Data Mining-Aided Automatic Landslide Detection Using Airborne Laser Scanning Data in Densely Forested Tropical Areas

Abstract

Landslide is a natural hazard that threats lives and properties in many areas around the world. Landslides are difficult to recognize, particularly in rainforest regions. Thus, an accurate, detailed, and updated inventory map is required for landslide susceptibility, hazard, and risk analyses. The inconsistency in the results obtained using different features selection techniques in the literature has highlighted the importance of evaluating these techniques. Thus, in this study, six techniques of features selection were evaluated. Very-high-resolution LiDAR point clouds and orthophotos were acquired simultaneously in a rainforest area of Cameron Highlands, Malaysia by airborne laser scanning (LiDAR). A fuzzy-based segmentation parameter (FbSP optimizer) was used to optimize the segmentation parameters. Training samples were evaluated using a stratified random sampling method and set to 70% training samples. Two machine-learning algorithms, namely, Support Vector Machine (SVM) and Random Forest (RF), were used to evaluate the performance of each features selection algorithm. The overall accuracies of the SVM and RF models revealed that three of the six algorithms exhibited higher ranks in landslide detection. Results indicated that the classification accuracies of the RF classifier were higher than the SVM classifier using either all features or only the optimal features. The proposed techniques performed well in detecting the landslides in a rainforest area of Malaysia, and these techniques can be easily extended to similar regions.

1. Introduction

Landslide is a natural disaster that adversely affects lives and properties (Jebur et al., 2014). Therefore, continuously developing detailed and updated landslide inventory maps is important (Lin et al., 2013). These maps are important data sources for landslide susceptibility mapping and risk assessment(Ardizzone et al., 2007; Fiorucci et al., 2011; Guzzetti et al., 2012; Van Westen et al., 2008). Landslides must be accurately detected to produce high-quality landslide inventory maps (Siyahghalati et al., 2016). However, mapping a landslide inventory in tropical areas is challenging because the dense vegetation cover in these regions obscures underlying landforms(Pradhan et al., 2016; Chen et al., 2015).

Landslide inventory maps are traditionally produced by visually interpreting aerial photographs. This technique inevitably requires several field surveys; thus, it is time-consuming and costly (Van Den Eeckhaut et al., 2005). Field mapping may also fail to provide a complete view of large-scale landslides in certain areas, particularly in densely vegetated areas. Old landslides are also sometimes difficult to visualize using aerial photographs or satellite images because of the vegetation cover or the alteration caused by other slope failures and human activities(Miller et al., 2012). Consequently, this method enables researchers to map and recognize single and small groups of landslides (Galli et al., 2008).

Usually, visua interpretation of aerial photographs is used in the traditional techniques in the construction of landslide inventory maps which requires multiple field surveys. This approach could be quite expensive and time consuming. According to Brardinoni et al.(2003) traditional techniques include visual interpretation of stereoscopic aerial photographs and geomorphological field mapping. However, geomorphological field mapping experience certain level of limitations such as degree of landslide which is usually too big to be completely studied in an area. Therefore, research activities are limited in terms of perspective in distinguishing all the features of landslide in detail. (Miller et al., 2012) reported that most at times old landslides are usually covered by vegetation or have experienced slope feature changes. Therefore, this method gives investigators avenue to recognize and map single landslide or minor groups of landslides(Galli et al., 2008). Lee et al.(2012)reported that some of the oldest techniques used for landslide detection are still in use today despite substantial progress made in the technology of aerial photographs. The deployment of vertical exaggeration using stereoscope enables the morphological structure of land to be amplified to ease detection of changes in slopes. According to Nichol et al.(2006) this method does not require any sophisticated technical capabilities. Also, Malamud et al.(2004)reported that an aerial photograph having fine scale and a arge size can cover the entire location of landslide in a particular period. This sets of aerial photographs for similar regions could serve as a valuable resource for research works to conduct a temporal evaluation of landslides (Miller et al., 2012). The use of aerial photography in detecting slope failures is an uncertain technique that requires training, an organized methodology, skills and proper interpretation principles (Antonini et al., 2002). However, there seems to be no standard procedure yet available to identify and categorize landslide based on the investigation and knowledge of set of features that can be to recognized images(Pradhan et al., 2016). Besides, vegetation thickness and height and changes influence the way slope failure are recognized when using aerial photographs (De Blasio, 2011).

Optical remote sensing and synthetic aperture radar (SAR)-based remote sensing have led to significant progress in landslide inventory mapping. Remote sensing data that are useful for landslide studies include SAR images, high-spatial-resolution multispectral images, and digital elevation models (DEMs), which are obtained from space-borne sensors and airborne laser scanning systems (Ardizzone et al., 2007; Guzzetti et al., 2005; Jebur et al., 2014; Stumpf and Kerle, 2011). However, it is difficult to identify landslides in rough topographies and dense vegetation cover using aerial photographs, SAR and VHR and imageries, due to the fact that the morphologic featues revealing landslides could be subdued (McKean and Roering, 2004; Wills and McCrink, 2002). Second, data interpretation is frequently based on the expert knowledge and experience of an analyst as well as his or her familiarity with the area (Chen et al., 2014; Malamud et al., 2004). Third, additional errors can be introduced while translating image interpretation results into thematic maps (Malamud et al., 2004). High resolution LiDAR-derived DEMs can depict ground surfaces and provide valuable information about the topographic features of possible landslide-affected areas that are covered by dense vegetation (McKean and Roering, 2004). LiDAR technology is considered an effective tool for detecting landslides and mapping the features of densely vegetated areas because of its capability of obtaining high-resolution topographic data by penetrating through the canopy and thick vegetation (Van Westen et al., 2008; Borkowski et al., 2011; Pradhan et al., 2016; Wang et al., 2013). LiDAR data and their derivatives, such as hillshade, surface roughness, slope, and contour maps, provide significant and valuable information about active geological processes, such as landslides, which reshape the topography of an area (Booth et al., 2009; McKean and Roering, 2004; Van Den Eeckhaut et al., 2011).

In the area of remote sensing and geoscience, image analysis methods are commonly used in studying landslide. According to Gao and Mas(2008), pixel based and object-based image aalysis (OBIA) methods have beenCompared in numerous studies. OBIA can be applied at different scales, unlike the pixel-based analysis (PBA). Various objects sizes that depict different land features can be produced depending on the selected application such as environment under analysis and underlying input imagery. OBIA can develop contextual semantic features and additional geometry that can be used for classification studies (Duro et al., 2012). In heavily forested areas, OBIA using LiDAR data have evolved to be an alternative approach due the difficulty in the use of optical image-based analysis vegetated in rugged and terrain (Li et al., 2016). Conversely, the pixel-based approach (Fei and Lee, 2009; Rau et al., 2012) usually create pepper-and-salt effect that makes onsite identification difficult with poor transferability (Drăguţ and Blaschke, 2006).

A sufficient number of training areas should be used to represent the variability of a class (Pal and Mather, 2003). Furthermore, time and cost efficiency should be considered when designing a sampling scheme to achieve high accuracy (Lippitt et al., 2008). An adequate training set size is needed to obtain high classification accuracy (Foody and Mathur, 2006; Pal and Mather, 2003). Most of the cited studies used a random sampling method to obtain the reference data for training and assessment(Puissant et al., 2014; Zhen et al., 2013). However, this sampling method is not always suitable because it leans toward undesampling, but increasing the training set size is imperative for mapping categories.

Feature subset selection is crucial in data mining (Karegowda et al., 2010). The high-dimensional dataset, however, makes it difficult for testing and training the classification methods. Few object-based studies have handled the features selection for landslide detection by using LiDAR data (Chen et al., 2014; Li et al., 2015). Karegowda et al.(2010) demonstrated the significance of feature selection by using Correlation-based Feature Selection (CFS) and gain ratio algorithms. Aladesote et al.(2016) applied two feature selection techniques, namely, Gain Ratio Feature (GR) selection and principal component analysis (PCA) to an intrusion detection system. Their results indicated that both algorithms efficiently selected highly relevant features from a dataset. Chen et al.(2014) successfully applied RF for feature selection. Venkateswaran et al.(2016) compared Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) and found that the former was effective on LISSIV Madurai imagery and more accurate than the GA. Dou et al. (2015) proposed an automatic landslide detection method using an integrated approach consisting of object-oriented image analysis, a GA, and a case-based reasoning technique. They found that the GA was the best algorithm. Imani et al. (2012) integrated GA and Ant Colony Optimization (ACO), and the integrated model presented better performance than the standar GA and ACO did. However, Dadaneh et al.(2016) applied ACO to feature selection and asserted the effectiveness of ACO; However, they did not use a LiDAR intensity data. For object based image analysis (OBIA), optimizing multiresolution segmentation parameters for landslides plays a key role in exploiting the spectral and spatial information. In OBIA, feature selection is considered as a significant step because it improves the performance of the classifier and reduces the complexity of the computation by removing redundant information (Pedergnana et al., 2013). In this regard, few studies have examined feature selection algorithm for landslide detection through an object-based approach and only LiDAR data. Therefore, this study aims to optimize multiresolution segmentation parameters by using FbSP optimizer and evaluated six feature selection algorithms in order to find the best combination subset for detection landslide.

This study aimed to investigate suitable algorithms for selecting features in landslide detection using airborne laser scanning data. The specific objectives were as follows: 1) to optimize the multiresolution segmentation parameters, 2) to evaluate the six feature selection algorithms for landslide detection, and 3) to determine the appropriate algorithms for selecting features by using RF and SVM classifiers. The studied algorithms have not been tested in existing former studies, particularly in landslide detection. Selecting the suitable algorith for landslide detection could improve the accuracy of results.

2. Study Area

The Cameron Highlands is a rainforest area characterized by a dense vegetation cover and frequent occurrences of landslides. The study region covers an area of 26.7 km². Geographically, the Cameron Highlands is located in the north of peninsular Malaysia at a latitude range of 4°24′32″N–4°24′43″N and a longitude range of 101°22′54″E–101°23′11″E. The annual average rainfall in the area is approximately 2,660 mm. Its average temperature is approximately 24℃ and 14℃ during daytime and nighttime, respectively. A large part of the area (approximately 80%) is forested, and the slope inclination in the area range from flat terrain (0°) to hilly area (80°). Three sites were selected for the analysis of the proposed method, as shown in Fig. 1. These sites were selected according to information obtained from an inventory map. The site (a) was used for developing the method of landslide detection, whereas sites (b) and (c) were used for testing the method in terms of model transferability. Considerations were taken into account in selecting the sites to avoid the missing in land-cover classes.

Fig. 1. Shows study is that consist of [red] Analysis area; [black] Test site-1; and [yellow] Test site-2.

3. Methodology

According to Lu and Wong (2008), digital suface model(DSM), digital elevation model(0.5 m)(DEM), and intensity featue are produced by converting LiDAR point cloud into raster data through inverse distance weighting (IDW) interpolation. DEM is then used to generate other LiDAR-derived products (i.e., slope, aspect, hillshade). Subsequently, normalized digital surface model (nDSM) also known as Height were produced by subtracting the DSM from DEM. Then, the LiDAR-derived products and orthophotos were combined by correcting their geometric distortions, integrating them into a coordinate system, and then prepared in a GIS for feature extraction. Next, the FbSP optimizer developed by Zhang et al. (2010) was used to acquire the suitable parameters (scale, shape, and compactness) at different segmentation levels. Subsequently, the training samples were evaluated using a stratified random scheme following the procedure adopted by Ma et al. (2016). Relevant features were selected using the six algorithms and ranked from the most important. Two supervised learning models, namely, SVM and RF, were used to identify the locations of the landslides. The transferability of each classifier model was then verified by applying it to other sites (i.e., Test site-1 and Test site-2). Finally, the results were validated and compared using a confusion matrix and overlaid with the slope and aspect derived from the LiDAR DEM data to identify other characteristics of the landslide, such as direction, runoff, width, and length. The flowchrt of the proposed method is illustrated in Fig. 2.

Fig. 2. Fow chart of processing steps that were followed in the current study.

1) Data Used

The LiDAR point cloud data were collected in an area of 26.7 km² over the Ringlet and surrounding area of the Cameron Highlands at a flying height of 1510 m. The LiDAR data and were captured on January 15, 2015. The point density was 8 points per square meter and the pulse rate frequency was 25,000 Hz. The absolute accuracy of the LiDAR data should meet the root-mean-square errors of 0.15 m in the vertical axis and 0.3 m in the horizontal axis as standardized by Department of Survey and Mapping Malaysia (JUPEM). The same system for the collection of LiDAR point cloud data in the study area was used to collect the orthophotos. A DEM with 0.5 m spatial resolution was interpolated from the LiDAR point clouds after the non-ground points were removed using inverse distance weighting, with GDM2000/ Peninsula RSO as the spatial reference.

Subsequently, the LiDAR-based DEM was used in generating a number of derived layers to facilitate the detection of landslides and their characteristics(Miner et al., 2010). The slope is considered an important factor of land stability because of its direct impact on landslide phenomenology (Martha et al., 2011). Moreover, the slope is the principal factor affecting the landslide occurrences (Pradhan and Lee, 2010). Hillshade map provides a good mage showing terrain movement, and this map facilitates landslide mapping (Olaya, 200). The accuracy of a DEM accuracy and its capability to represent the surface are affected not only by terrain morphology and sampling density but also by the interpolation algorithm (Barbarella et al., 2013). In the current study, hillshade, intensity, height (nDSM), slope, and aspect were derived from the LiDAR-based DEM, orthophotos, and texture information were utilized for detecting landslides, as shown in Fig. 3.

Fig. 3. Shows LiDAR derived data (A) Orthophotos (B) DTM (C) DSM (D) Intensity (E) Height (F) Slope (G) Aspect.

2) Image Segmentation

is the initial and prerequisite step in object-based analyses because it determines the sizes and shapes of image objects (Duro et al., 2012). The selection of the appropriate parameters of image segmentation relies on the selected application, the environment under analysis, and underlying input imagery (Blaschke et al., 2010). Multiresolution segmentation is a bottom-up region-merging technique that merges the most similar adjacent regions as long as the internal heterogeneity of the resulting object does not exceed the user-defined threshold of the scale factor (Benz et al., 2004). The multiresolution segmentation was carried out on the basis of color, scale and shape including compactness and smoothness of the shape using the eCognition software (Definiens et al., 2007). Considering the complex charcteristics of landslides such as variations in land cover, differences in illumination, divers it of spectral behavior, and size variability, it is difficult to delineate each individual landslide as a single object (Martha et al., 2010).

Three parameters(i.e., scale, shape, and compactness) should be identified in this algorithm. The values of these parameters can be determined using the traditional trial-and-error method, which consumes considerable time and demands extensive work (Pradhan et al., 2016). Therefore, various automatic and semiautomatic methods for identifying the optimal parameters have been explored (Sameen and Pradhan, 2017; Martha et al., 2011; Anders et al., 2011; Belgiu and Drăguţ, 2014; Drăguţ et al., 2010). The Taguchi optimization method proposed by Pradhan et al. (2016) and the fuzzy logic supervised approach (i.e. Fuzzy-based Segmentation Parameter optimizer (FbSP optimizer) presented by Zhang et al. (2010) are among the advanced methods for the automatic selection of segmentation parameters. Nevertheless, delineating image objects at various scales remains a challenge. Furthermore, not all selected features are completely exploited using a particular segmentation scale. Accordingly, an automatic method should be directly implemented.

3) Generating Training Datasets

According to Ma et al. (2016), landslide inventory should be used to obtain prior knowledge of all the sites, and this step is a prerequisite in the stratfied random sampling method. Thus, in this present study, the segmentation scale was optimized, and the landside inventory map was overlapped with the segmented layer in order to label the classes. The sample sets were constructed automatically at optimal scale of the analysis area by using ArcGIS 10.3 software. Subsequently, stratified random sampling was implemented on the labeled objects. This process was conducted and repeated 20 times at optimal scale. In this study, the training sets were evaluated, a training set with 70% of the training samples (i.e., 30% testing set). Stratified random sampling is recommended to acquire an adequate training set size for every class without any bias during sample selection (Ma et al., 2016). Therefore, this sampling model was adopted in this study to evaluate the training samples and achieve improved results without strong bias.

4) Image Analysis Approaches.

According to Li et al. (2016), Support Vector Machines(SVM) and Random Forest(RF) are suitable for object-based techniques. Hence, the tendency of the overall accuracies declining with increase segmentation scale is confirmed. Therefore, SVM and RF classifiers were used for evaluating performance of six feature selection methods.

(1) SVM Model

A supervised non-parametric statistical learning technique was used to categorize the data set into groups in the manner consistent with training examples. SVMs are gaining popularity in the remote snsing field, including landslide mapping (Heleno et al., 2015; Van den Eeckhaut et al., 2012; Chang et al., 2012; Moosavi e al., 2014), due to their ability to handle data with unknown statistical distributions and small training data sets as obtainable in the field (Mountrakis et al., 2011). Dou et al. (2015) found that SVMs with a small training dataset was more accurate and stable than the maximum likelihood, decision tree, and artificial neural network classifiers with large training data sets. SVMs are binary classifiers whose aim is to find the decision region boundary that separates the data set characteristics or features into two regions in the feature space. The SVM chooses the boundary optimal hyperplane that exhibits the maximum safety margin to the closest training features refer to as support vectors which maximizes the margin between the classes (Heleno et al., 2016). The linearization of the decision boundary was achieved through the use of kernel functions that maps the training data into higher-dimensional space capable of linearly separating the two classes of hyperplane (Pawłuszek and Borkowski, 2016). The SVMs perform a nonlinear transformation of covariates into high-dimensional feature space (Pradhan and Lee, 2010). In case of linear data separation, a separating hyperplane could be defined as follows (Eq. 1):

y_i (w.x_i + b) ≥ 1 – δ_i       (1)

i is the input vector, y_i is the desired output. The. In order to determine the optimal hyperplane, the Lagrangian multipliers must first be solved (Samui, 2008) (Eq. 2 and Eq. 3).

 

\(\text { minimize } \sum_{i=1}^{n} \alpha_{i}-\frac{1}{2} \sum_{i=1}^{n} \sum_{i=1}^{n} \alpha_{i} \alpha_{j} y_{i} y_{j}\left(x_{i} x_{j}\right)\)       (2)

\(\text { subjected to } \sum_{i=1}^{N} \alpha_{i} y_{j}=0,0 \leq \alpha_{j} \leq C\)       (3)

where a_i are Lagrange multipliers, C denotes the penalty, and the slack variables δi allow violation of the penalized constraint. The decision function, which is used to classify new data, is illustrated in Equation 4.

\(g(x) \operatorname{sign}\left(\sum_{i=1}^{N} y_{i} \alpha_{i} x_{j}+b\right)\)       (4)

In some cases, determining the separating hyperplane is impossible through the four available basic kernel – linear(LF), polynomial(PF), radial basis (RBF), and sigmoid (SF) functions. LF is the simplest one; PF is non-stationary and well suited when all trainng data are normalized; SF is from the field of neural networks; and RBF depends on the distance from the origin as shown in Eq. 5.

\(K\left(x_{i}, x_{j}\right)=\exp \left(-\gamma\left\|x_{i}, x_{i}\right\|^{2}\right), \gamma>0\)       (5)

In this study, SVM was implemented using e1071 package Meyer et al. (2014) within the R statistical computing software (RDevelopment CORE TEAM, 210). The performance of a SVM classifier depends on its hyperparameters. Therefore, selection of these parameter was optimized and their sensitivity was analyzed. In the case of SVM, three parameters were evaluated namely kernel function, penalty parameter (C) and gamma parameter (γ). The best prediction accuracy was achieved with the Radial basis function (RBF), using Gamma parameter (γ) 0.9 and penalty parameter of 300. This was carried out in an expedite manner, by visual inspection of the match between results and reference data. The 70% of the inventory map together with all the features were selected as training sets to train the RF model.

(2) RF Mode

The RF algorithm developed by Breiman et al. (2001) is a nonparametric ensemble learning method based on several decision trees for classification or regression. This supervised method has been successfully applied in the detecting landslide using various types of remote sensing data (Chen et al., 2017; Chen et al. 2014; Stumpf and Kerle, 2011). The algorithm constructs multiple decision trees based on randomly selected subsets of the training dataset. In a classification problem, the RF exploits the high variance among individual trees. It assigns the respective class according to the majority votes. The main advantage of this method is the reasonable performance on complex datasets with less efforts of fine-tuning (Stumpf and Kerle, 2011). A RF is considered a random subset of the original set of features, whereas a classification and regression tree considers all variables in each node. Users can estimate the number of variables per node by using the square root of the total variable number. Two mechanisms, namely, sampling and using random variables in each node, generate significantly different uncorrelated trees. Moreover, having a relatively large number of trees is necessary to derive the variability of the training data and achieve high-accuracy classification. A feature is assigned to a class by considering the votes of all the trees in the forest. The class will then be assigned on the basis of the majority vote. (see Eq.6).

\(c T \sqrt{M N \log N}\)       (6)

where c is a constant dependent of data complexity (i.e., small or large dataset), T is the number of tree, M is the number of variables, and N is the number of instances (Breiman, 2003).

In this study, the RF package(Liaw and Wiener, 2002) an open-source statistical language R (R Development Core Team 2013) was used. Two parameters were used: the number of variables in the random subset at each node and the number of trees in the forest. The number (500) was selected for trees in this study which is usually used for the RF classifier as reported by Stumpf and Kerle (2011), while one randomly split variable was used to make the trees grow. The 70% and 30% were used for the inventory map together with all the features as training sets to train the RF model and evaluation of the classification accuracies respectively. The mean and stdev values of the classification accuracies were then drawn from 50 random runs.

5) Features Selection

Feature selection methods are divided into a filter, wrapper, and embedded methods (Ladha and Deepa, 2011). The first method is suboptimal and independent of the classification algorithm, and this method requires less computation for a large dataset(Ladha and Deepa, 2011). The second method measures the feature set using the classification method itself; thus, the selected features depend on the classifier model used. This method is time-consuming and complex because each considered feature should be evaluated with the classifier algorithm used (Saeys et al., 2007). The performance of embedded method declines as the number of introduced irrelevant features increases. Compared with the wrapper method, this method depends on the classifier algoithm and requires short computational time, and it is relatively robust against overfitting (Srivastava et al., 2014). Feature selection algorithms can effectively reduce the number of features and enhance the accuracy of results (Li et al., 2015). Handling a large number of features is undesirable because the irrelevant input features may lead to overfitting (Chen et al., 2014). By contrast, the selection of a small(possibly minimal)feature set leads to the best possible classification results (ursa and Rudnicki, 2010). Significantfeaturesshould be selected to improve the results of landslide detection in a certain area (Kursa and Rudnicki, 2010). Van Westen et al. (2008) asserted that selecting relevant features is important in distinguishing landslides from nonlandslides and classifying landslides. Stumpf and Kerle (2011) reported that the results obtained after reducing features present improved accuracy. Numerous studies (Borghuis et al., 2007; Chen et al., 2014; Danneels et al., 2007; Moine et al., 2009) have examined different feature selection techniques for landslide detection. The results of these studies revealed that valuable information could be obtained using relevant features.

(1) ACO

TheACOis a powerful metaheuristic and optimization technique used for parameter optimization due to its ability to eliminate influence of expert subjectivity. It superior performance can be attributed to its parameters such as mutation, crossover and srvival of chromosomes. Also, derivative information and step size calculation are not necessary in ACO (Ladha and Deepa, 2011). (Dorigo and Stützle, 2003)reported that pheromone evaporation aids in the prevention of rapid convergence of the algorithm toward a suboptimal region. It can perform robust and flexible search for a good combination of terms involving values of the predictor attributes (Parpinelli et al., 2002). This approach has been applied conveniently in many remote sensing applications like parameter selection (Alwan and Ku-Mahamud, 2012), feature extraction (Li et al., 2012), feature selection (Sameen et al., 2017) and image segmentation (Cao and Xia, 2007).

The overall workflow of ACO-based feature selection is presented in Fig. 4 and the process begins with the generation of a number of ants, which are then placed randomly on a graph, i.e., each ant starts with one random attribute. The number of ants located on the graph may be set equal to the number of attributes within the data in which each ant initiates path construction at a different attribute. From their initial positions, the ants traverse nodes probabilistically until a traversal stopping criterion is satisfied and resulting subsets are collected and evaluated. If an optimal subset is found in certain number of times, the process stops, and the best attribute subset encountered is outputted. If none of these conditions hold, the pheromone is updated and a new set of ants is created an the process is reiterated.

Fig. 4. ACO-based attribute selection workflow.

(2) GR

The gain ratio is an extension of the information gain measure, which attempts to overcome the bias that the information gain measure is prone to selecting features with a large number of values (Han et al., 2011). Thereby, the information gain measure is used as an attribute selection measure of the decision tree and is obtained by computing the difference between the expected information requirement, cassifying a tuple in tuples, and the new information requirement for attribute A after the partitioning. The measure of the expected information requirement is given by (Han et al., 2011) (Eq.7)

\(\operatorname{Info}(D)=-\sum_{i=1}^{m} p_{i} i \log _{2}\left(p_{i}\right)\)       (7)

where m is the number of distinct classes; pi indicates the probability by calculating the proportion of belonging to classC_i in tuples D. The new information requirement for attribute A is measured by (Eq. 8).

\(\operatorname{Info}(D)=-\sum_{j=1}^{v} \frac{\left|D_{j}\right|}{|D|} \times \operatorname{Info}\left(D_{i}\right)\)       (8)

where v indicates that D was divided into v partitions or subsets, {D1, D2, …, Dv}. Thus, the information gain measure Gain(A) for attriute A can be calculated by the formula (Eq. 9).

Gain(A) = Info(D) – Info_A(D)       (9)

Then, a ‘split information’ function was used to normalize the information gain measure Gain (A). The split information function was defined by (Eq. 10).

\(\text { SpliteInfo }_{A}(D)=-\sum_{j=1}^{v} \frac{\left|D_{j}\right|}{|D|} \times \log _{2}\left(\frac{\left|D_{j}\right|}{|D|}\right)\)       (10)

Finally, the gain ratio is calculated as the information gain measure Gain(A) divided by the split information measure SpliteInfo(A), as shown in (Eq. 11).

\(\operatorname{GainRatio}(A)=\frac{\operatorname{Gain}(A)}{\operatorname{SpliteInfo}_{A}(D)}\)       (11)

The larger the gain ratio obtained, the more important the represented features are.

(3) PSO

In 1995, Kennedy and Eberhart proposed PSO as a technique which was motivated by social behavior like fish schooling and birds flocking. PSO relies on optimization through social interaction in a population which depends on personal and social behavior. The useful features from the available features of eCognition software were selected using PSO optimization implemented in MATLAB which was used to minimize the error rate. The fitness function is gien in (Eq. 12) which is used to minimize the classification error rate obtained by the selected features during the evolutionary training process and the number of selected features(Xue et al., 2013; Sameen and Pradhan, 2017).

\(F=\left\{\begin{array}{l} w \times \operatorname{Train}_{E R_{s}}+(1-w) \times \operatorname{Train}_{E R_{s}} \\ \times \frac{\# \text { Feature }}{\# \text { All Features }}+(1-\alpha) \times \frac{E R_{s}}{E R_{\text {all }}} \end{array}\right.\)       (12)

where w is the weight of the classification error rate obtained from the training data and w ∈ [0,1], \(Train_{ER_s}\)  represents the classification error rate gotten from the selected feature subset and the training subset data, \(Test_{ER_s}\) denotes the classification error rate obtained from the selected feature subset and the testing subset data, α is the weight of the number of selected features, #Features represents the number of selected features, #All Features represents the total number of features available or classification, ER_s is the classification error rate obtained from the selected features, and ER all is the classification error rate obtained from all the available features. The error rate of the classification results are calculated from (Eq.13) (Sheikhpour et al., 2016), as follow:

\(\text { Error Rate }(E R)=\frac{F P+F N}{T P+T N+F P+F N}\)        (13)

where TP, TN, FP, and FN represent true positives, true negatives, false positives, and false negatives, respectively.

(4) GA

Goldberg (1989) defined GA as a stochastic class search and optimization techniques based on evolutionary principles and natural selection. When using GA for feature optimization, the feature attributes are coded as chromosomes in a type of binary string. The populations are initially randomized before the search process commence and the searched would then determine the encoded chromosomes to take full advantage of the optimal fitness function, which was computed for each of the randomly originated chromosomes. Because designing, the optimal fitness function remains a key player in improving efficiency of the search space. If improper fitness function are selected in a local optimum, it can lead to decrease the search effectiveness (Tang et al., 2005). The fitness function f(x) can be expressed as shown in (Eq. 14 and Eq. 15) which facilitates the assigning of optimal fitness value for each chromosome.

\(f(x) \frac{\sum_{i=1}^{n} \sum_{j=1\left(\omega_{i}=\omega\right)}^{n} \delta\left(x_{i}, x_{i}\right)}{\sum_{i=1}^{n} \sum_{j=1\left(\omega_{i} \neq \omega\right)}^{n} \delta\left(x_{i}, x_{i}\right)}\)       (14)

\(\delta\left(x^{i}, x^{i}\right)=\sqrt{\sum_{k=1}^{n} \omega_{k}}\left(x_{k}^{i}-x_{k}^{i}\right)\)       (15)

where xⁱ represents an n-dimensional feature vector of image object i, xⁱ =\(x_{1}^{i}, x_{2}^{i}, \ldots, x_{n}^{i}\) and δ(xⁱ , xⁱ ) represents the Euclidean distance between vectors xⁱ and \(x_{k}^{i}\), which is k-th feature value of the i, ω_k is the weight of the k-th feature, and n is the number of objects in feature optimization. GA optimization was carried out using MATLAB software and are used to compute the optimal fitness value of each individual and only the optimal individuals survives under this condition. Therefore, an optimized generation process can be used to reproduce generations through crossover or mutation. Eventually, to passage a discrimination related to the fitness, the optimal individuals were decoded for use and corresponded to feature selection as inputs for landside detection and classification in the CBR process.

(5) CFS

The feature subset was evaluated using filter algorithm unlike the feature evaluation methods aforementioned which is a correlation-based Feature Selection (CFS. The CFS measured the worth of a set of features using heuristic evaluation function based on the correlation of features which is consistent with assertion by Hall and Holmes(2003) who reported that a superior subset of features should be correlated with classes highly uncorrelated to each other. Thus, the criterion of a subset can be evaluated using (Eq. 16).

\(r_{c z}=\frac{K r_{z i}}{\sqrt{K+K(K-1) r_{i i}}}\)       (16)

where r_zc represents the correlation between the summed feature subsets and the class variable, k is the number of subset features, r_zi is the average of the correlations between the subset features the class variable, and r_ii is the average inter-correlation between subset features. In addition, the best first search was used to explore the feature space, and the five consecutive fully expanded non-improving subsets were set to a stopping criterion to avoid searching the entire feature subset space. In this study, the WEKA package was used to implement this feature selection algorithm.

(6) RF

The feature evaluation method was based on random forest known as an embedded method (Pal and Foody, 2010) which provides a variable importance criterion for each feature by computing the mean decrease in the classification accuracy for the out of bag (OOB) data of the bootstrap sampling (Verikas and Gelzinis, 2011). Assuming bootstrap samples b = 1, … B, the meandecrease in classification accuracy D_j for variable x_j as the importance measure is given in (Eq. 17).

\(D_{j}=\frac{1}{\mathrm{~B}} \sum_{b=1}^{B}\left(\mathrm{R}_{b}^{\mathrm{OO} b}-\mathrm{R}_{b j}^{\mathrm{OOb}}\right)\)       (17)

where \(R_b^{OOb}\) denotes the classification accuracy of OOB data \(l_b^{OOb}\) using the classification model T_j; and \(R_b^{OOb}\) is the classification accuracy of OOB data \(R_b^{OOb}\) permuted the values of variable x_j in \(l_b^{OOb}\) (j = 1, … , N). Finally, a z-score of variable x_j representing the variable importance criterion could be computed using the formula \(z_{j}=\frac{D_{j}}{S_{j} \sqrt{B}}\) , after the standard deviation s_j of the classification accuracy decrease is calculated. In this study, the feature evaluation procedure was performed automatically using the R package ‘RRF’.

4. Results of Proposed Methodology

1) Optimizing Segmentation using Supervised Approach

Supervised approach (i.e. Fuzzy-based Segmentation Parameter optimizer (FbSP optimizer) was used to optimize the parameters like shape, scale and compactness of the ultiresolution segmentation algorithm shape, and compactness, respectively in the analysis area (Fig. 5(a) and (b)). These parameters were visually assessed to achieve an over segmentation. After a few iterations with these initial values, the optimal values obtained for scale, shape, and compactness were 65.37, 0.34, and 0.58, respectively (Fig. 5(c)).

Fig. 5. Shows the result of the segmentation using optimized parameters for the analysis area. It can be seen that the landslide objects were accurately delineated highlighted by red color.

According to the various man-made and natural objects found in the scene such as landslide, cut slope, trees, and bare land, three segmentation levels were found to be necessary. In each class, there are also several types of the same object, which suggested using different levels of segmentation. For example, the cut slope class which represents the previously landslides that maintained by slope engineers have different sizes according to the type and size of the landslide occurred.

The supervised approach presents more significant improvement in terms of time used for the segmentation. It only requires few minutes to achieve optimal segmentation parameters for the landslide segmentation as demonstrated in Fig. 4 which shows improvement in terms of speed. The optimized parameters allowed increasing the overall classification accuracy of delineating the landslide boundaries. As the boundary of he objects could be accurately delineated, the computation and utilization of spatial and textural of the image objects were improved.

2) Selection of the best Classification Features by Various Algorithms

Once the image objects have been created, among the several classifications features available in eCognition software, the best subsets were selected through ACO, GR, PSO, RF, GA, and CFS methods. The best feature subset selection aimed to distinguish the landslide objects from non-landslide objects with high classification accuracy. The number of features in a subset to be selected was set to be lower than the number of samples in the landslide inventory map to avoid overfitting and reduce model complexity (Yu et al., 2006).

Overall, in order to detect the landslide locations 82 features were Mean and StdDev (Intensity, DEM, DSM, Slope, Aspect, Height ), texture information All directions,0o,45o,90o,135o (Gray-level co-occurrence matrix (GLCM) correlation, GLCM Dissimilarity, GLCM angular second moment, GLCM StdDev, GLCM Mean, GLCM Contrast, GLCM Entropy, GLCM Homogeneity, GLDV angular second moment, Grey level difference vector (GLDV) Mean, GLDV Entropy and GLDV Contrast) and Mean and StdDev (Red, Green and Blue, Max. diff, and Brightness). The values of these features are expressed in mean and standard deviation (StdDev). Many the features were initially removed from the analysis due to the landslide class of the study area and only those that ave the possibility of transferability were selected. The six feature selection algorithms selected the same features but provided different ranks (i.e. combination) and subsequently yield different landslide detection accuracy (Table 1).

Table 1. Multi- resolution segmentation parameters

The RF and SVM model were used in evaluation the process, and the inventory map was divided into training (70%) and testing (30%) sets. The values of their parameters should be placed in other to apply the six feature selection methods. In accordance with the preliminary analysis and previous studies (Sameen et al., 2017; Chen et al., 2016; Li et al., 2016; Duo et al., 2015; Karegowda et al., 2010; Kumar et al., 2006), these parameters were selected and are found to be suitable for this research. Results of ACO, GR, PSO, RF, GA, and CFS approaches with 70% of the inventory data were evaluated according to the overall accuracy (Table 2). The highest landslide detection accuracy (91.00%) was achieved by using the features selected by CFS trained and evaluated with RF and SVM models. Furthermore, the algorithms ACO and RF indicate better performance than the GR and PSO methods in both models (RF and SVM). This implies that large number of features does not signifies accurate in the landslide inventory map. It was also observed that the ideal number of features is 11 among 82 available features (see Table 2).

Table 2. Optimal features slection for detecting landslide using various algorithms

Table 2 showed that several features, such as the StdDe DTM and GLCM Homogeneity were found to be the most important features for distinguishing landslide objects from other objects in the scene. In addition, based on the results of other feature selection methods, the most important features were slope and texture information represented inGLCMHomogeneity, GLCM correlation, and GLCM angular second moment.

The effect of the number of iterations on feature selection was also analyzed. Table 2 shows the 11 features selected in each iteration by using six methods. This experiment was executed for the best subset (11 features) and high accuracy was achieved. Different features were selected as optimum in each iteration. The result indicates that the same classification accuracy can be achieved with different feature combinations. Therefore, it is not sufficient to select only the significant features, combination of features should be selected as implemented in this study.

3) Results of the SVM and RF Models

First, the 70% training set was used to train the SVM model on the analysis area,site-1, and site-2. When all the features were utilized, the qualitative assessment results were poor. Furthermore, the quantitative assessment showed that the overall accuracies of the results were 74.73%, 71.09%, and 66.57% for the study area, site-1, and site-2, respectively. On the contrary, te SVM model that used the optimal features generated high-quality results in the qualitative assessment and performed accurate identification of the locations of landslides. The quantitative assessment demonstrated that the overall accuracy of the SVM model using the optimal features was 87.34% for the analysis area, as shown in Fig. 6. Fig. 6 also shows that the classification results for site-1 and site-2 achieved overall accuracies of 86.82% and 84%, respectively.

Fig. 6. Shows the results of support vector machine (A) analysis area (B) Test sit-1 (C) Test sit-2.

The results of the qualitative assessment of the RF model were of poor quality. The overall accuracies in the quantitative assessment were 77%, 72.83%, and 68.78% for the analysis area, site-1, and site-2, respectively. These results were obtained when the 70% of the training set and all the features were used to train the RF classifier. When the same training set ratio (i.e., the 70% of the training dataset) and only the optimal features were used, the RF model produced high-quality results and accurately identified landslide locations in the qualitative assessment. In the quantitative assessment of the analysis area, the overall accuracy and kappa coefficient were 91% and 0.84, respectively (Fig. 7). Fig. 7 Shows that the classification results for site-1 and site-2 achieved overall accuracies of 88.68% and 86%, respectively.

Fig. 7. Shows the results o random forest (A) analysis area (B) Test sit-1 (C) Test sit-2.

The results of the RF and SVM models demonstrated that using the six algorithms in feature selection and applying the optimized segmentation parameters with the use of high-resolution iDAR, orthophotos, and texture information enhanced the models performance and improved the transferability of the RF and SVM models.

4) Transferability of Models

The training samples were evaluated using a stratified random sampling method. The 70% of training set was applied to train the RF and SVM models with either all features or only the optimal features. When the RF and SVM models using all the features were applied, overall accuracies of 77% and 74% were achieved, respectively (Table 4). As shown in Table 3, the overall accuracies of the RF model (SVM model) for site-1 and site-2 were 72.83% (71.09%) and 68.78% (66.57%), respectively. When RF and SVM models using only the optimal features were used for the analysis area, the overall accuracies of the results were 91% and 87.34%, respectively. The overall accuracies for site-1 and site-2 of the RF model (SVM model) were 88.68% (86.82%) and 86% (84%), respectively (Table 3).

Table 3. Results comparison based on Overall accuracy and Kappa coefficient for important features and full features using RF and SVM algorithms

Table 4 shows the results of the user’s and producer’s accuracies of the RF nd SVM classifiers using either only the optimal features or all features for the analysis area,site-1, and site-2. The results showed that the RF classifier exhibited higher accuracies for all the mentioned areas.

Table 4. Results comparison based on User's Accuracy and Producer’s Accuracy for important features and full features using RF and SVM algorithms

5) Evaluation via Precision/Recall method

One of the well-known methods for quantitative counting accuracy assessment is Precision/Recall method. The proposed method was evaluated using field observation for each block (Nyland, 1996) (see Eq. 18, Eq. 19 and Eq. 20).

\(\text { Precision }=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}\)       (18)

\(\text { Recall }=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}\)       (19)

\(\mathrm{F}-\text { measure }=\frac{(1+\alpha) \times \text { Precision } \times \text { Recall }}{\alpha \times \text { Precision }+\text { Recall }}\)       (20)

Where a True Positive (TP) is the number of correctly detected landslide. A False Negative (FN)is a landslide that is not detected. A False Positive (FP)shows a pixel that is recognized as a landslide but it is something else. The &lpha; is a non-negative scalar. In this study, α is set to 0.5 as suggested in Lin et al. (2011). Also, the success rate can be computed using another equation which is to determine the positive counted rate by dividing segmented numbers with total trees.

Table 5 displays the results of the proposed method, which achieved very high accuracy assessment in all studies. This demonstrates that the proposed method can be transferrable to another spatial dataset with the same climate and condition.

TABLE 5. Performance evaluation of RF and SVM Models.

In this current study, the accuracy assessments also were performed in two categories of quantitative and qualitative to measure the precision of applied methods. Precision/Recall method was applied for measuring the accuracy of landslide detection quantitatively. The actual number of landslide events were collected via field surveying as a reference. Then, the results of RF and SVM models were compared to landslide inventory. The landslides would be counted as a corrected detection, as long as it is recognized in segments by weather bigger size of segment border or smaller. The key point in landslide counting is having even a single segment on occurred landslide and area of landslides are not considerable in mentioned assessment.

F-measure stands for overall accuracy in counting landslide detection, showed a consistence result in analysis area and tested areas(i.e. Tessite-1 and 2)for two models(i.e.RF and SVM). RF however, exhibited the highest results for landslide detection in all aforementioned areas as illustrated in Table 5. While, SVM achieved low accuracy in detecting the landslide (see Table 5). Thus, based on the assessed accuracy measurements proposed methodology improve the landslide detection analysis quantitatively and qualitatively.

5. Discussion

This research uses six feature selection methods, object-based technique and LiDARdatato improve the accuracy of landslide inventory mapping. It was sufficient to optimize segmentation parameters like scale, shape, and compactness using FbSP optimizer in delineating landslide boundaries. Because, optimized segmentation parameters facilitate the generation of accurate objects segment and uses spatial and texture features to distinguish another land cover classes and reduces the influence of under and over segmentation. Since landslides can be classified according to their features, accurate segmentation is essential for differentiating between the classes. Even though segmentation results could prove difficult sometimes due to shape of objects, optimization approaches can be used to improve its accuracy (Pradhan and Mezaal, 2017). Among these methods, the selection of a subset with optimal features can significantly improve the results of classification because non significant features could have redundant information and subsequently degrade classification accurac.

On the basis of the results of the RF and SVM classifiers, three of algorithms, namely, CFS, ACO, and RF, exhibited the highest ranks in landslide detection, respectively. The results of CFS showed the high classification accuracy, it was a rapid and time effective method. Among all the compared feature selection algorithms, ACO is second powerful technique for selecting a subset from available features effectively. For the RF algorithm, the ranks of RF also illustrated the significance of subset features for enhancing the results of accuracy. These methods are easier as their mathematical operations can be solved with the primitive mathematical operators. They are cost effective as their application does not require high speed or memory. Moreover, their basic concept of these above-mentioned methods is simple that their ideas can be summarized in simple code which are made up of few lines.

The result indicated the importance of this step in detecting landslides in the OBIA framework. Using feature selection for object detection can reduce computational complexity, eliminate the irrelevant features, reduce the dependence on subjective expert knowledge, simplify of the developed rules, and improve the model. Distinguishing between landslides and other landcovers in densely vegetated terrains and hilly areaslikeCameron Highlands(cutslope, bare soil and man-made slopes) could be quite challenging. Therefore, the transferability results of the feature selectionform of the analysis area to the sites (site-1 and 2) were tested as presented in Fig. 6B, C and 7B, and C. It was observed that the location of landslides was separated by using the relevant feature as shown in Table 3. According to Stumpf et al. (2011), the overall accuracy of landslide detection applied to other areas could decrease even if the same method was used in the development of the model. This reduction in accuracy could be due to difference in landslide characteristics and environmental conditions Also, spatial resolutions of images, differences in the sensors used, and illumination conditions could be contributed to the challenges reported recently (Rau et al., 2014).

In this context, landslide detection techniques are appropriate for generating a well-organized landslide inventory map that is useful for qualitative/quantitative hazard assessment. Various methods for landslide mapping have been proposed; however, no method has effectively revealed the ideal results. Li et al. (2015) identified landslide by using LiDARdata, object based image analysis and random forest. The overall accuracy of their study was 89%. Dou et al. (2015) integrated object-based approach and a Genetic Algorithm (GA) algorithm with help of LiDAR data for landslide detection. The quantitative assessment of showed that the overall accuracy of their study was 87%. The reasons are that LiDARintensity data was not involved in their work and not all spectral and spatial informatin of object because unfitting delineation of the segmentation for landslide object. Therefore, many misclassifications can be seen in the results.

It is absolutely necessary to take the required measures in other to avoid landslide separation from the most similar land cover classes(i.e.man-made, bare soil and cut slope). The morphological characteristics of landslide map differs. Forinstance, slope, the shape, and other characteristics such as dip direction, texture, width and length of the surface terrain could change after landslide. Therefore, by using relevant features derived from very high resolution LiDAR data and texture and geometric features can be used to distinguish between landslides and bare soil. In addition, applying different optimization techniques helped us to improve the classification accuracy in landslide detection over other landcover classes, such as bare land, cut slope, etc., as described previously by Pradhan and Mezaal(2017). Their results demonstrated that using optimized techniques with very high resolution LiDAR data enabled them to separate landslide from other types of land cover. Furthermore, Mezaal et al. (2017a) suggested that using the object feature from LiDAR data is suitable for resolving landslide identification issues.

6.Validation

The reliability of the proposed method was further validated by conducting a field investigation. A handheld GPS device (GeoExplorer 6000) was utilized to identify landslide locaions, as shown in Fig. 8. A more detailed information like source area, direction, run out, volume and deposition was obtained from insitu measurements which proves the reliability of the inventory map produced in the field using GeoExplorer 6000 handheld GPS. All the information obtained from the field measurements allowed for the assessment of the precision and reliability of the produced landslide inventory map. The field investigation confirmed that the landslides detected using the proposed method was accurate. Thus, the proposed method can identify landslide locations and produce a credible landslide inventory map for the Cameron Highlands in Malaysia.

Fig. 8. Landslides locations in the study area; (a) Taman Mawar, Kuala Terla and (b) Jalan Tapah-Ringlet.

7. Conclusion

This study indicated the importance of using the optimized parameters of multiresolution segmentation to achieve the highest overall accuracy, as they allow for the accurate delineation of landslide boundaries. The RF results model were much more accurate than the results of SVM model when either all features or only the optimal features were used.

The quantitative assessment revealed that the overall accuracy of the RF model and (SVM model) using the optimal features were 91% and (87.34%) for the analysis area, while, the results for site-1 and site-2 achieved overall accuracies of 88.68% and 86%, (86.82%) and (84%), respectively. Moreover, the resultsof the transferability model showed that, RF and SVM models were used for the analysis area, the overall accuracies of the results were 91% and 87.34%, respectively. The overall accuracies of RF model (SVM model) were 88.68% (86.82%) and 86% (84%), for site-1 and site-2 of the, respectively.

The algorithm with the highest ranks in feature selection for landslide detection were CFS, ACO and RF. The feature selection algorithm reduced the dimensionality of the object features, expedited the training of RF and SVM classifies, and imprved the classification accuracy of these classifiers. The SVM classifier was more sensitive to the feature selection than the RF classifier. In addition, field investigation was applied for performing the second round of validation. Therefore, the proposed method is suitable for the accurate identification of landslide locations and production of reliable inventory maps, which are crucial to avoid disasters in urban areas. The results indicate that the significance of the relevant selection achieved from very high-resolution airborne laser scanning data, visible bands, texture features for improve the detecting of the locations of landslide. Overall, using various feature selection algorithms and a supervised approach based on RF and SVM models yielded robust results and increased the efficiency and cost effectiveness of the development of landslide inventory maps with the use of high-resolution LiDARderived data, orthophotos, and texture iformation.