Introduction

Landslides are considered one of the most extensively distributed mass developments in hilly terrain all over the world (Blothe et al. 2015; Carey et al. 2015; Nicolussi et al. 2015; Paulin et al. 2015; Pham et al. 2016a). Due to their unexpected and seasonal characteristics (Vilimek and Smolikova 2015; Yang et al. 2015), landslides always present great risks to human life and economic stability (Kritikos and Davies 2015; Panek 2015), especially in industrial and other social activities such as mines, water resources facilities and hydropower stations (Vranken et al. 2015; Zhang et al. 2015). Therefore, over the last two decades, many scientists have engaged themselves to predict landslide locations spatially, which is critical for development and planning purposes (Peng et al. 2015; Posner and Georgakakos 2015). In the formation of landslides, many factors play key roles; some are natural factors, and some are human factors. It has also been observed that climate change and landslide activities have a significant relationship (Sewell et al. 2015). Some recent studies have revealed that turbidity currents and submarine landslides will be the main cause of future rapid global climate warming. However, some research has indicated that accelerated climate change does not undoubtedly increase activity and thus does not provide a direct confirmation of climate change as a dominant triggering factor (Clare et al. 2015). In fact, heavy rainfall and snow melt have been supported as great indirect triggering factors due to complex hydrogeological effects and isolated groundwater organization. As such, it is difficult to determine the triggering parameter, especially when it is particularly affiliated with oblique river erosion in a landslide appendage (Abolmasov et al. 2015; Kirschbaum et al. 2015).

Preparation of a high-quality landslide inventory map is very critical in landslide studies. Field surveys and remote sensing-based analysis are commonly used, but both of these approaches have merits and demerits. Field investigation is highly time-consuming and very laborious. On the other hand, satellite image processing techniques also need rigorous validation steps to produce an authentic landslide inventory map (Hong et al. 2016; Lin et al. 2015). High-resolution imagery is an essential information base for properly evaluating and certifying landslide features (Akcay 2015; Yusof et al. 2015). Similarly, selecting landslide-conditioning factors is also an important and complex process. In this regard, the statistical index method is generally used for the validation of different landslide-conditioning factor sets. The method expresses how forceful every single landslide-conditioning factor enhances or diminishes the objective task and then excludes a few factors to acquire superior inputs (Bellugi et al. 2015; Meinhardt et al. 2015).

In recent years, extensive application of Geographical Information System (GIS) and Remote Sensing (RS) technologies has been used to assess landslide susceptibility, hazards and risk extent (Ciabatta et al. 2015; Oliveira et al. 2015; Pham et al. 2016c; Shahabi and Hashim 2015; Tan et al. 2015). There are numerous state-of-the-art approaches and methods that have been adopted in landslide susceptibility mapping. For example, the step-wise weight assessment ratio (Dehnavi et al. 2015), multivariate adaptive regression spline (MARS) (Chen et al. 2017d; Conoscenti et al. 2015; Wang et al. 2015b), artificial neural network (ANN) (Chen et al. 2017e; Dou et al. 2015; Bui et al. 2016), random forest (RF) (Chen et al. 2017g; Trigila et al. 2015), multicriteria evaluation (Ahmed 2015), kernel logistic regression (Chen et al. 2017f; Hong et al. 2015; Bui et al. 2016), spatial multicriteria evaluation (SMCE) (Gaprindashvili and Van Westen 2016), and adaptive neuro-fuzzy inference system (ANFIS) (Chen et al. 2017a; Nasiri Aghdam et al. 2016; Bui et al. 2012) have all been used. However, there still remains a question of the selection of the best method in landslide susceptibility assessment (Boue et al. 2015; Goetz et al. 2015; Pham et al. 2015). Thus, in order to achieve the most suitable and optimum model, many scientists have used various models in geographically different study areas worldwide (Chen et al. 2017g; Bui et al. 2017; Tsangaratos et al. 2017; Youssef et al. 2016).

The principal aim of the present study is to analyze probable utilization of the novel hybrid integration method of support vector machines (SVM) and random subspace (RS), named RSSVM, for landslide susceptibility assessment. Random subspace is an integrated algorithm, whereas RSSVM is a SVM-based classifier. Using statistical index-based assessment and receiver operating characteristic curve (ROC) means, the accuracy of the RSSVM model has been assessed. In addition, other approaches, such as Artificial Neural Network (ANN), Naïve Bayes (NB) and Support Vector Machine (SVM), have been applied, and the results were compared to the results of the RSSVM model. All the aforementioned models were applied in the Wuning area situated in the Jiangxi province (China). The analysis process was carried out using Weka 3.7 and ArcMap 10.0.

Model background

Support vector machines

A support vector machine (SVM) is defined as a supervised learning method and is widely applied in the statistical categorization and regression test (Chen et al. 2017b; Pham et al. 2016b). It leads to a map vector in a higher dimension space, within which there is a maximum separation hyperplane (Feng et al. 2016; Promper and Glade 2016). On the opposite side of the hyperplane, SVM separates data from two parallel hyperplanes from each other (Chen et al. 2017c). In addition, SVM separates a super flat surface to maximize the distance of two parallel hyperplanes (Bezak et al. 2016; Iadanza et al. 2016). Assuming that the distance between parallel gaps is greater, the classifier of the total error is smaller. The main idea of SVM can be summarized as follows (Li et al. 2016; Ma et al. 2016):

  1. 1.

    It is used to solve the problem of linear separable analysis, and in the case of a linear integral, nonlinear mapping algorithms are generally used, which makes a low spatial input space linear case into a high spatial feature space linear separable case (Alvioli and Baum 2016; Bennett et al. 2016; Mertens et al. 2016). Therefore, it makes possible a high spatial feature space by using a linear algorithm on nonlinear characteristics.

  2. 2.

    It is based on a structural opportunity minimization theory to create an excellent aggressive plane segmentation in the feature space (LaHusen et al. 2016; Osadchiev et al. 2016; Yamao et al. 2016).

  3. 3.

    More detailed information about the SVM algorithm can be seen in Chen et al. (2016a, 2016b, 2017h) and Pradhan (2013). The equation of SVM algorithm can be described as follows:

    $$ y_{i} \left( {w \cdot x_{i} + b} \right)/ \ge 1 - \xi_{i} $$
    (1)

Random subspace ensemble

Random subspace is defined as a classic integrated algorithm which was proposed by Ho in 1998 (Tin Kam 1998). The algorithm is similar to the bagging algorithm and is randomly selected by the original training set to construct the training subset (Kotsiantis 2011; Kuncheva et al. 2010; Mielniczuk and Teisseyre 2014). However, the difference is that random subspace is randomly chosen from the original training set of features (Bertoni et al. 2005; Skurichina and Duin 2001, 2002). Then, the series features a subset of each subclassifier training at the final forecast results obtained by a combination of voting methods (Lai et al. 2006; Sun and Zhang 2007; Tao et al. 2006). The performance of the results depends on integrated learning differences, the bagging method subcategories to obtain the difference between the subclassification among different samples of each subclassifier training, and the random subspace ensemble learning method to take samples at different spatial characteristics to obtain differences between subclassification (Nanni and Lumini 2008; Zhang and Jia 2007; Zhu et al. 2009).

Novel hybrid integration of support vector machines (SVM) and random subspace (RS) (RSSVM)

The novel hybrid integration of the support vector machines (SVM) approach and the random subspace (RS), ensemble (RSSVM), for spatial forecasting of landslides appearing in the Wuning area, is displayed in Fig. 1.

Fig. 1
figure 1

The flowchart of the proposed RSSVM model

Full size image
  1. 1.

    Data collection and processing in GIS The landslide database is created for building and certifying the proposed integrated landslide susceptibility model. The landslide inventory map with 445 landslides was randomly split into 70% (311 landslides shown in the yellow color) for coaching the models and the remaining 30% (134 landslides shown in the red color) for validation purposes (Fig. 2). The coaching and validating datasets were applied to build the landslide susceptibility model, whereas the testing dataset was employed to check the RSSVM model. For this purpose, by employing the frequency ratio method, fourteen landslide-influencing factors were reclassified from categorical classes into continuous values. (Bui et al. 2015). Subsequently, all 14 landslide-conditioning factor maps were transformed to a raster format with a pixel of 25 m.

    Fig. 2
    figure 2

    Landslide location map of the Wuning area [Map of China from the National Geographic World Map (ESRI 2010)]

    Full size image
  2. 2.

    Random Subspace ensemble (RSS) optimization Random subspace (RSS) can improve the performance of the SVM classifier by dividing a dataset with a large dimensional feature space to lower datasets, such as in training data. Generally, these ensemble data mining classifiers have better accuracy than a single predictor. These methods first produce via training the dataset base classifications. Second, the real classification is achieved by combining the results of base classifiers with the previous base classifier (Piao et al. 2015). The process (1) divides the original feature space (FR) into L feature subsets (FS) of p-dimensionality, (2) submits each subset to a base classifier (BC) in the ensemble, and (3) makes a final decision on the class of the form achieved by connecting the decisions of these BCs using a connecting order such as “most votes” (Kuncheva and Plumpton 2010). Because the feature subsets submitted to the BCs are diverse, an aligned ensemble is a common selection (Kuncheva and Plumpton 2010). By evaluating the fitness of each result, the algorithm constantly searches the solution space and pursues the most appropriate set of the model criteria.

  3. 3.

    Training support vector machines model (SVM) As we know, the SVM algorithm is a supervised learning method; a training dataset with input components and matching desired class labels must be supplied. Using the RBF kernel function and based on the training dataset, the SVM algorithm draws the input data from the original input space into a high-dimensional feature space (Chen et al. 2016a). Accordingly, the result of this algorithm builds up a hyperplane. SVM can separate the input data of influencing factors into two distinctive decision areas: “landslide” (Y = 1) and “non-landslide” (Y = − 1). When the training case is finished, the SVM algorithm for classifying input patterns is displayed.

  4. 4.

    The optimized RSSVM model The SVM-based boost method continues until the maximum number of generations is obtained (Piao et al. 2015). When the exploring process completes, an appropriate set of the RSSVM adapt parameters is established. Then, it is used to compose a classification model for spatial evaluation of landslide susceptibility in the Wuning area. The suggested model is prepared to forecast invisible input arrangements in the validation dataset.

  5. 5.

    Comparison with data mining techniques The obtained results from the novel hybrid integration of support vector machines and random subspace ensemble (RSSVM) were compared to other well-known data mining techniques such as Multiple Perceptron Neural Network (MLPN), Naïve Bayes (NB) and Support Vector Machine (SVM) for spatial assessment of landslide susceptibility in the Wuning area.

Experiment and analysis

Description of study region

The Wuning area is almost 3507 km2 and overlays the entire area of the Wuning region. The Wuning area is situated in the western part of Jiangxi province. It lies in the South hilly area of Mubu between the longitudes of 114°29′E and 115°27′E and the latitudes of 28°53′N and 29°35′N (Fig. 2). According to a report of geological survey results (http://www.cgs.gov.cn/), there are more than 56 geologic groups and units recognized (Table 1). In the Wuning area, the main lithologies are Tuffaceous slate, phyllite, spilite, quartz keratophyre, (two long; K-feldspar) granite, and two-mica adamellite (Fig. 3). Geologic data for the Wuning area were acquired from the China Geology Survey (http://www.cgs.gov.cn/). Over the last 20 years, many mines have been developed in the Wuning area.

Table 1 Types of geological formations in the study area
Full size table
Fig. 3
figure 3

Geology map of the study area

Full size image

The Wuning area is in the subtropics monsoon climate region. The wet period is usually from April to August. The total average precipitation for 3 months (April–June) is approximately 638 mm. The largest daily precipitation is more than 100 mm during the wet period. The dry period is commonly from September to January with average precipitation approximately 65.6 mm/month. The annual average precipitation in the Wuning area varies from 578 to 1183 mm, with an average of 162.4 wet days. The annual average temperature is 16.6 °C. The highest and lowest average temperatures are 39.1 and − 5.2 °C, respectively. The annual average sunshine in the Wuning area is 1700.5 h.

Data preparation

Historical landslide events

There are numerous different methods and techniques being used to construct landslide inventory maps. These methods are field survey, satellite image interpretation, aerial photography and historical landslide records (Benoit et al. 2015; Palamakumbure et al. 2015; Varilova et al. 2015; Wang et al. 2015a). However, until now, there has been no agreement among scientists on the best suitable primary method for producing an accurate landslide inventory map ((Hong et al. 2017; Uhlemann et al. 2016).

In this research, the landslide inventory map was generated by combining field survey data with satellite image data. The landslide inventory map for the Wuning area had 445 landslide locations, which were provided by the Jiangxi Province Meteorological Bureau (http://www.weather.org.cn) and the Department of Land and Resources of Jiangxi Province (http://www.cgs.gov.cn; Fig. 2). Figure 4 shows Google Earth photographs of landslides in the study area. The landslide inventory map demonstrates that the volume of the smallest landslide is 20 m3, the volume of the largest is 96,000 m3, and the average volume of landslides is 1761.3 m3. In the study area, large-sized landslides (> 1000 m3) account for 8.1% of the total landslides, and 254 people have been threatened. Medium-sized (200–1000 m3) landslides account for approximately 16.0% of the total landslides and have threatened 121 people in the study area. Small-sized landslides (< 200 m3) have threatened 551 people and account for 75.9% of the total landslides. Heavy rainfall is major reason for these landslide occurrences; there have been no reports about earthquake-induced landslides. Approximately 38.5% of the landslides occurred when the measured rainfall was approximately 100 mm per day. The other landslides occurred when the daily rainfall was greater than 105 mm.

Fig. 4
figure 4

Google Earth photographs showing landslides of the study area

Full size image

Landslide-conditioning parameters

In the current study, based on the analysis of the landslide inventory map and some literature, a total of 14 landslide-conditioning factors were selected as follows: slope, aspect, altitude, topographic wetness index (TWI), stream power index (SPI), sediment transport index (STI), soil, lithology, normalized difference vegetation index (NDVI), land use, rainfall, distance to roads, distance to rivers and distance to faults.

A digital elevation model (DEM) for the Wuning area was produced from ASTER GDEM Version 2 (http://gdem.ersdac.jspacesystems.or.jp). Based on this DEM, slope, altitude, aspect, TWI, SPI and STI were extracted Using Arcgis 10.2 software (Fig. 5a–f). The soil map was compiled in 1995 by the Institute of Soil Science, Chinese Academy of Sciences (ISSCAS) China (http://www.issas.ac.cn; Fig. 5g). The lithology map was categorized into eight groups (A, B, C, D, E, F, G, H, I and J) (Fig. 5h). The NDVI and land use map were obtained from Landsat 7 ETM+ satellite images that were acquired on 10 December 1999 (Fig. 5i, j). These images were obtained from the Computer Network Information Centre of Chinese Academy of Sciences (http://www.gscloud.cn). The NDVI was calculated using the common formula:

$$ {\text{NDVI}} = \left( {{\text{NIR}} - R} \right)/\left( {{\text{NIR}} + R} \right) $$
(2)

where NIR and Red are the near infrared and red bands, which are from 0.7 to 1 and 0.6 to 0.7 lm, respectively, of the electromagnetic spectrum. The maximum likelihood supervised method was used for the land use classification, with a classification accuracy of 90.7%. Rainfall data are from the Jiangxi Province Meteorological Bureau (http://www.weather.org.cn). For the period 1960–2012, there were 37 rainfall stations that were used to create the rainfall map. The mean annual precipitation was divided into five classes (Fig. 5k) using the inverse distance weighted method (Zhu et al. 2012). Distance to roads and rivers and distance to faults were produced from topographic maps and geological maps, respectively (Fig. 5l–n). The detailed information of the classes for each landslide-influencing factors is provided in Table 2. Finally, all these maps were transformed into the same resolution of 25 m × 25 m.

Fig. 5
figure 5figure 5

Landslide-conditioning factor maps of the Wuning area: a slope degree, b aspect, c altitude, d topographic wetness index (TWI), e stream power index (SPI), f sediment transport index (STI), g soil, h lithology, i normalized difference vegetation index (NDVI), j land use, k rainfall, l distance to roads, m distance to rivers, n distance to faults

Full size image
Table 2 Classes of landslide-influencing factors
Full size table

Results and discussion

Feature selection of linear support vector machine (LSVM)

The Linear Support Vector Machine (LSVM) is a classifier of the Support Vector Machine (SVM) algorithm, which has been widely used in landslide susceptibility modeling (Chou et al. 2016; Conte et al. 2016; Gullà et al. 2016). According to a one-vs-the-rest arrangement, this class holds both sparse and dense input, and the multiclass hold is controlled (Fan et al. 2016; Lora et al. 2016; Romano et al. 2016).

It is very meaningful to assess the predictive ability of an assembling training dataset using fourteen landslide-conditioning causes. In this study, we use the Linear Support Vector Machine method with tenfold cross-validation. Figure 6 presents the predictive ability of landslide-conditioning factors in the Wuning area. It was demonstrated that slope angle has the best predictive ability in landslide susceptibility models (AM = 13.8). Rainfall also has a very high offering in landslide susceptibility models (AM = 12.8). TWI (AM = 12) and STI (AM = 11.4) have relatively high offerings in landslide susceptibility models as well. Slope aspect (AM = 10) and distance to road (AM = 9) have moderate offerings in the modeling. Other factors, such as distance to faults (AM = 7.3) and distance to rivers (AM = 7.2), had nearly similar predictive abilities. Altitude (AM = 5.4), land use (AM = 4.9) and lithology (AM = 4.9) had similar offerings. In contrast, NDVI (AM = 2.3) and SPI (AM = 2.1) had low predictive ability, and soil type (AM = 1.9) had the lowest predictive ability.

Fig. 6
figure 6

Predictive ability of landslide trigging factors using the LSVM approach

Full size image

In sum, all fourteen landslide-conditioning factors contributed to the landslide susceptibility models (AM > 0). Overall, these fourteen landslide-conditioning factors have been used in landslide susceptibility.

Preparation of dataset and training the RSSVM model

Performance of the RSSVM model significantly depended on the selection of the calculating parameter, which is the number of iterations. Thus, a test of the performance of the RSSVM model was accomplished with different numbers of iterations to filter the optimal parameter. For this purpose, the ROC curve method was used to evaluate the performance of RSSVM.

Figure 7 shows the analytical results of the ROC curve with various numbers of iterations for training and testing the RSSVM model. It can be seen that when the number of iterations is 14, this results in the best performance of the RSSVM model. Thus, in the present study, the number of iterations has been set to 14 for training the RSSVM classifier in the novel classifier framework.

Fig. 7
figure 7

Analysis of the results of the RSSVM model using ROC curve with various numbers of iterations

Full size image

Validation of predictive ability of the RSSVM model

Performance of the RSSVM model for landslide susceptibility assessment using statistical index-based evaluations is shown in Table 3. It can be seen that the RSSVM model achieved good classification of both landslide and non-landslide pixels. The positive predictive value is 77.39% in the training dataset and 78.18% in the testing dataset, indicating that the probability the RSSVM model accurately classifies pixels to the landslide class is 77.39 and 78.18%. The sensitivity is 88.62% for the training dataset and 78.18% for the testing dataset indicating that 88.62 and 78.18% of the landslide pixels are accurately classified into the landslide class. Overall, the performance of the RSSVM model for classification of landslide pixels (specificity = 79.93%) is slightly better than those of non-landslide pixels (specificity = 78.18%).

Table 3 Performance of the RSSVM model using statistical index-based evaluations
Full size table

The receiver operating characteristic (ROC) curve was also applied to evaluate the general performance of the RSSVM model. The ROC curve is widely used in landslide susceptibility mapping. In general, the AUC value varies from 0.5 to 1; if the AUC value is equal to 1, the result of the landslide model is excellent; otherwise, if the AUC value is equal to 0.5, the result of the landslide model is imprecise. Figures 8 and 9 show the results of the RSSVM model for landslide susceptibility assessment using the ROC curve technique. In this study, the RSSVM model executed very well based on the analysis of the ROC curve (AUC = 0.918); additionally, the ROC curve for testing the RSSVM model is 0.857, which is reasonably satisfactory.

Fig. 8
figure 8

Analysis of the ROC curve for training the RSSVM model

Full size image
Fig. 9
figure 9

Analysis of the ROC curve for testing the RSSVM model

Full size image

Comparison of the RSSVM model with popular landslide models

In this study, other popular landslide susceptibility models, such as Multiple Perceptron Neural Network, Naïve Bayes and Support Vector Machine, have been applied and compared to the result of the proposed hybrid model.

Multiple Perceptron Neural Network (MLP) Based on the technique of biological nervous systems, which contain the brain and process information, artificial neural networks are defined as an information processing method (Haeberli et al. 2001; Satorra and Bentler 2001). The real function and effectiveness of neural network algorithms demonstrate their capability to perform both linear and nonlinear connections and to master these connections directly from the modeling data (Carlini et al. 2016; Gutiérrez and Lizaga 2016). Classic linear models are naturally poor when they input modeling data that includes nonlinear information (Andrews 1988; Ye and Chen 2001). As we know, neural network algorithms are being adapted to an expanding number of real-world problems (Dickson and Perry 2016; Wang et al. 2016). Their basic convenience is that they can address issues that are too complicated for normal methods (Leung et al. 2000; Rao and Scott 1987). In general, neural network algorithms are well suited to address problems that include pattern recognition of trends in data (Ogneva-Himmelberger et al. 2009; Song et al. 2014). The most ordinary neural network algorithm is the multiple perceptron. Due to its need for a desired output to study, this type of neural network algorithm is famous as a supervised system (Moore and Sawyer 2016). The goal of it is to build a model that correctly maps the input and output data so that it can then be applied to obtain the output result even if the output is unknown (Webster et al. 2016). Back propagation is the most common algorithm adopted by the multiple perceptron neural network algorithm (Shi et al. 2016). With back propagation, the input data are often given to the neural network algorithm (Ciurleo et al. 2016). The neural network algorithm always adjusts the weights and decreases with each iteration error, moving closer and closer to acquiring the coveted result (Gutiérrez and Lizaga 2016).

Naïve Bayes (NB) The definition of a Bayesian network is a directed acyclic diagram with a probability annotation where every node in the graph represents random variables; two nodes in the graph occur if there is an arc, and the two nodes correspond to the probability of whether a random variable is dependent and conversely indicates that two independent random variables are the conditions (Pirdavani et al. 2014a, b). An arbitrary node in the network of X has a corresponding Conditional Probability Table (Conditional aim-listed Probability Table, CPT), and nodes X in the father obtain all the possible values of the Conditional Probability (Chen et al. 2012; Harris et al. 2010). If nodes are without father X, then X CPT for the prior probability is the distribution (Wei and Qi 2012). A Bayesian network structure of the nodes and the CPT define the likelihood of allocation of each variable in the system. Naïve Bayes models originated through classical mathematics theory and have a solid mathematical foundation as well as the stability of classification efficiency (Zhang et al. 2011). At the same time, the NB model requires only a few estimated parameters, is less sensitive to missing data, and the algorithm is simpler (Zhang and Mei 2011). In theory, the NB model of minimum error rate compares well with other classification methods (Hadayeghi et al. 2010). However, this is not always the case because the NB model assumptions are independent of each other between attributes. This assumption is often not established in practical applications, and this brings a certain influence to the correct classification of the NB model (Koutsias et al. 2010). Depending on the number of attributes or when the attribute correlation is large, the NBC model classification efficiency is better than the decision tree model. If the attribute correlation is small, the performance of the NB model is good (Sharma et al. 2011). The Naïve Bayes theorem is a kind of unsupervised learning with no iteration; the learning efficiency is high, and it is easy to implement under large sample sizes with better performance (Kumar et al. 2012; Lukawska-Matuszewska and Urbanski 2014). However, because the conditional independence assumption is too strong, assumption features associated with the condition of the characteristics of the input vector of scenarios do not apply (Martinez-Fernandez et al. 2013; Paez et al. 2011).

Using the results from analyzing the performance of the ROC curve for different landslide models (Fig. 10), it can be found that the RSSVM model (AUC = 0.857) has the highest performance, followed by the SVM model (AUC = 0.814), MLPN model (AUC = 0.823) and NB model (AUC = 0.783).

Fig. 10
figure 10

Comparison of predictive ability of landslide models

Full size image

Delineating landslide susceptibility maps

In this study, 311 landslides were used for training data (70%), and 134 landslides were used for validation data (30%). First, the training dataset of the SVM and NB models were run using Weka 3.7.12 software. The polygon of the study area in ArcGIS 10.2 was then converted to rasters with a pixel size of 25 × 25 m, which was similar for all conditioning factors. The raster polygon was sampled in GIS with all conditioning factors, and this layer is the test dataset used in the Weka program for landslide susceptibility mapping. Consequently, according to the samples for each point raster of the study area, a probability of landslide incidence was obtained and transferred to GIS, and a landslide susceptibility map for the RSSVM model was produced. Ultimately, the range of values of the susceptibility map was classified into five categories based on the natural break classification method (Basofi et al. 2015), including very low susceptibility (VLS), low susceptibility (LS), moderate susceptibility (MS), high susceptibility (HS) and very high susceptibility (VLS) (Dai and Lee 2002; Table 4). The landslide susceptibility map of the RSSVM model is shown in Fig. 11.

Table 4 Landslide bulk on landslide susceptibility map
Full size table
Fig. 11
figure 11

Landslide susceptibility map generated using the RSSVM model

Full size image

Interpretation of the landslide susceptibility map generated using the RSSVM model shows that the very high class covers only 12.98% of the total study area but contains 56.12% of the total landslide locations. In contrast, the low and very low classes account for 49.14% of the total study area; however, they contain only 3.57% of the total landslide locations. This indicates that the RSSVM model produces a high accuracy result and that the map fits well with the landslide inventories.

Conclusions

Landslides are a most dangerous and hugely destructive disaster all over the world. For this reason, landslide susceptibility research is most important for local management and town planners. Many scientists have utilized different methods to develop landslide susceptibility maps in various regions worldwide. However, until now there has been no agreement about the best method to use in landslide susceptibility modeling. Thus, the current study aimed to discover a new ensemble method to achieve this target.

In this study, a novel hybrid integration approach was used by integrating support vector machines (SVM) and random subspace (RS) for landslide susceptibility assessment. The result shows that the RSSVM model performs very well in landslide susceptibility mapping in the Wuning area (China). The predictive ability of a base classifier of SVM is significantly improved through the newly proposed RSSVM model. In comparison with Multiple Perceptron Neural Network (MLP), Naïve Bayes (NB) and Support Vector Machine (SVM), the RSSVM model has the best performance. In general, the landslide susceptibility map of this model is very beneficial for decision makers and land use planners in the Wuning area.