Landslide Detection Using a Saliency Feature Enhancement Technique From LiDAR-Derived DEM and Orthophotos

This study proposes a new landslide detection technique that is semi-automated and based on a saliency enhancement approach. Unlike most of the landslide detection techniques, the approach presented in this paper is simple yet effective and does not require landslide inventory data for training purposes. It comprises several steps. First, it enhances potential landslide pixels. Then, it removes the image background using slope information derived from a very high-resolution LiDAR-based (light detection and ranging) digital elevation model (DEM). After that, morphological analysis was applied to remove small objects, separate landslide objects from each other, and fill the gaps between large bare soil objects and urban objects. Finally, landslide scars were detected using the Fuzzy C-means (FCM) clustering algorithm. The proposed method was developed based on datasets acquired over the Kinta Valley area in Malaysia and tested on another area with a different environment and topography (i.e., Cameron Highlands). The results showed that the proposed landslide detection technique could detect landslides in the training area with a Prediction Accuracy, Kappa index, and Mean Intersection-Over-Union (mIOU) of 71.12%, 0.81, and 68.52%, respectively. The Prediction Accuracy, Kappa index, and mIOU of the method based on the test dataset were 65.78%, 0.68, and 56.14%, respectively. These results show that the proposed method can be used for landslide inventory mapping and risk assessments.


I. INTRODUCTION
Landslide is a destructive natural geohazard that poses significant damage to human life and property every year worldwide, for instance, 187 casualties were recorded by landslide events in Iran, imposing US$ 12,700,000 up to September 2007 [1]. A landslide in Northern Iran caused a huge loss in infrastructure estimated to be US$ 5,000,000 in January 2007 [2]. In Malaysia, from 1973 to 2007, landslides caused losses of approximately one Billion US$ for the period from 1973 to 2007 with over 100 deaths [3]. It is defined as The associate editor coordinating the review of this manuscript and approving it for publication was Weimin Huang . a mass movement under of gravity of earth, debris, or rock down a slope [4]- [6]. Even though preventing natural disasters is impossible, great efforts have been put into reducing their impact on society [7]. The preparation of landslide inventory maps is the basic step for landslide susceptibility mapping, hazard and risk assessment [8]. Landslide inventory maps show locations, along with attribute information, of landslides that occurred in a particular area [9]. Nevertheless, these sorts of complete inventory data do not always exist (i.e. impacts of landslides). A complete landslide inventory database contains precise locations, type, and volume of mobilized materials, date of occurrence, and the impacts of past landslides. In contrast, an incomplete landslide inventory database shows only the location of past landslides that have occurred in an area. The traditional way of landslide mapping is the interpretation of aerial photographs and field surveys. However, with the advancement of remote sensing systems such as very high-resolution satellite sensors and laser scanning systems, landslide mapping has been changed to include automated procedures and a large area mapping in relatively short times [8], [10]- [12].
The current approach consists of (i) Enhancement of potential landslide pixels, (ii) Removing image backgrounds using slope information, (iii) Removing small objects and separating landslide objects from non-landslide objects via morphological analysis, and (iv) Identifying landslide scars by Fuzzy C means (FCM) clustering method. The main contribution of this research is to evaluate the integration of the above methods in a framework that can detect landslides with limited or no inventories data.
The remainder of this article is organized as follows: Section II summarizes the related studies. Section III explains the proposed framework and methodology. Sections IV describes the experimental results. Section V provides the discussion. Lastly, section VI includes the conclusion and suggested future works.

II. RELATED STUDIES
In recent years, several techniques and algorithms have been proposed for landslide mapping using remotely sensed data. The most popular methods are based on object-based image analysis (OBIA) and pixel-based change detection [13]. The basic concept of OBIA methods is the delineation of colour, space, texture, and context at the object level and uses them for the detection of landslide scarps. Reference [14] presented an OBIA-based method for landslide detection in high-resolution, Resourcesat-1 Linear Imaging, and Self Scanning Sensor IV (5.8 m) multispectral image. The method accurately detected five types of landslides with an overall recognition accuracy of 76.9%. Reference [15] proposed a semi-automatic OBIA approach for landslide detection using SPOT images and a digital elevation model (DEM). The method was validated by Persistent Scatterer Interferometry and their results were promising. Moreover, [16] developed an object-oriented landslide detection approach using highresolution satellite images. They showed that their method could be effectively applied to areas with similar environments. Reference [17] presented an automatic case-based reasoning approach for landslide detection. In their method, OBIA concepts and genetic optimization were integrated to achieve their goal. In their work, the integrated method outperformed the traditional OBIA method reaching from 75% to 87%. Although OBIA methods can achieve reasonable accuracies in landslide mapping, they have some limitations such as segmentation optimization [14], transferability, and scale dependency [14], criteria for the optimal selection of remote sensing data [18] and shadow shortcoming in visual interpretation using NDVI (normalized difference vegetation index) and 3D digital stereoscopy for mapping [19].
Besides, landslide mapping by change detection methods requires multi-temporal remotely sensed data. For instance, [11] improved the landslide classification based on change detection in landslide events using a semi-automatic technique. They implemented a maximum likelihood probability task and a multi-threshold strategy using an automatic histogram division. The evaluation of the EI (Error Index) for three approximations, namely T-global, T-grid, and T-SU (Slope Unit) was carried out with metrics of 0.512, 0.510 and 0.487, respectively. The result showed that SU outperformed the regular grid results. The strategy improved the presentation of the landslide characterization with constrained training samples. However, the transferability of the technique was required. In another study, [20] developed a landslide inventory mapping method using change detection and image fusion technique. They used SPOT and IKONOS images to detect landslides at the regional scale. Their results indicated that the proposed method could achieve 70% accuracy. Furthermore, [21] proposed an object-based change detection method for landslide inventory mapping; the method achieved an accuracy of 81.8% applied to very high-resolution optical images. Reference [22] developed a local similarity measure for landslide detection and compared the results with the image differencing method; the method showed to be more reliable than the simple image differencing method. These methods required bitemporal satellite images, which are difficult to obtain in some areas such as tropical regions or require multi-scans of point clouds (e.g. multi-view), which are often very expensive and not practical. Subsequently, [23] developed a new technique based on saliency enhancement and morphological operations for landslide detection in large areas; the method was carried out using Landsat images acquired over Central Nepal and it performed reasonably well by detecting 99.1% of the landslides in the image; however, the authors indicated that more statistical methods should be developed to accurately separate background objects from landslide regions.
In another study, [24] developed a rule-based semiautomated approach for detection of landslides in Himalaya. The authors used IRS LISS3 satellite images by implementing decision tree methods; the results showed that the overall accuracy of the training and test areas were 76% and 75%, respectively. In another work [11], separating the riverbank from the background object to avoid misclassification of the riverbank was carried out. Authors in another research [25] classified shallow landslides from a high-resolution multispectral image and topographic data using a Bayesian framework; the framework was employed to enhance landslide mapping in Southern Taiwan, showing some advantages compared to other semi-auto and automatic approaches [8]. In the last decade, optical images, DEM, and LiDAR data have been used extensively in developing reliable algorithms for landslide detection. However, in recent years, LiDAR data is being more attractive due to its advantages over other data sources. Reference [26] compared LiDAR and aerial photographs for landslide detection. Their analysis showed that LiDAR is VOLUME 8, 2020 highly effective for accurate landslide delineation, is easier to interpret than aerial photographs, and is highly manipulatable concerning shadows and vertical exaggeration [27] evaluated the suitability of an aerial laser scanner for generating an optimal DTM for mapping landslides in the Cameron Highlands, Malaysia. Authors in [28] reviewed the use of LiDAR for landslide investigations. They discussed the feasibility of using accurate DEMs in landslide applications and inventory mapping. They pointed out that LiDAR data provides promising information for landslide scarps and displaced materials. Accordingly, geomorphological features derived from LiDAR-based DEM (scarps, mobilized materials and foot) can improve the mapping of landslide locations and characteristics.
In tropical areas, several methods based on LiDAR data were proposed for landslide detection. Authors in [29] developed an automated method for detecting landslides using LiDAR data and geomorphological indicators. They used the surface roughness index, the vegetation index, and break lines as indicators for landslide scarps in the study area; their results showed that the best spatial resolution of DEM is 2 and 3 m. Reference [30] presented an expert-based semi-automated method for geomorphological mapping in a mountainous area using LiDAR-based DEM. The authors found that high-resolution DEM is useful for extracting features that are important for landslide mapping in tropical areas. Reference [31] presented a data fusion technique using wavelet transform and Taguchi optimization for landslide detection in a tropical area. Weitao Chen et al. [32] developed a framework for landslide detection in Three Gorges, China utilising features derived from LiDAR such as mean aspect, DTM, slope textures, standard division filter of aspect. The features were combined with the Random Forest algorithm, causing a reduction in features set by 74% and improving the overall accuracy of the model by 0.44%. Tao Chen et al. [33], used the same reduction strategy, identifying significant landslide features (26 out of 124 features) and then integrated the model with mathematical morphology analysis to map the landslide locations; however, some uncertainties in the segmentation process and tuning parameters were reported. More recently, [34]- [36] used optimization techniques to detect landslides in dense vegetation regions using very highresolution LiDAR data. The LiDAR data and QuickBird satellite images were fused to detect landslides using the OBIA. The method achieved an overall accuracy of 90.06%, with 65.65% accuracy before fusion.
Despite the success of the landslide detection methods above, they have limitations such as (1) the need of a complete inventory data, which are not always accessible, and (2) they are an expensive and time-consuming process. 2) landslide types and their triggering factors vary from one area to another and different environment achieves relatively lower accuracy in most cases. 3) For accurate landslide detection, segmentation optimization [14] and transferable rules are necessary. 4) Refinement of OBIA rule-sets and other supervised landslide detection methods often require an adequate number of training landslides to avoid the problem of overfitting and to improve the generalization capacity of the method [37]- [40].
The current study attempts to develop a quick and practical technique for landslide detection, especially when inventories are scarce or not available. The proposed technique does not require landslide inventory data in the training process and is computationally more efficient than the presented methods especially in emergency cases [23]. Thus, the main aim of this paper is to introduce the efficiency of the presented landslide detection technique and evaluate its performance with different datasets.

III. METHODOLOGY A. STUDY AREA
In this research, the training area was selected in Kinta Valley, located in Perak state, East of Malaysia ( Fig.1). It is about 200 km north of Kuala Lumpur. Geographically, it is bounded between 103 • 3 5 E and 101 • 5 47 E latitudes and 4 • 39 11 N and 4 • 41 53 N longitudes. The area is classified as the tropical rainforest and the temperature is the same with little variations throughout the year with an average temperature of 28 • C (Meteorological Service Department of Malaysia). The area sees high rainfall throughout the year with an average of 200 mm each month and the average annual rain is 2428 mm. The region is susceptible to landslides and rockfall because of the loose sands and soft clays in the mining areas [41]. The land use map of the area shows that it includes residential, commercial, and industrial buildings, forest, mixed-grass, and open lands, as well as other infrastructures such as roads, power stations, and drainage sewers.
The test area was selected in the Cameron Highlands, which is 90 km away from the training area. Approximately 80% of the area is forest land, and it has an undulating topography, which makes it effective for performance assessment of the landslide detection models. The average altitude of the test area is 1,200 m above the mean sea level, and the total land area is 2.65 km 2 (Fig. 1). The geology of the Kinta Valley consists of a high percentage of igneous rocks, sedimentary (limestone) and metamorphic rocks (marble).

B. DATASET USED
This research used LiDAR data as the main source for deriving the digital elevation models and slope angle parameters for the training and test area. The details of the data acquisition for both areas are given in Table 1. Additionally, the landslide inventories which are needed for validation purposes were obtained and implemented through remote sensing techniques and technical researches according to previous studies for Kinta Valley and entire Malaysia. The data was collected from archive data of the Mineral and Geoscience Department of Malaysia [42]- [44]. A total of 13 and 21 landslides were identified in the training and test areas, respectively.

C. THE PROPOSED LANDSLIDE DETECTION TECHNIQUE
The proposed landslide detection technique and its performance evaluation and transferability assessment consist of three steps of data processing, as shown in Fig. 2. The first step included data acquisition. The second step was data preprocessing where the input data were prepared for analysis. In this step, LiDAR point clouds, orthophotos and the landslide inventories were geometrically corrected based on the Malaysian GDM 2000 coordinate system to ensure a proper analysis and validation task. Then, a very high-resolution DEM (0.5 m) was produced from the point clouds using the multi-scale curvature algorithm of ESRI's ArcGIS 10.2 software [45]. In the last step of data pre-processing, a raster map that contains the terrain slope angle, pixel-by-pixel was derived from the LiDAR-DEM with the same spatial resolution. The third step (landslide detection) method consists of four main steps including (1) image saliency enhancement, (2) image background removal, (3) morphology analysis, and (4) Fuzzy C-means (FCM) clustering. None of these steps needs supervision, thus no landslide inventories were used directly in training and developing the model. The available inventory data was only used for validation purposes. The first stage (image saliency enhancement) converted the orthophotos into a new image where the landslide features were enhanced using the symmetric surrounding regions [46]. The second stage removed the image backgrounds (nonlandslide regions) using slope information of the study area as an auxiliary step for the detection process. In this step, the slope map was reclassified into two classes, with slope either above or below 25 • .This indicator (threshold) was selected based on practical research in the same study area which referred that a slope above 25 degrees triggers landslides in the region [47]. The value of this threshold is region dependent, so it is expected that the transferability over the new test area may affect the validation accuracy.
In addition, the morphology analysis included a three-time iteration of dilating operation to remove small features and fill the gaps in the large non-landslide regions [48]. Dilating and eroding are two essential tasks of mathematics morphology to structure some compound activities and bring some viable morphology calculation [49]. After that, the FCM unsupervised clustering method was applied to classify the result into landslides and non-landslides where the label of the classes was determined based on orthophoto interpretation. There are two key parameters used in FCM clustering: the fuzzy exponent (m), which defines the fuzziness among each cluster; and (C) which is the number of clusters, which specifies the total number of clusters to be used. The combination of the two parameters affects the results of FCM clustering. The values of FCM clustering parameters were selected based on preliminary examination and previous landslide studies [50], [51] these parameters were selected as m = 1.9 and C = 3. Nevertheless, sometimes there are no optimal values for these parameters [50], [50]. Finally, the landslide detection technique was evaluated based on some statistical evaluations and its transferability to another area was investigated.

D. IMAGE SALIENCY ENHANCEMENT
Within the existence of urban features, it is difficult to separate between uncovered soil and urban highlights utilizing just spectral details. Also, the pixel-based examination frequently cannot precisely recognize landslide zones in the image. Notwithstanding, since particular landslides happen in uncovered (bare) soil areas, it can boost this area over urban features so that the identification process may succeed more efficiently via the pixel-based approach. This examination received a saliency improvement approach dependent on the most extreme symmetric surround which is developed by [50]. This method is mainly used to enhance nonvegetation regions, in particular, the bare soil areas of the current study area. The key advantage of this method is that it retains well-defined boundaries by preserving more frequent content from the original image. Other advantages involve in the implementation, simplicity and computational efficiency, particularly at emergency conditions that need a quick response [23], [50].
This saliency algorithm uses surround regions that are symmetric in pixels for which saliency needs to be computed. The advantage of doing so is that each pixel is treated to be at the centre of its sub-image (Fig. 3). For the orthophotos, the symmetric surround saliency value for a given pixel S ss (x, y) is obtained from the (1) as per the reference [46]: where I µ (x, y) is the average value of the CIELAB vector of the sub-image and I f (x, y) is the corresponding CIELAB image pixel vector in the Gaussian filtered version of the original image computed using (2): Note: CIELAB is a popular colour space to measure numerical changes in values of transitive and reflective objects.
where x o , y o , and A, are the offsets and area of the sub-image and computed using the following expressions: where w and h are the image width and height, respectively. Further details about the algorithm can be realized withing reference [46].

E. SLOPE-BASED IMAGE BACKGROUND REMOVAL
Each geographical location has particular predominant factors causing instability such as curvature, type of soil, manmade activities, etc. [52]. The significance of each factor (e.g. slope) varies with the other factors [2], [7], [53]. The slope of prone areas often plays an essential role in landslides occurrence. Accordingly, the slope raster map can be involved as an indicator to separate the landslide region from other bare soil regions. Several studies on the same study area suggested that regions with a slope higher than 25 • are highly prone to landslides [47], [54], [55]. Therefore, the current study reclassified the slope map into a binary raster (0, 1) using the slope indicator of 25 • as high prone-landslide areas; 0 indicates the other bare soil regions and 1 indicates the landslide regions. Subsequently, the non-landslide regions were removed as much as possible. Additional information about slope statistics in the training and test areas such as mean, standard deviation (SD), minimum, maximum, and range are provided in Table 2. Moreover, Fig.4 shows the slope distribution in both training and test areas.

F. LANDSLIDE DETECTION WITH MORPHOLOGICAL AND FCM CLUSTERING
In the two previous steps, we attempted to separate landslide regions from other bare soil and urban regions. However, there are still many objects that are not of interest, and they should be removed. This step requires geoscience expertise about landslide objects so that they visually can be differentiated from other non-landslide objects. In other words, bare soil and urban objects have relatively large areas, whereas landslides are often small in area, in which these two objects could be distinguishable from each other. Considering this fact, a three-time iteration of a dilating operation with a structural element of 11 × 11 pixels was applied to further remove non-landslide regions from the result. After that, the generated raster was segmented using the mean shift algorithm [56] and the attributes (i.e. colour, standard deviation, compactness, area) of each object were calculated. The mean shift technique is a nonparametric technique originally presented by [57] for the calculation of modes in a multivariate of the likelihood function. It was extended to a mode estimation of the joint spatial and spectral domain [58]. In this technique, the Gaussian kernels are usually chosen in most applications [59]. In this stage, this algorithm was selected due to its popularity and main advantages over other algorithms including well-performing with different remote sensing data including medium to very-high-resolution, multivariate nature, simplicity of the filtering steps, flexibility and availability in different implementations [60], [61]. At the end of the process, each pixel is assigned the estimated spectral signature and spatial location of the local mode of the probability density function it belongs to. Finally, the objects were automatically classified into two classes using the FCM clustering method [62], [63]. The landslide objects were labelled using the visual interpretation of LiDAR imageries.

G. PERFORMANCE EVALUATION
The detection accuracy of the proposed method was measured based on four commonly used statistical measures, namely, Overall Accuracy (OA), Prediction Accuracy (PA), Kappa Index, and Mean Intersection-Over-Union (mIOU).
The OA measures the accuracy of the method by comparing the number of landslides or the number of landslides pixels in the map produced by the proposed method and the number of landslides or the number of landslides pixels in the ground truth data [64]. This measure is a rough accuracy estimate, as it does not account for false detections or spatial agreement. OA was calculated based on (4). (4) where No. The method is the number of landslides or the number of landslide pixels in the result of the proposed method and No. Inventory is the number of landslides or the number of landslide pixels in the inventory data. The PA and Kappa index are the measures that account for incorrect predictions, and they are a form of reliability measure. A value of PA = 1, Kappa index = 1 indicates a perfect performance for the landslide detection method.
The PA and Kappa Index were calculated using the following expression: where the TP (true positive) value is the number of pixels that have been predicted correctly as ''landslide'', the FP (false positive) value is the number of pixels that have been predicted incorrectly as ''landslide'', the TN (true negative) value is the number of pixels that have been predicted correctly as ''non-landslide'', and the FN (false negative) value is the number of pixels that have been predicted incorrectly as ''non-landslide''. Kappa = P p − P exp 1 − P exp (6) where, P p is the proportion of pixels that have been classified correctly as landslide or non-landslide, and P exp means the expected agreements [65]. Furthermore, while none of the methods above measures the spatial agreement between the detection map and the ground truth, this research also used the mean intersection-over-union (mIOU) approach which is widely used in segmentation studies to further assess the performance of the proposed method. The mIOU = 1 indicates a perfect prediction and greater than 0.5 is normally considered a ''good'' prediction [66]. mIOU was computed using (7) [67].
where k + 1 (k = 1 in this research) is the number of classes, i is the label of the ground truth, and the j is the label of the prediction. The p ij is the total number of pixels labelled as j but predicted as i. Fig. 5 shows the result of an initial image saliency enhancement process (Fig. 5a) and after classifying the enhanced image into two classes (Fig. 5b) indicating the initial VOLUME 8, 2020 landslide regions. It can be seen that in orthophotos, after application of saliency enhancement, the contrast of the bare soil and urban regions was enhanced, whereas for vegetated land the contrast was subdued (muted). To separate the potential landslide regions from other regions, the result of image saliency enhancement was classified into two classes with the threshold of 0.54 -1.75% selected by the quantile classification scheme (which is suitable for ordinal data or ranking) utilising Arc GIS software addressed by [34] ( Fig. 5b). This threshold may differ in other areas and data sets; however, it can be automatically selected using common classification algorithms such as quantile, natural breaks, or others [68]. With additional geoscience knowledge about the landslide in the study area, this threshold can be further optimized to reduce the miss-classification between the landslide regions and other bare soil and urban regions. The initial landslide region contained many objects belonging to the bare soil and urban areas from where they needed to be removed.

A. RESULTS OF LANDSLIDE DETECTION
To remove the misclassified bare soil and urban objects as landslides, the slope raster layer was used. As it was referred previously, in the study area of this research objects with a slope of less than 25 • were removed as there is a very low possibility for these to be subject to the landslide event; this was addressed by a recent study for the same area by [47]. The results of this process are shown in Fig. 6. The slope of the study area ranged from 0 to ∼83 • where the highest slopes were found in the northwest part of the area. These areas are mostly forest and cliffs. After removing the regions with a slope of less than 25 • , the result showed that potential landslides were mostly in the south and northwest parts of the study area (Fig. 6). However, the resulting map contained many small objects distributed over the study area and needed further processing, which is discussed in the following section. The values of this slope threshold contributed to separating landslides from other land cover classes such as bare soil. It can be credited to the characteristics of a landslide in the area under consideration. This indicates that the LiDAR-derived data and orthophotos are efficiently operative in revealing the location of landslide.
The dilate operation with a structural element of 11 × 11 pixels as a morphological processing step was employed to refine the landslide detection result. The landslide regions detected after morphological processing are shown in Fig. 7a. Overall, 120358.82 m 2 (0.97% of total area) were detected as landslides (shown in red colour). After that, the generated raster was segmented using the mean shift algorithm [55] and the optimal segmentation parameters (i.e. scale, shape, and compactness) were calculated. The general goal of this segmentation is to partition an image into semantically meaningful regions. This can be accomplished by clustering the pixels in the image using a Gaussian kernel [59] (Fig. 7).
Considering the great diversity of landslides, environmental settings and accessible images, feature selection in high-dimensional datasets is a significant task which was recommended by previous landslide detection researches to target a better performance of data classification [33], [69], [70].
To obtain the optimum features [32], [70], correlationbased feature selection was used to select the most important features for identifying landslide locations with 20 iterations. The total input features contained 82 features derived from LiDAR data (height, slope, and intensity), texture features (GLCM homogeneity and GLCM StdDev), and the  visible band. The highest accuracy was achieved when the features were reduced to 10 features (most important features). Table 3 represents the most important features selected for this study. The values of these features were contributed to separate landslides from other land cover classes.
The result was further refined using the FCM clustering method, where the objects were automatically classified into two classes using spectral and spatial attributes. The result is shown in Fig. 7c. The overall area of landslide regions was 64780.95 m2, which was only 0.52% of the study area. Also, the number of detected landslides was 17. The minimum and maximum areas of the detected landslides were 111 and 11147.29 m 2 , respectively.

B. PERFORMANCE EVALUATION
The comparison of the number of landslides and the number of landslide pixels between the inventory data and the result of the proposed technique is shown in Table 4. The landslide inventory data of the training area had 13 landslide locations. The number of landslides detected by the proposed method was 17. Therefore, the accuracy was calculated as 69.23%.
With a simple but computationally efficient method, an accuracy of 69.23% in terms of detecting the number of landslides seems acceptable, given that no inventory data were used for the detection process compared to traditional and other methods proposed in the literature. Additionally, the number of landslide pixels in the inventory data of the training area was calculated as 85067 for 13 landslide objects, whereas for the detected landslides the number of landslide pixels was 98125. This resulted in an accuracy of 84.64% for the proposed technique using the training area. On the other hand, the test area contained 21 landslide locations and 604911 landslide pixels. The number of landslides and landslide pixels detected in the test area by the proposed method was 37 and 858973, respectively. Thus, the calculated accuracy for the detection of the number of landslides and landslide pixels was 28% and 58%, respectively.
In addition, Table 5 shows the accuracy assessment using a landslide inventory data containing 13 landslide locations of the training area and 21 of the test area and statistical evaluation indices. The performance of the landslide detection technique was evaluated using PA, Kappa index, and mIOU. For the training area, the PA, Kappa index, and mIOU are 71.12%, 0.81, and 68.52% respectively. However, the PA, Kappa index, and mIOU of the method in the test area are 65.78%, 0.68, and 56.14%, respectively. The lower accuracy of the method on the test area is mainly because the terrain (topography and geomorphology) of the Cameron Highlands area is much more complicated. Additional factors that contributed to lowering the accuracy of landslide detection in the test area included the presence of man-made slopes, which are very similar to natural landslides in terms of spectral and spatial characteristics. Considering prior knowledge about these man-made slopes, more undesirable objects can further be removed from the image, accordingly, and results of the landslide detection might be improved.
Results of the landslide detection for the test area are shown in Fig. 8. The use of the proposed technique promises positive progress in the detection accuracy of landslides, VOLUME 8, 2020  especially when inventory data are limited and not available. Moreover, high-resolution LiDAR and visible bands contributed to the simplification in the development of the current research.

V. DISCUSSION
Landslide inventory mapping is an important task in hazard and risk assessments. This study proposed a simple but effective landslide detection technique using the saliency feature enhancement and some other traditional algorithms such as morphological operations, mean shift segmentation, and FCM classification. The main data used are LiDAR-derived slope map and orthophotos and the proposed landslide detection technique could be a useful tool for landslide mapping and monitoring specifically when inventories are scarce as well as there is a need of quick response time.   Fig. 9a and Fig.9b. Also, the salient objects resulting from the saliency enhancement for the area and the final detected landslide by the proposed technique are shown respectively in (Fig. 9c, d). The area of the shown landslide was 295.72 m2 surrounded by a highway and dense vegetation. The low slopes can be observed at the top of the landslide (initial source) and the boundary of the landslide scar. In contrast, Fig. 9c shows the enhanced image using the symmetric surround saliency enhancement technique. It can be seen that a large part of the landslide region has gained the highest brightness values (>200) while the surrounding areas had low values with some high values for the unpaved road in the east part of the area. Furthermore, Fig. 9d shows the detected landslide by the proposed method. It can be observed that some pixels have not been detected as landslides because of their low slope value, as shown in Fig. 9b. The overall accuracy of this landslide detection was 87%. This example shows the efficiency of the proposed techniques for landslide mapping and monitoring. However, to improve the accuracy of landslide boundary delineation, further processing and enhancement of the technique are needed.
Different objects of similar spectral characteristics of landslides would largely weaken the detection rate making it hard to meet the requirements for further landslide analysis. In the proposed approach, as most bare soil and urban zones are not of our interest (background/ unwanted objects), the morphology analysis was employed to further remove and refine these unwanted objects. Using the morphological analysis, the detection process was further improved so that the non-landslide regions were refined (decreased) reaching 0.120 Km 2 , which comprised 0.97% of the total area (Fig 10). To evaluate the advantages of this morphology, we further examined the landslide detection by removing the morphological analysis from the main process. It was noticed that by eliminating this step, the presence of unwanted objects was raised by nearly two-fold. Fig. 10a clearly shows that the nonlandslides objects were increasingly presented in the analysis, indicating a negative impact on the entire process. This is because the morphological operation provides the value of each pixel in the output image based on a comparison of the matching pixel in the input image with its neighbours to enhance the analysis. Furthermore, the combination of dilation and erosion tasks in the same structuring element provides functions to remove the unwanted/ small objects from an image and smooth the boundary of large objects. It can be inferred that involving the morphology step (Fig. 10b) was a successful strategy taken to separate the landslide and nonlandslide regions. This was agreed with suggested studies in [23], [33]. Nevertheless, careful selection of scales and mask dimensions is vital in the design of comprehensive algorithms for all these tasks.
There are numerous features such as spectral, LiDARfactors, spatial and textural that can be used in OBIA methods, however, when the number of features is more than the number of training samples, over-fitting is more likely to happen [71]. Nevertheless, with many landslide inventory locations, OBIA-methods can be more robust for the following reasons. Data from various sources and scales can be effectively integrated at the object level. They capture the complexity of the natural phenomenon and geomorphologic processes such as landslides better than pixel-based methods. Also, results produced by the OBIA methods contain much less ''salt and pepper'' noise than pixel-based methods [63]. In general, instead of using landslide inventory data for model training, general knowledge about the study area is often used.
In the current research, the landslide events could be detected using both geomorphological characteristics (i.e., slope), as well as assisting the expert's advice regarding the study area [43], [47], [54], [55], [72]. Even though knowledge representation in computer programming is not an easy task of modelling, the result of this technique could be more generalizable than data-driven techniques.
Despite the advancement of the spatial resolution of DEM and orthophotos, reliable landslide detection from remote sensing data is still challenging. The proposed method resulted in a promising performance, given that no inventories were used in the training process (unlike the traditional methods). It could statistically achieve average accuracies of 68.45% PA, 0.745 Kappa index, and 62.33% mIOU. Compared with other unsupervised detection methods, the average reliability accuracy obtained in this study scored a Kappa index of 0.745, which was higher than [23]. Scientists in a previous study [23] utilized the Landsat 8 image by employing the saliency and morphological analysis, for which the highest Kappa index of 0.42 for training and 0.25 for tests was obtained. In comparison with another unsupervised detection using K-means clustering via slope feature extraction [73], the accuracy of 54.54 for two clusters, and 85. 47 for five clusters were obtained; however, misclassification in some landslide locations was reported.
Machine learning methods have a core disadvantage that they need the spatial distribution of the test data to be similar to that of the training data [74]. In data science generally, by having more training data (landslide inventories), better accuracy can be achieved. But because the main contribution of the proposed study is exclusively not to use inventory data in the training processes, the result indicates a convincing prediction ratio, unlike most of the existing methods [16]. This can be considered a justifiable advantage over the traditional methods, especially when the disaster's response time is the dominant issue, providing a faster way to map regional hazards. Only a few studies have tried to detect landslides without inventory data, similar accuracies were obtained; however, some uncertainties were reported [17].

VI. CONCLUSION
This study introduced a semi-automated landslide detection technique that does not require landslide inventory data for the training purpose and provides a quick practical solution for landslide mapping and monitoring, particularly for emergency conditions and monitoring. The technique used symmetric surround saliency enhancement, morphological operations, mean shift segmentation, and fuzzy-based clustering methods. The experimental results showed that the proposed technique is simple and effective and its average accuracy can reach 71% for areas with complicated topography and geomorphology. The technique performed well both in the training area (Kinta Valley) and the test area (Cameron Highlands) located in a high landslide-prone region in Malaysia.
Overall, the proposed solution of landslide detection using LiDAR-DEM and orthophotos is promising for landslide inventory mapping and landslide monitoring, especially in areas where data is scarce. Despite the acceptable accuracies VOLUME 8, 2020 achieved for landslide detection by the proposed method, some points need to be improved in future work. First, an optimized combination of RGB and additional bands such as near and shortwave infrared could be used to enhance the geologic features before the saliency enhancement process. Moreover, a combination of landslide factors (e.g slope, altitude, curvature, type of soil) could be utilized as an indicator/threshold to further enhance the background removing process.
HUSAM A. H. AL-NAJJAR received the bachelor's degree in civil engineering and the master's degree in remote sensing and GIS from University Putra Malaysia (UPM), in 2014. He is currently pursuing the Ph.D. degree with the Centre for Advanced Modelling and Geospatial Information Systems (CAMGIS), University of Technology Sydney (UTS). His research interests include the application of deep learning techniques for natural hazard modeling, including landslides, feature extraction from satellite imagery, and UAV-based remote imaging for spatial decision support systems. In 2019, he was awarded the Presidents' Scholarship from UTS. In 2018, he was awarded the International Research Scholarship for the same course. He has several peer-reviewed publications on the application of machine learning methods. He was awarded the Best Paper Award, in 2019, ISPRS Geospatial Week in Enschede, The Netherlands.