K}^{(M\_k)} \end{bmatrix} \tag{5}$$

For each sub training dataset extracted by the rotation matrix *Ri <sup>a</sup>*, average grouping method is adopted to obtain the coefficients of each class in a given test sample [85]:

$$\mu\_{j}^{(\mathbf{x})} = \frac{1}{L} \sum\_{i=1}^{L} d\_{ij}(\mathbf{x} \mathbb{R}\_{i}^{\mathbf{a}}), \ j = 1, \ldots, \mathbf{c}. \tag{6}$$

where μ*<sup>j</sup>* (*x*) is the maximum confidence specified on the class, classifier probability allocation *Di*, and the *dij* regression *dij*(*xRi <sup>a</sup>*) [85].

#### *3.5. Dagging*

Dagging is a well-known resampling integration technique originally proposed by Ting and Witten to generate many disjoint hierarchical folds from a dataset, and each data partition can be sent separately to the basic classifier [86]. The final forecast is based on a majority vote [86]. The main principle is to use a majority vote to combine multiple classifiers to improve the prediction accuracy of the basic classifier [86].

For a given training dataset, which has *n* samples, the dagging algorithm constructs M datasets (M is a free parameter) from the original training dataset [87]. Each dataset contains *n* samples [87], and no two datasets have the same sample. A basic classifier is trained for each dataset to build a classification model [87]. Therefore, the M dataset can be summarized into M classification models [86,87].

#### **4. Results**

This section consists of the detailed description of the results of the present study, which includes the following four sections: (1) the correlation between landslide and conditioning factors, and then the CF values are used as input to weight the classes of conditioning factors; (2) selection of landslide conditioning factors that are positive to the modeling process; (3) application of four hybrid models and generate landslide susceptibility maps; and (4) validation and comparison of models using ROC and Chi-squared methods.

#### *4.1. Correlation Analysis of Landslide and Conditioning Factors Using the CF Method*

The landslide density at each class was calculated by combining each thematic map and landslide inventory map. Meanwhile, this paper summarizes the spatial relationship between the landslides and conditioning factors using the CF method (Table 2). According to the calculation results in Table 2, the highest CF value (0.661) is found in the elevation category of 1500–1574 m, which indicates that the probability of landslide is the highest. Among the six classes classified by the slope, 40◦–50◦ (0.324) is the highest CF value of the six categories. As far as aspect is concerned, the CF values of slopes facing south (0.309) and southwest (0.242) are the largest. Among the five classes classified by plan curvature, the classes of (−9.24)–(−1.79) have the lowest CF value (−0.495), and the classes of 1.44–7.56 have the highest CF value (0.244). Among the five classes classified by profile curvature, the classes of (−1.65)–(−0.46) have the lowest CF value (−0.346), and the classes of 0.58–1.97 have the highest CF value (0.277). For STI, the frequency of landslide occurrence is the most relevant in 20–30 categories, with the largest CF value (0.220). In TWI, the CF value is the largest in the classes of 2–3 (0.164) and the smallest in the classes of >5 (−1). For NDVI, the lowest CF value (−0.326) was found in the classes of 0.01–0.04, and the highest CF value was found in the categories of 0.07–0.09 (0.223). In terms of land use, landslides mostly occur in residential areas (0.465). Among the five types of lithology, the groups 2 and 4 were relatively more sensitive to landslide occurrence, with CF values of 0.430 and 0.465, respectively. For soil, the majority of landslides occurred in red clay soils with a CF value of 0.712. It can be seen from a distance to roads that the closer the distance is, the more sensitive the landslide. CF value is the largest in the categories of 0–100 m (0.452). For distance to rivers, CF value is the largest in the categories of 0–200 m (0.585).


**Table 2.** Relationship between landslides and conditioning factors using the certainty factor (CF) method.


**Table 2.** *Cont.*

#### *4.2. Selection of Landslide Conditioning Factors*

In order to ensure the accuracy of landslide prediction results, it is necessary to remove unimportant or unrelated factors [88,89]. In this study, the Pearson correlation method [90,91] with 10-fold cross-validation was used as an effective feature selection method for evaluating the predictive ability of conditioning factors. The distance to rivers, slope, and lithology has the highest predictive abilities (Table 3). Since a no conditioning factor has a null predictive value, all are included in this analysis.

#### *4.3. Application of Landslide Susceptibility Models*

In this study, the training data and CF values were used to construct four models, namely the functional trees (FT) model, bagging-functional trees (BFT) model, rotation forest-functional trees (RFFT) model, and dagging-functional trees (DFT) model, respectively. To get the best performance of the model, the iteration times of the FT model and the minimum number of instances considering the separation of nodes from the training dataset are optimized to 15 and 36, respectively. When building the BFT, RFFT, and DFT models, the two parameters mentioned above were fixed firstly. After completing the above work, the optimized models were applied to the whole research area to create landslide susceptibility maps. The calculated landslide sensitivity index (LSI) values can be interpreted as the probability in the range of 0 and 1, and all LSI values can be converted to ArcGIS to generate the final landslide susceptibility map.


**Table 3.** Correlation attribute of landslide conditioning factors.

Four landslide susceptibility maps generated by FT, BFT, RFFT, and DFT models are shown in Figure 3a–d respectively. The landslide susceptibility maps were reclassified into five classes, namely very low, low, moderate, high, and very high using the natural break method [92]. The comparison of area sizes for each category of the four models is shown in Figure 4. For the FT model, the largest area is the very low class (27.92%), followed by high class (23.47%), very high class (20.21%), low class (17.55%), and the smallest area is the moderate class (10.86%). For the BFT model, the percentages of very low, low, moderate, high, and very high classes are 24.02%, 22.87%, 19.88%, 18.10%, and 15.12%, respectively. The results of landslide susceptibility zoning using the RFFT model show that these percentages are 37.62% (very low), 21.41% (low), 7.79% (moderate), 12.25% (high), and 20.93% (very high), respectively. For the DFT model, the percentages of very low, low, moderate, high, and very high classes are 19.70%, 30.59%, 23.72%, 16.50%, and 9.49%, respectively.

#### *4.4. Model Performances and Comparisons*

In this study, the landslide susceptibility models were evaluated by using the areas under the ROC curves (AUC), standard error, 95% confidence interval, and significance level *p*-value. The ROC curve can be used as a useful tool to indicate the quality of deterministic and probabilistic prediction system [93–95]. The sensitivity (true positive rate) is shown as y-axis and 1-specificity (false positive rate) as x-axis [94,96]. The AUC values are in the range of 0.5 to 1 [97], and the excellent attributes of the model increase with the AUC values [98].

Using the training dataset, the performance of the landslide susceptibility models was evaluated (Table 4). The BFT model has the highest AUC value (0.947), the lowest standard error (0.011), and the narrowest 95% confidence interval (0.925–0.969). It is followed by the RFFT model, the FT model, and the DFT model. For the validation data, the calculation results are shown in Table 5. The BFT model has the highest AUC value (0.804), the lowest standard error (0.035), and the narrowest 95% confidence interval (0.736–0.871). It comes before the DFT model, the FT model, and the RFFT model. These results show that all performance in the validation dataset is slightly worse than those of the training data. These results show that the BFT model is the best model among the four models, and the ensemble model is not necessarily superior to the single model.

**Figure 3.** Landslide susceptibility maps: (**a**) functional trees (FT) model; (**b**) bagging-functional trees (BFT) model; (**c**) rotation forest-functional trees (RFFT) model; (**d**) dagging-functional trees (DFT) model.

**Figure 4.** Percentages of landslide susceptibility classes.


**Table 4.** Parameters of receiver operating characteristic (ROC) curves with the training dataset.

**Table 5.** Parameters of ROC curves with the validation dataset.


A Chi-squared test was used to analyze the significance of the four models (Table 6). It can be seen that only the comparison of FT and RFFT exhibits lower Chi-squared value (0.044) and higher *p*-value (0.834), which indicate no significant difference between the two models. The other five groups all present larger Chi-squared values and lower *p*-values. The significant differences between the models indicate that the differences between the models are good, which is more conducive to the modeling work and enables this study to obtain the susceptibility results smoothly.

**Table 6.** Pairwise comparison of four models.


#### **5. Discussions**

In this current study, the correlation analysis between conditioning factors and landslides was carried out by the CF method. The probability of landslide occurrence is in inverse correlation with elevation. This may be related to local rainfall and loess and may be related to human engineering activities. With the increase of slope angle, the degree of certainty of landslide occurrence decreases. This may be due to the larger slope angle, the less loose material or more weatherproof material. At the same time, it can be observed that most landslides occur on slopes facing south and southwest with the highest probability. This is mainly because more rain and sunshine are available to the south and landslides are prone to occur. The curvature of plan and profile shows anomalous results. The curvature of the plan (near zero) and convex plan (positive value) are highly sensitive. This anomaly may be related to the overweight effect [28,99,100]. In terms of land use, the probability of landslides in residential areas is the largest, which can explain the impact of human engineering activities on landslides. For the lithology, the second group (Tertiary (T): mudstone, conglomerate) and the fourth group (Triassic (T): mudstone, sandstone, songlomerate) are more sensitive to landslide occurrence. There is groundwater flow in the relatively fractured saturated sandstone and fractured conglomerates, resulting in additional load on the mudstone, resulting in landslides [28,101]. The linear characteristics of the road and river buffers are inversely correlated with landslide susceptibility in the distance. Such an important result has been repeated in many kinds of literature [6,102–104]. However, the remaining five variables make little contributions to the occurrence of landslides.

According to ROC curve analysis (Figures 5 and 6) and statistical index analysis (Tables 4 and 5), it can be concluded that the four machine learning methods selected in the training and testing data assemble a very small *p*-value and significant high performance in the 95% confidence interval. The BFT

model has the highest AUC value (0.947), the lowest standard deviation (0.011), 95% confidence interval (0.925–0.969), and *p*-value (<0.0001). However, the DFT model has the worst results in this study area. The DFT model has the lowest AUC value (0.797), the highest standard deviation (0.023), 95% confidence interval (0.752–0.842), and *p*-value (<0.0001). There is no doubt that most ensemble models are superior to single models. However, there is still a phenomenon that the performance of hybrid machine learning methods is not always better than a single model. In order to find more optimal solutions, much more different set models should be applied to the research field.

**Figure 5.** ROC curves of the models using the training dataset. AUC: area under the curve.

**Figure 6.** ROC curves of the models using the validation dataset.

According to the paired comparison of the performance of the models (Table 6), the Chi-squared test shows that the Chi-squared values are relatively large. Among them, the Chi-squared value of the FT and RFFT models is smaller, the *p*-value is larger, and the difference between these two models is not significant. The good results obtained from the other three groups can serve as a powerful basis for modeling in this study. At the same time, the BFT model is compared with the other three models in pairs, and the difference is significant. According to the evaluation results of various evaluation criteria, the performance of the BFT model is better than that of the RFFT model, FT model, and DFT model. As a final recommendation, the obtained results can be useful for policy planning and decision-making in areas prone to landslides. The proposed BFT model, based on performance and prediction accuracy, is suggested in the study area and other regions over the world where they have similar geo-environmental conditions with a logical caution.

#### **6. Conclusions**

This study applied functional tree-based ensemble techniques (FT model, BFT model, RFFT model, DFT model) for landslide susceptibility spatial modeling in Zichang County, China. Fourteen conditioning factors and the occurrence of landslides were used to analyze the correlation. Meanwhile, the ROC curve and statistical parameters were used to evaluate and compare the accuracy of the model results. The results showed that the prediction rate of the BFT model is the highest. Therefore, the BFT model is the best optimization ensemble model in this study, and it can be used as an advantageous and promising method for landslide susceptibility modeling. Finally, the landslide susceptibility map generated by this study can be used as an effective tool for future land planning and monitoring by government officials or research experts and scholars.

**Author Contributions:** X.Z. and W.C. contributed equally to the work. X.Z. and W.C. collected field data and conducted the landslide susceptibility mapping and analysis. X.Z. and W.C. wrote and revised the manuscript. W.C. provided critical comments in planning this paper and edited the manuscript. All the authors discussed the results and edited the manuscript. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was supported by the National Natural Science Foundation of China (Grant No. 41807192), Natural Science Basic Research Program of Shaanxi (Program No. 2019JLM-7, Program No. 2019JQ-094), China Postdoctoral Science Foundation (Grant No. 2018AT111084, 2017M61318), and project funded by Shaanxi Province Postdoctoral Science Foundation (Grant No. 2017BSHYDZZ07).

**Conflicts of Interest:** The authors declare no conflicts of interest.

#### **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

## *Article* **Spatiotemporal Dynamics and Obstacles of the Multi-Functionality of Land Use in Xiangxi, China**

**Hui Xiang 1,2, Qing-Yuan Yang 1,2,\*, Kang-chuan Su 1,2 and Zhong-Xun Zhang 1,2**


Received: 28 June 2019; Accepted: 22 August 2019; Published: 4 September 2019

**Abstract:** The multi-functionality of land as the basis of land use and utilization is under increasing investigation. This study assesses the spatiotemporal dynamic multi-functionality of land use and analyzes the obstacle indicators in Xiangxi using two methods, i.e., the analytic hierarchy process (AHP) and the hierarchical weighting method (HWM). First, we found that the total function level of land use in Xiangxi was constantly optimized. Spatial heterogeneity was clearer. Land use had a trend toward diversification, with a focus on production or living function. The coordination of multifunctional land use has undergone certain changes. It was more apparent in the south than in central and northern regions. Second, we discovered that production function of land use in Xiangxi grew slowly and spatial differentiation was enhanced. The living function fluctuated with the trend of spatial equilibrium. Changes in ecological function and any spatial differentiation were not clear. Third, land use can be divided into living-ecological, production-living, and production obstacle types. Lastly, we state that, to narrow the gap between urban-rural areas and reduce the non-point pollution from agriculture in living-ecological barrier areas, we need to develop production and social public utilities in production-living barrier areas, and develop production and eliminate poverty in production barrier areas.

**Keywords:** land use; multi-functionality; production-living-ecology function; spatiotemporal dynamics; obstacle factors; Xiangxi

#### **1. Introduction**

Land is a complex system including topography, soil, hydrology, biology, climate, and other elements [1]. In addition to providing food, fresh water, and other material resources for humans, it is also essential for habitation, transport, leisure, and other activities [2]. Land use reflects the type and intensity of human activities, which directly affects the biodiversity of ecosystems [3], water security [4], and human health [5]. The changes of land use are dominated by human activities [6], which, in turn, affects people's livelihood and sustainable development of the economy. Several models have been used to assess land use change [7,8], but most of them have only focused on the economic or ecological functions of land [9]. Land resources are multifunctional, which is not only an important factor of the ecological environment, but also the main resources of human production and life. Land multi-function evaluation is a very effective method that takes economic, social, and ecological factors into consideration [10]. The concept of multi-functionality in land use originated from agriculture and refers to the ability of land to provide diversified products to meet various needs. This concept has gradually expanded into the non-agricultural sectors. In 2004, the SENSOR project, under the sixth framework of the European Union, considered that the function of land use refers to the various uses of land for the provision of multiple products and services [11], to meet the needs of humans for economic, environmental, cultural, and social services. The multi-functionality of land as the basis

for land use and allocation is an important factor affecting regional land decision-making and spatial planning, and is related to the sustainable development of society and economy.

Research concerning the versatility of land use have become increasingly abundant. These studies have mainly focused on the types of land function, the multiple functions of agriculture, and the evolution of land function. Land multi-function includes regulation function, habitat function, production function, information function, and so on [12]. Scholars have generally discussed the relationship between agricultural versatility, agricultural policy [13], and sustainable development, believing that the main functions for agricultural land were food production and environmental sustainability [14], which had a positive impact on the life of the higher social capital class [15]. An index system for quantifying the versatility of agricultural land was constructed [16,17]. Based on results of this evaluation, the land functions were divided into different areas [18], and the zoning optimization strategy was proposed. Scholars have analyzed the rubber agroforestry system in Sumatra and China [19], concluding that the factors restricting multifunctional change are complex and diverse. A survey in Australia has found that ranches have changed from a single production function to a variety of functions [20], such as protecting biological communities and developing tourism. The study has discovered that family farms in Missouri provide different leisure services and play a diversified role in land use [21]. Since the 21st century, the land use of parks in Poland has undergone significant changes [22]. The reasons for the changes of forest landscape functions in Southwest Parks of Poland include population change, intensive agriculture, urbanization, and land-use policy change [23]. Evidence from Denmark shows that recreational hunting contributes to multi-functional maintenance and change of land use [24].

It can be seen from the above that scholars have carried out extensive studies concerning the types of functions, the current multi-functional use of agricultural land, and have made beneficial attempts to understand the spatio-temporal evolution of multi-functional land use. However, research concerning multi-functional dynamic changes to land use focus mainly on macro (national) and middle (provincial) scales, with insufficient research carried out on a micro (county) scale, and there are few discussions regarding the ethnic regions at the county-scale. Most research studies are taken from the perspective of the dimensions of time or space. Therefore, research using a combination of space and time with regard to land use is required. In addition, land use change has also brought a series of adverse effects [25,26], which restricts the sound development of the land systems. For measuring the function of land, it is also necessary to clarify factors that create obstacles for the versatile use of land, but few studies have conducted such analyses. Xiangxi is typically representative of less developed areas in China, and it is also a minority-concentrated area. Based on the land-use history of this area, we arrive at three questions.


To answer these questions, bases on the logic of "analysis of functions-diagnosis of barrier factors-policy recommendations," this study is conducted from the following aspects. First, we classified the land function types and constructed an evaluation index system. Second, we evaluated the temporal and spatial dynamics of land multifunction in the last five years. Third, we explored the obstacles in depth. Lastly, through comparative analyses between Xiangxi and other regions, we put forward the corresponding policy recommendations.

The aims of this study are: (1) to realize efficient positioning, quantitative expression [27], (2) identify obstacles to constrain land use efficiency in multifunctional land use, and (3) provide policy suggestions for the rational land use in minority areas.

#### **2. Data and Methods**

#### *2.1. Research Area*

Xiangxi (full name: Xiangxi tujia and miao Autonomous Prefecture) is located in the northwest part of the Hunan province (Figure 1). It is bordering the Hubei, Guizhou, and Chongqing provinces. It is the bridge connecting the central and western regions, and the link of communication between the Han and minority nationalities. Geographically, Xiangxi lies between 109◦10 ~110◦22.5 E and 27◦44.5 ~29◦38 N. The total area of the region is approximately 1.547 million hm2, accounting for 7.3% of the Hunan province. It has seven counties and one city under its jurisdiction (Fenghuang county, Luxi county, Huayuan county, Guzhang county, Baojing county, Yongshun county, Longshan county, and Jishou city—a county-level city). Xiangxi is a typical winding territory of Wuling mountain, with high altitude terrain in the northwest, and low terrain in the southeast. Ethnic minorities live in compact communities in Xiangxi. In 2017, 80.73% of the total population of Xiangxi belonged to ethnic minorities, mainly Tujia and Miao.Xiangxi, with a relatively backward economy, which is one of the 14 areas included in the national pilot project for poverty alleviation and is currently in a critical period of new rural construction and modern industrial development. From 2013 to 2017, the economic and social development in Xiangxi increased rapidly, with an average annual GDP growth rate of 7.82%, a per capita GDP growth rate of about 7.26%, and an increase in the urbanization rate from 38.8% to 44.97%. The regional ecological environment is superior, with 70.24% forest coverage in 2017. The land use in Xiangxi is dominated by agriculture, with the area of land used for agriculture reaching 1.4118 million hm2 (accounting for 91.26%), and an area of 0.0555 million hm<sup>2</sup> has been used for construction (accounting for 3.59%). Driven by economic interests, the land use structure has changed significantly, and the area of land used for construction has increased dramatically. For example, the total urban land in 2016 was 961.34 hm2, which represented an annual increase of 15.8%. Therefore, the efficient use of land in Xiangxi currently faces two major problems. First, the agricultural land area is large but the production efficiency is low. Second, urban construction is limiting the space available for production and ecology, and the contrast between land supply and land demand is increasingly prominent. Therefore, it is necessary to improve the function of production in land use in Xiangxi, and to raise people's awareness that changes in the structure of land use lead to changes in land function.

#### *2.2. Data Sources*

The data used in this study concerning the current situation of land use comes from the file issued by the Xiangxi state bureau of land and resources "Suggestions for general land use planning (2006–2020)." The DEM digital elevation data comes from the geographical spatial data cloud (http://www.gscloud.cn/), and all other data (population, land area, land for traffic area, the output value of farming and animal husbandry, etc.) comes directly or indirectly from the Xiangxi statistical Yearbook (2013–2017) and the Hunan statistical Yearbook (2014–2018) (Xiangxi statistical Yearbook is based on the data of that year, and Hunan statistical Yearbook is based on the data of the previous year).

**Figure 1.** Map of the study area.

#### *2.3. Research Methods*

#### 2.3.1. Classifying the Multifunctional Types of Land Use

The land system is composed of economic, social, and ecological subsystems [28], and the land is rendered whole with comprehensive functions via the organic coupling of each subsystem [29]. This study is based on the economic, social, and environmental dimensions of sustainable development, combined with the idea of a national "production-living-ecology (PLE)," based on the optimization of land use, by subdividing the total functions (TF) for which land is used as PLE sub-functions (Table 1): production functions (referred to as PF), living functions (LF), and ecological functions (EF). Production functions are based on the key needs for human survival. Land provides agricultural products to guarantee the capacity of non-agricultural economic output and transportation. These functions are, therefore, measured as the three aspects of agricultural production, economic growth, and transportation security [30]. Living functions are the ability of the land to meet the needs for human development [31,32], which are mainly reflected by four aspects: employment support, social security, cultural leisure, and the residential home. Ecological functions are related to the high-quality production and living needs of human beings. This function is evaluated from the three dimensions of maintaining ecological balance, providing resources, and keeping the environment clean. This study, therefore, divides the function types of land use into three levels from top to bottom: total function, sub-function, and single function. The number of functions at each level is 1, 3, and 10, respectively.


#### **Table 1.** Data sources.

Note: The meaning of X1–X23 is detailed in Table 2.

#### 2.3.2. The Establishment of the Evaluation Index System

The selection of indicators for evaluating multifunctional land use follows the basic principles. (1) Indicators should take the regional development situation into account. For example, the output from agriculture and animal husbandry in Xiangxi accounts for more than 95% (the data of 2017) of the regional output of agriculture, forestry, animal husbandry, and fisheries. The two indexes of per capita agricultural output value and per capita animal husbandry output value were, therefore, selected, according to the needs of regional production and development. (2) Indicators should be chosen that can be quantified and easily obtained. Most indicators selected come either directly or indirectly from the public websites of government departments, which can be directly accessed. (3) The indicators were independent and complementary (Sun et al., 2017). For example, the employment security function for rural and urban land was, respectively, represented by the number of rural employees and the average annual wage of urban employees. (4) Using direct effective indicators. Specific indexes were selected that could minimize the total quantity of indexes, which directly reflects the functional level of land. From the above principles, a total of 23 typical sensitive indicators representing land use function in Xiangxi were selected (Table 2).

#### 2.3.3. Determination of the Index Weight

The analytic hierarchy process (AHP) was used to calculate the weight of each factor. Through modeling and quantitative analysis, the AHP simplifies complex problems and is widely used in the field of land evaluation.

The software of Yaahp (full name is yet another AHP) was invented by Zhang Jianhua. It can be download from this website: http://www.yaahp.cn. It is simple and efficient, and widely used to determine weights [33,34].

First, Yaahp v.10.3 software is used to build a four-level hierarchical structure model, which is a function-subfunction-single function-index layer.

Second, the evaluation factors are scored. The judgment matrix Amk was constructed. Experts were invited to evaluate the relative importance of two factors (m, k) at the same level. The evaluation results were divided into five levels: absolutely important, very important, relatively important, slightly important, and equally important. They were assigned 1, 3, 5, 7, and 9 points, respectively. If the evaluation results landed in the middle, the median value was taken.



*Appl. Sci.* **2019**, *9*, 3649

The third step involved consistent checking and weight calculation. To test the rationality of the pairwise judgment matrix, consistency testing is required. If the test value is ≤0.1, it shows that the matrix evaluation is reasonable. However, the results have to be revised. Running software found that the evaluation results passed the consistency test.

Lastly, the weight of all factors was calculated (details in Table 2).

#### 2.3.4. Measuring the Versatility of Land Use

The first step was to standardize the data. According to the land use function evaluation index system, a sample matrix *X* of m evaluation indexes in n areas from 2013 to 2017 was constructed, where *X* = (*Xij*)n×m, *i* = 1,2, ... N, *j* = 1, 2,...M. In order to make an evaluation of the quantitative comparison of different attributes and dimension indexes, the optimal value of each index (the maximum value of the positive index and the minimum value of the negative index) during the study period was selected as the reference value *Xo* in order to conduct a dimensionless quantization of *X*. The calculation formula used was as follows.

$$\mathcal{Y}\_{ij} = \begin{cases} X\_{ij}/X\_{oj} \text{ the positive index} \\ X\_{oj}/X\_{ij} \text{ the negative index} \end{cases} \tag{1}$$

where *Yij* was the standardized value of the *j* evaluation index in area *i*, *Xij* was the original value of the *j* valuation index in area *i*, and *X0j* was the optimal value of the *j* evaluation index. Evaluation samples *Y* were obtained after standard treatment, *Y* = (*Yij*)n×m, *Y* ∈(0 1].

The second step was to calculate a value for the land use function. Value (*F*) represents the value for the level of land use with regard to functions. The larger *F* is, the better the level of land use is, and vice versa. According to the standardized values of the evaluation samples and the weight of factors at all levels, the land single functional value (*F1*), sub-functional value (*F2*), and total functional value (*F3*) were calculated. The formula for calculating the functional values is shown below.

$$F1 = \sum W\_{\bar{j}} Y\_{\bar{i}\bar{j}} \ F2 = \sum W\_{\bar{i}} F1 \ F3 = \sum W\_n F2 \tag{2}$$

where *Wj* was the weight of evaluation index, *Wi* was the weight of a single function (*F2*), and *Wn* was the weight of a sub-function (*F3*).

The third step was to calculate the degree of dynamic change in land use and the degree of functional advantage of the land. The degree of dynamic change in land function (*d*) refers to the degree of change in land function levels within a certain period. The overall degree of dynamic change in land function was, therefore, calculated for the study period of five years. The calculation used was as follows.

$$d = \frac{F\_{t+4} - F\_t}{F\_t} \times 100\% \tag{3}$$

where *t* represents the year, *Ft* represents the land function value of the year *t*, *d* < 0 indicates that the land function remains unchanged, *d* > indicates that the land function is enhanced, and *d* < 0 indicates that the land function is degraded.

The dynamic dominance of land function reflected the differences in land use, calculated with:

$$s = \frac{|d|\_{\max}}{\sum |d|} \tag{4}$$

where |*d*|*max* was the maximum absolute value of the degree of dynamic change in land function, and Σ|*d*| was the sum of the absolute value of the dynamic changes of land function. The larger *S* is, the more diverse the land function change is, which means that the land use tends to be simplified. Smaller values for *S* point to diversification.

The fourth step was to measure the multifunctional coordination degree of land use. The difference in the standard deviation of the functional value of land use (σ) reflects the coordination of the functional level of the land. The calculation formula was as follows.

$$
\sigma = \sqrt{\frac{\sum \left( F - \overline{F} \right)^2}{N}} \tag{5}
$$

where *F* was the average value for land function and *N* was the function number. The value of σ is inversely related to the coordination degree of land function. The higher the value is, the lower the coordination degree of land function is, and vice versa.

#### 2.3.5. Diagnosing the Obstacles to Land Use Function

The degree of land dysfunction is represented by the degree of the total function that has been hindered. The higher the degree of land dysfunction, the stronger its influence. The calculation formulas for the obstacle degree (*Qi*) of a single function and the obstacle degree (*Ki*) of a sub-function are shown below.

$$Q\_i = \frac{\left(1 - \chi\_{ij}\right) \mathcal{W}\_j}{\sum \left(1 - \chi\_{ij}\right) \mathcal{W}\_j} \times 100\% \tag{6}$$

$$\mathcal{K}\_i = \sum Q\_i \tag{7}$$

where (1 − *Yij*) represents the gap between the *i* index of land use function and the goal of function, which is the difference between the index standard value (*Yij*) and the optimal standard value (1).

At last, we placed the calculation results into the attributes of each research units, and created the column chart using the spatial analysis function of software ArcMap v.10.2 (ArcMap is one of the three user desktop components of ArcGIS, and it was developed by the Environmental System Research Institute in 1978. Its official website is http://argmaps.com). The length of the column represented the size of each data, and labeled the values to visualize the results of the research.

#### **3. Results**

#### *3.1. The Spatio-Temporal Dynamic Evaluation of the Total Function of Land Use*

We got the map of total function of land use change in Xiangxi over the past five years (Figure 2), to analyze the trends in spatio-temporal dynamic evolution.

(1) The total function level of land use was constantly optimized and regional differences were found to be greater. During the study period, the value of land function level in Xiangxi showed an upward trend (Figure 2a). The highest increase occurred in Jishou (24.05%), and the lowest (8.90%) increase occurred in Longshan, with the other areas falling in between (10%–20%). In 2013, Xiangxi Autonomous Prefecture committee put forward the developmental policy of "5-4-2" (i.e., the "five constructions," industry, infrastructure, new towns, ecological construction for civilization, "four Xiangxi"—green, civilized, an open and harmonious Xiangxi, "two take the lead," take the lead in development, take the lead out of poverty). Under the lead of an open strategy for Hunan province and land use activities in the area remained stable and improved. With these changes in the total function of land use, the differences in the land function level in Xiangxi were increasingly clear, and the standard deviation in the value for total land function between the counties increased from 0.0344 to 0.0515. On the whole, the land use function level in Xiangxi presented a spatial pattern of high levels in the north and south (Jishou, Luxi, Fenghuang, and Huayuan in the south, Longshan, and Yongshun in the north), and low levels in the middle (Guzhang and Baojing).

(2) The functions for which land was used tended toward diversification, focusing on those associated with production or living. During the study period, the function value for the change of land use in each region was characterized by rapid growth of the lower (production and living) and slow change of the higher (ecological) functions, which reflects the trend in land use diversification. Production and living functions (Figure 2b,c) underwent the largest changes in value, with the highest degree of dynamic changes in production functions observed in Jishou (37.48%) and the fastest change in living functions in Guzhang (39.57%). Ecological functions changed slowly or even declined (dynamic change degree-3.09%~12.08%). Over the five years studied, the dominance of land production or living functions in eight areas was 40% to 65%. The west of the region (except Huayuan) was dominated by diversified land use that focused on production functions, which was strongest in Longshan (64.46%). The east and west of Huayuan were marked by diversified land use dominated by living functions, and the dynamic dominance of living function was found to be the highest in Guzhang (55.90%). This indicates that the current notion of land use in Xiangxi is in the promotion of economic growth and the development of social undertakings to improve the livelihood of its residents.

**Figure 2.** Changes of land use total function from 2013 to 2017 in Xiangxi. (**a**) The value of the function. (**b**) The dynamic degree. (**c**) The dynamic dominance. (**d**) The standard deviation.

(3) The coordination of multi-functional land use was found to be changing slowly, with high coordination in the south and low in the central and northern regions. The standard deviation of land function in Jishou, Huayuan, Fenghuang, Baojing, and Yongshun underwent increasing fluctuations, which indicates a volatile decrease in functional coordination (Figure 2d). The standard deviation of land function in Luxi, Guzhang, and Longshan decreased, which reflects that the degree of functional coordination increased. However, the standard deviation for the land function in each area did not change significantly during the study period (variable rate −5.89% to 5.51%), which reflects the slow changes in the coordination of regional land use. The southern regions (Jishou, Huayuan, Fenghuang, and Yongshun) demonstrated a high degree of land use coordination, with the highest coordination seen in Jishou (the standard deviation was as low as 0.0036 in 2014), and the central (Baojing and Guzhang) and northern regions (Yongshun and Longshan) had a low degree of land use coordination, of which Yongshun was the lowest (with a standard deviation as high as 0.1567 in 2017). This was mainly because of the strengthening of the approach to development within Jishou (the core of Xiangxi) and the policy guidance and technical support of the local government, which promotes coordinated development in economy, society, and environment. However, the northern areas are far away from Jishou, and the effect of such policy changes reaching this area is limited. In addition, the industrial structure is unbalanced, with traditional industries, agriculture, and animal husbandry accounting for a large proportion. However, the degree of coordination in land resource utilization is low.

#### *3.2. The Spatio-Temporal Dynamic Analysis of the Sub-Function of Land Use*

The sub-function of the spatial and temporal evolution of land use during the study period was analyzed from three aspects: the value of functions, the dynamic degree of the functions, and the dynamic dominance of functions, using the data processed from the evaluation samples (Figure 3).

**Figure 3.** Changes in land use sub-function from 2013 to 2017 in Xiangxi. (**a**) Production functions. (**b**) Living functions. (**c**) Ecological functions.

#### 3.2.1. The Production Function

Production grew slowly, with unbalanced development over the whole region. The production functions of the region were generally not high and all increased slowly, with a faster growth rate in the south and the middle than in the north (Figure 3a). Jishou demonstrates the best production function and the fastest development, because it is the political, economic, and cultural center in that region, with relatively superior production conditions and rapid development in transportation and economy. The production function values of Baojing, Guzhang, Yongshun, Longshan, Huayuan, and Fenghuang were always at a low level and have changed slowly, which is due to the fact that these areas have been dominated by traditional agriculture and animal husbandry production, which is a decline in transportation and slow economic growth. The function of security from transportation was clearly differentiated over the region, but the spatial differentiation of agricultural and animal husbandry production and economic growth was not so clear. In 2012, the Aizhai bridge was opened to traffic and the Chongqing-Hunan highway was completed, which strengthened the economic connection and material exchange between Jishou, Chongqing, and Changsha. This promoted development in the flow of people and cargo, which rendered Jishou the regional transportation hub. Luxi is the only area through which this route passes, and the proportion of freight volume moving through the state has increased significantly (from 18.02% to 35.98% within five years). Therefore, Jishou and Luxi have both undergone a relatively rapid increase in traffic function. The expressways connecting other areas in the prefecture were developed later, such as that of LongYong and YongJi, which were completed in 2016 and 2017, respectively, and this has limited the external exchanges between these regions to a certain extent. Within the 5 years studied, the per capita agricultural output value and per capita output value from animal husbandry both increased. The agricultural and animal husbandry production of Guzhang underwent the largest increase rate (28.79%). The economic density of all areas increased, and the proportion of the secondary and tertiary industries has also risen in most counties (Huayuan and Luxi declined slightly). However, the overall economic growth of such industries within Xiangxi was slow during the study period. Jishou has always had a high proportion of secondary and tertiary industries (with 10.58% higher than the average level of the whole state in 2017), and the economic density has been growing rapidly (with a dynamic degree of 43.13%), so the dynamic degree of economic growth was the highest (24.54%).The production function of the whole prefecture focuses on agriculture and animal husbandry. Except for Jishou and Luxi with clear benefits from the development of transportation, the production functions of agriculture, and animal husbandry have developed rapidly in other areas.

#### 3.2.2. The Living Function

The living function fluctuated and tended toward a balance. The values for the living function of the land in the eight areas of Xiangxi have fluctuated, but have increased overall (Figure 3b). The living function of Yongshun has always been at the forefront (reaching the highest value of 0.7563 in 2017). Longshan and Jishou also have certain advantages in this area. The value for the living function in Guzhang was always low, but growth has been significant and the gap with other regions has narrowed. The level of employment support, cultural leisure, and residential homes have all improved to a certain degree. However, the urban-rural income balance index and the incidence of poverty in all areas of Xiangxi (except Jishou) have also increased to some extent. The wealth gap between urban and rural areas has widened, and the social security function has deteriorated. With the intensification of the urbanization process, the imbalance between urban and rural development has become increasingly prominent, and the incidence of rural poverty has increased. In recent years, local governments have encouraged the development of commercial housing, along with the renovation of dilapidated houses in rural areas and shanty towns in urban areas. The living conditions and living environment of residents have constantly improved. During the study period, the growth rate per capita in the area of Xiangxi was 13.77% to 112.61%, with the highest in Guzhang. With increasing communication with the outside world, job opportunities provided by tea production and tourism

in Xiangxi have increased, and the land employment guarantee function is, therefore, constantly optimized. The growth rate in the southern and central regions was faster than that of the north, which can be associated with the development of the regional production function. The improvement to production and the optimization of the industrial structure have enhanced the ability for increased employment opportunities. The functional level of cultural and leisure in regions other than Longshan and Fenghuang has been improved to a certain extent. The per capita green park space in Guzhang has increased 2.23 times, which is a key factor in the rapid growth of the cultural and leisure functions. The center of the living function was mainly residential homes (Longshan, Yongshun, Huayuan, Guzhang, and Luxi), which was followed by the cultural and leisure functions (Jishou, Baojing), and the employment support function (Fenghuang).

#### 3.2.3. The Ecological Function

The ecological function has changed slowly, and regional differences in this functional level have narrowed. The level of ecological function has slightly decreased in Guzhang, whereas it has increased in all other areas, albeit not significantly (Figure 3c). The value for the ecological function in land use over all eight areas of Xiangxi was always high, and was closely related to the significant supply of resources (due to the sparse population) and ecological maintenance (from high green coverage and forest coverage). The ecological advantage of land use in Yongshun was relatively clear because the three associated functions of resource supply, ecological maintenance, and environmental purity in Yongshun were all excellent. (In 2017, the three functions of Yongshun in all areas were respectively ranked 2, 2, and 1). Fenghuang was found to require improvement, since the level for ecological function in the area during 2015 and 2016 was the lowest. The standard deviation of the ecological function for each area decreased as a whole from 0.054 (2013) to 0.048 (2017), which reflects that the regional ecological function tended to be balanced. The three functions of ecological function (resource supply, ecological maintenance, and environmental purification) were clearly differentiated during the study period. With the promotion of ecological environmental construction by a local government, the rate of increase in green coverage and forest in all regions has been greatly improved, which means that the growth of ecological maintenance function is the clearest. The growth rate was 13.38% to 68.63% (except for Guzhang). In some counties (Fenghuang, Luxi, Yongshun), the resource supply capacity decreased, which reflects the lack of regional reserve resources and restricts the consequences of regional development. With the increased input of agricultural chemicals and the high intensity use of chemical fertilizers (taking Huayuan as an example, the index increased by 170.83% in the study period), the ecological purification capacity decreased, which posed a significant threat to the ecological environment of the regional land resources. In addition to the high dynamic dominance of resource supply in Guzhang, the ecological function of land use in the other seven areas focused on ecological maintenance.

#### *3.3. The Analysis of Obstacle Factors*

According to Equations (5) and (6), the degree of obstacles affecting both single functions and sub-functions was calculated, and the obstacle factors for the sub-functions and the total function were, respectively, obtained (Table 3). The obstacle factors of a sub-function were obtained in the following way: the obstacle degree of 10 single functions was ranked from high to low, with 1–3 as obstacle factors, 4–7 as intermediate factors, and the rest as dominant factors. The obstacle factors for total function were obtained according to the average value of the three sub-functions (production, living, and ecological functions). Those higher than the average value were counted as obstacle factors.


**Table 3.** The main obstacles of land use function in Xiangxi.

Note: PF is production function, LF is living function, and EF is ecological function. PF1 is the function for agriculture and animal husbandry production, PF2 is the function for economic growth, and PF3 is the function for transportation security. LF1 serves as employment support function, LF2 as the social security function, LF3 as cultural and leisure function, LF4 as the residential home function, EF1 as the resource supply function, EF2 as the ecological maintenance function, and EF3 as the environmental purification function.

According to the frequency of total dysfunction factors in the past five years, this study divided the land use function of the eight areas of Xiangxi into three types: living-ecological obstacle, production-living obstacle, and production obstacle.

#### 3.3.1. The Living-Ecological Type of Obstacle

Jishou is representative of the living-ecological obstacle type. Except for 2014, the degree of obstacles against production in Jishou was slightly higher than average (35%). In other years, the obstacle degree of production and living dysfunction (or one of the two) was notably higher, with significant room for improvement. From the perspective of this sub-function, Jishou suffers from a high degree of obstacles to the agricultural and animal husbandry production function (PF1), social security function (LF2), and environmental purification function (EF3). Statistics show that, in the recent five years, the per capita agricultural output value and per capita animal husbandry output value in Jishou were both in the middle and lower reaches. In 2013, the two indicators were respectively 2334.70 yuan/person and 609.58 yuan/person, which are both lower than the average value of the whole state. The urban-rural income balance index in Jishou was low, in the range of 31.29% to 34.86%, which was lower than the statewide rankings. Meanwhile, within five years, the amount of chemical fertilizers and pesticides applied per unit of cultivated land in Jishou was two to four times higher than that of other areas, which results in greater pressure on the ecological environment of the land, and the capacity for environmental purity was, therefore, weak.

#### 3.3.2. The Production-Living Type of Obstacle

Luxi and Guzhang were both undergoing obstacles to production and living. The main factors restricting the land versatility of the two areas were the economic growth function (PF2), the transportation security function (PF3), the employment support function (LF1), and the cultural leisure function (LF3). According to statistical data, in 2017, the economic density of Luxi and Guzhang accounted for 25.60% and 13.86% of Jishou, respectively, and the low output value per unit of land was one of the main factors restricting land use in the two areas. In 2017, passenger transport turnover in Luxi and Guzhang accounted for 7.58% and 4.98% of the total, respectively, and the backward transportation facilities also limited the development of the region. The average salary of workers

in both areas was not high (all were less than 88% of Jishou in 2017), which was less attractive for the labor force. In addition, farmers have to leave the area to work and do business frequently, and there are few rural employees (Guzhang retains less than one-third of the rural workers of Yongshun in 2017). The per capita green park space and the proportion of residents' cultural, educational, and entertainment expenditure in the two areas are less than the average level of the whole state, and there remains a significant room for improvement in land use for the cultural and leisure function (LF3).

#### 3.3.3. The Production Type of Obstacle

The geographical location of Xiangxi is in a remote region that restricts the efficient development and utilization of the land. Fenghuang, Huaguan, Baojing, Yongshun, and Longshan all suffer from production obstacles. The agriculture and animal husbandry production function (PF1), economic growth function (PF2), transportation support function (PF3), and environmental purification function (EF3) are all hindered by significant obstacles. The agricultural and animal husbandry production of Baojing, Yongshun, and Longshan have certain advantages, especially the agricultural output per capita and animal husbandry output per capita of Longshan and Yongshun in the last five years ranking in the top 3. However, the development of regional agricultural production is extensive, and is dominated by traditional agriculture and animal husbandry. The the level of productivity is low. The Wuling mountain area is high terrain, covering a large area with few people. It is limited by the influence of the surrounding economic influence, which limits the development of the economy and transport in the five areas. For example, in 2017, the sum of economic density of the five areas was 1.26 times that of Jishou. The sum of road area per capita was 1.04 times that of Jishou, and the sum of freight turnover was only 0.31 times that of Jishou. Although the land use intensity of Xiangxi is low, the ecological function has certain advantages. However, in the last five years, the large input of agricultural chemicals in these five areas (in 2017, the average amount of fertilizer applied in the five areas increased by 1.52 times that of 2013) led to an enhancement of human disturbances to the land and an increase in the environmental purification dysfunction degree of the land.

As a whole, the obstacle factors for land use in Xiangxi are part of the dynamic change in spatial and temporal dimensions. Although there are spatial and temporal differences in the types and degrees of obstacle factors, the main type of obstacles against efficient land use in Xiangxi are due to production, agriculture and animal husbandry (PF1), economic growth (PF2), transportation (PF3), and environmental purity (EF3), which are still the main factors restricting the overall land use function in Xiangxi.

#### **4. Discussion**

This study analyzed the multi-functional spatial and temporal evolution in the pattern of land use in eight areas of Xiangxi during the period of 2013 to 2017. This is based on an evaluation index system using production, living, and ecology as the main obstacle factors, which is divided into three obstacle types. To restrain the obstacles and promote sustainable land use in Xiangxi, it is necessary to learn from the experiences of other regions.

#### *4.1. Comparison in Land Use Management*

An assessment of 150 agricultural grasslands in Germany found that land intensive use increased feed production, but also led to loss of biodiversity and changes of land functions [35]. To improve the social and environmental functions of land as well as measurements of land management, irrigation and fertilization must be improved. Soil erosion, soil pollution, and soil degradation were serious problems in the corn belt of the United States, and agricultural production was seriously threatened. To coordinate the contradiction between grain production and environmental protection, and to achieve multi-functional agriculture, it was necessary to integrate market, technology, and policy measures [36]. From 1990 to 2010, the total land function of Guangzhou increased [37], but agricultural production and resource supply capacity became the main factors restricting its development. Consequently,

protecting farmland and improving resource utilization efficiency were the main measures to restrain functional impairment.

From the above cases, we can see that multi-functional land use is an inevitable trend. It is beneficial for the long-term use of land resources to formulate differentiated measures according to regional situations.

#### *4.2. Policy Suggestions for the Land Use in Xiangxi*

Based on the principle of a specific policy enforced in a certain city, we proposed the following suggestions.

(1) Jishou, which is a city with obstacles to both living and ecology, aims to narrow the urban-rural gap and reduce agricultural non-point source pollution. We will coordinate the construction of an urban and rural land use policy system, and establish a mobile market for urban and rural land elements. We will increase investment and policy for the support of infrastructure in rural areas, where the level of education, medical care, and social security will be enhanced. This will ensure that urban and rural residents both enjoy equal opportunities from development and receive equal protection for their rights. We will strictly protect basic farmland, delimit areas where basic farmland is concentrated, and transform medium-and low-yielding farmland. We will expand the planting of flue-cured tobacco, Chinese herbal medicines, fruits, and other cash crops, develop green and organic agriculture, and reduce pollution from non-point agricultural sources.

(2) To overcome the production-living barriers of Luxi and Guzhang, we should promote the development of production and public utilities. We will improve the distribution of residential areas and guide rural settlements to gather in towns and cities. We will make reasonable plans for the construction of transportation, ensure development of important transport routes such as the Zhang-Ji-Huai corridor, and promote the all-around development of highways, railways, and air transportation, etc. Natural resources and national cultural endowments will be instrumental in developing regional tourism. Featured agricultural products (bacon, alpine vegetables, etc.) will be intensively produced, to extend the industrial chain and expand related industries, so as to promote economic growth and improve the income level of employees. We will improve the land supply system, the planning and decision-making system, and optimize the structure of construction land. Land for cultural and recreational facilities in cities and towns shall be guaranteed. The renovation of old cities will be strengthened to create a beautiful urban environment. We will attract social funds, to increase the input for the livelihood of residents, and improve the functions of the central urban areas.

(3) The top priority for other counties with production barriers is given toward developing production and to quickly eliminating poverty. The mode of production in which smallholders operate should be changed by promoting the consolidation of farmland to achieve scale operation. We will develop modern urban agriculture and tourism agriculture, perfecting the agricultural industry system to improve the diversification of agricultural functions. The measures concerning transportation and economic development have been mentioned in the production-living obstacle countermeasures, and will not be repeated.

#### *4.3. Limitation and Future Research*

The research of land functions in Xiangxi is of great importance for optimizing land-use patterns, adjusting land use structure, and promoting coordinated development of regional economy and ecology. However, due to the difficulty of obtaining data, we did not select a countryside or a town as the case for analysis. No information of a specific case was obtained. Simultaneously, the research period is relatively short: only five years. Thus, there was a lack of a long-term historical evolution process analysis. Therefore, research of typical cases and long-term scales should be further studied.

#### **5. Conclusions**

The main conclusions are as shown below.


**Author Contributions:** Conceptualization, Q.-Y.Y. and H.X. Methodology, Q.-Y.Y. Investigation, H.X. Writing—original draft preparation, H.X. Writing—review and editing, K.-c.S. and Z.-X.Z.

**Funding:** The Key Project of Chongqing Key Research Base of Humanities and Social Sciences, grant number 14SKB014, funded this research.

**Acknowledgments:** The authors thank the X.X. Statistical Bureau for providing the data used in this research work. The authors thank the anonymous reviewer and W.Y.-Teacher of Southwest University, for their valuable comments.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
