Data Fusion Technique Using Wavelet Transform and Taguchi ...

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

1

Data Fusion Technique Using Wavelet Transform and Taguchi Methods for Automatic Landslide Detection From Airborne Laser Scanning Data and QuickBird Satellite Imagery Biswajeet Pradhan, Mustafa Neamah Jebur, Helmi Zulhaidi Mohd Shafri, and Mahyat Shafapour Tehrany

Abstract—Landslide mapping is indispensable for efficient land use management and planning. Landslide inventory maps must be produced for various purposes, such as to record the landslide magnitude in an area and to examine the distribution, types, and forms of slope failures. The use of this information enables the study of landslide susceptibility, hazard, and risk, as well as of the evolution of landscapes affected by landslides. In tropical countries, precipitation during the monsoon season triggers hundreds of landslides in mountainous regions. The preparation of a landslide inventory in such regions is a challenging task because of rapid vegetation growth. Thus, enhancing the proficiency of landslide mapping using remote sensing skills is a vital task. Various techniques have been examined by researchers. This study uses a robust data fusion technique that integrates highresolution airborne laser scanning data (LiDAR) with highresolution QuickBird satellite imagery (2.6-m spatial resolution) to identify landslide locations in Bukit Antarabangsa, Ulu Klang, Malaysia. This idea is applied for the first time to identify landslide locations in an urban environment in tropical areas. A wavelet transform technique was employed to achieve data fusion between LiDAR and QuickBird imagery. An object-oriented classification method was used to differentiate the landslide locations from other land use/covers. The Taguchi technique was employed to optimize the segmentation parameters, whereas the rule-based technique was used for object-based classification. In addition, to assess the impact of fusion in classification and landslide analysis, the rule-based classification method was also applied on original QuickBird data which have not been fused. Landslide locations were detected, and the confusion matrix was used to examine the proficiency and reliability of the results. The achieved overall accuracy and kappa coefficient were 90.06% and 0.84, respectively, for fused data. Moreover, the acquired producer and user accuracies for landslide class were 95.86% and 95.32%, respectively. Results of the accuracy assessment for QuickBird data before fusion showed 65.65% and 0.59 for overall accuracy and kappa coefficient, respectively. It revealed that fusion made a significant improvement in classification results. The direction of mass

Manuscript received September 18, 2014; revised May 8, 2015; accepted September 19, 2015. This work was supported by Putra Research Grant (GPI/2014/9439200) to stimulate research through the Research University Grant Scheme (RUGS) under Project 9439200. The authors are with the Department of Civil Engineering, Geospatial Information Science Research Center (GISRC), Faculty of Engineering, Universiti Putra Malaysia, 43400 Serdang, Malaysia (e-mail. [email protected]; [email protected]; [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TGRS.2015.2484325

movement was recognized by overlaying the final landslide classification map with LiDAR-derived slope and aspect factors. Results from the tested site in a hilly area showed that the proposed method is easy to implement, accurate, and appropriate for landslide mapping in a tropical country, such as Malaysia. Index Terms—Fusion, landslide, laser scanning data (LiDAR), Malaysia, remote sensing, rule-based, Taguchi, wavelet transform.

I. I NTRODUCTION

L

ANDSLIDE locations should be detected to produce landslide inventories. Such inventories are used for various purposes [1], [2], such as recording the landslide magnitude in a region; implementing the initial stage for landslide susceptibility, hazard, and risk analysis; examining the distribution, kinds, and shapes of slope failures; and studying the evolution of landscapes affected by landslides [3]–[7]. Producing a landslide inventory map is considered a challenging task because of rapid vegetation growth in tropical regions. Vegetation can cover the landslide location, such that this location cannot be detected by most available recognition techniques. Thus, a rapid and accurate technique is required. Given the importance of having an accurate landslide inventory map, the principles and standards for the construction of these maps and their assessment should be clearly defined [8], [9]. Accessibility of new remote sensing tools for the recognition and mapping of slope failures can aid in the establishment of landslide inventory maps [10], [11]. Some general assumptions should be made to detect landslide locations and produce landslide inventories. Landslides provide visible marks, such as variations in the form, location, or appearance of the topography, which can be identified, categorized, and mapped through the interpretation of aerial photographs and satellite images [8]. Furthermore, the morphological signature differs on the basis of landslide type [12], [13]. Another assumption is that landslides do not occur by accident or by chance [14]. Landslides are the consequences of the interaction of the physical procedures and mechanical regulations governing the constancy or failure of a slope [12]. The final assumption is that landslides are expected to occur under circumstances similar to those that led to past slope failures. Detection and mapping of slope failures are conducted on the basis of these assumptions [15]. Two major aspects should be considered to produce an efficient landslide inventory map. One

0196-2892 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 2

aspect is related to the data used, whereas the other is related to the method employed. This research considered both aspects in providing a reliable and accurate landslide inventory map. Traditional techniques for the construction of landslide inventory maps are mostly based on the visual interpretation of aerial photographs. Such interpretation requires multiple field surveys. Some disadvantages are that these methods are expensive and time consuming. Traditional techniques include geomorphological field mapping and visual interpretation of stereoscopic aerial photographs [16]. Some researchers have assumed that this type of mapping is more precise than remote landslide mapping [17]. However, geomorphological field mapping has some limitations. For instance, the extent of a landslide is usually too large to be completely observed in an area. Moreover, researchers have a limited perspective in recognizing all the aspects of a landslide in detail. In some cases, old landslides are covered by vegetation or have been changed by other slope failures and human activities [18]. Thus, the use of this method enables the investigator to recognize and map single landslides or small groups of landslides [15]. One of the most popular techniques for landslide detection is still being used despite considerable improvements in the technology for interpreting aerial photographs [19], [20]. With the use of the vertical exaggeration provided by a stereoscope, the morphological form of land is amplified, such that changes can be detected. The use of this technique does not require the researcher to possess sophisticated technical abilities [21]. An aerial photograph with a large size and fine scale can cover the entire landslide location in one scene [22]. The accessibility of several sets of aerial photographs for a similar region provides a researcher with the opportunity to conduct a temporal evaluation of landslides [19]. The issue is that detecting slope failures with the use of aerial photography is an uncertain technique that requires skills, training, an organized methodology, and proper interpretation principles [23]. No standard procedure is available, and the interpreter identifies and categorizes landslide types on the basis of knowledge and the investigation of a set of features that can be recognized on the images. Moreover, vegetation height and thickness, as well as their changes, influence the recognition of slope failure when using aerial photographs [24]. Several emerging methods have recently been used on the basis of satellite, airborne, and remote sensing technologies to aid and simplify landslide detection and mapping procedures. These technologies require less resources and time for detecting and recording information. Guzzetti et al. [12] compared traditional and advanced techniques for landslide recognition. They stated that the new techniques can enhance the precision of landslide maps, with positive influences on all derived outcomes and investigations, such as landslide modeling, susceptibility evaluation, and risk assessments. Subsequently, numerous methods have been proposed and examined by various researchers [25], [26]. The advantages and limitations of the new remote sensing data and technologies have been likewise discussed. One of the common methods is analyzing the surface morphology with the use of very high resolution digital elevation models (DEMs). The accessibility of very high-resolution DEMs attained by high-resolution airborne laser scanning data (LiDAR) sensors provide opportunities for researchers to iden-

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

tify and map slope failures [27]. A significant advantage of LiDAR lies in its capability to penetrate vegetation areas and acquire valuable information on topographic conditions [28]. This capability makes LiDAR data distinct from data acquired from other sources, such as aerial photographs, in terms of the detection of slope failure in forested regions [29]. Based on the literature, the derived DEM from LiDAR sensors is mostly used for the visual assessment of topographic surfaces [30] and the semiautomatic identification of landslide characteristics. Ardizzone et al. [31] reported on the improvement in detected landslide locations when using a LiDAR-derived DEM over the interpretation of aerial photographs. Mann et al. [32] developed a method to automate the process of classifying rotational landslides from a high-resolution LiDAR-based DEM. Furthermore, the characteristics of landslides were investigated and described. The limitations of using LiDAR data are related to the cost and time required to acquire and examine data. Terrestrial laser scanning (TLS) data were also used for landslide detection and inventory mapping [33], [34]. TLS is one of the most promising remote sensing techniques for landslide characterization and mapping because of its capability to acquire the precise dense 3-D coordinates of the terrain [35]. The TLS approach shows some bias in the recording of landslide phenomena. Thus, the method should be combined with other approaches [33]. Another popular technique used to recognize and map landslides is satellite imagery interpretation and examination [36]. The variations that occur in the spectral signature of ground objects can be recognized by satellite sensors. Such variations can be used to detect slope failure. The identification of landslides on the basis of satellite imagery was initiated in the 1970s using the Satellite Pour l’Observation de la Terre (SPOT) and Landsat optical images [12]. The use of satellite data has increased because of the accessibility of high-resolution and very highresolution sensors, the availability of active and passive data, and the advances in computer technology for the analysis, representation, and examination of satellite images [25], [37]. The literature reveals three main attempts in the field of interpretation and examination of satellite imagery to recognize slope failure. These attempts are visual interpretation of optical images [19], analysis of multispectral images using image classification methods [38], and analysis of synthetic aperture radar (SAR) images [39]. Optical satellite images are an alternative to traditional aerial photographs and can be utilized to provide projected and orthorectified images. This efficiency of this approach has been proven in [19] and [40]. The use of multispectral imagery to detect landslides has an advantage over aerial photos, panchromatic images, and LiDAR data. Such advantage lies in the capability to record multispectral information of the terrain in specific portions. This capability facilitates landslide recognition and enables the implementation of various classifications, such as index thresholding [41], supervised and unsupervised classifications [42], change detection methods [43], and object-oriented analysis [44]. Another attempt made in the domain of satellite technology is the analysis of SAR data [39], [45]–[47]. Terrain deformations and temporal changes can be modeled by using SAR sensors [48]. This idea was first implemented in [45] using

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. PRADHAN et al.: DATA FUSION USING WAVELET TRANSFORM AND TAGUCHI METHODS FOR DETECTION

3

Fig. 1. Hill-shaded map of the study area and landslide locations.

Radarsat images and airborne SAR data to recognize large landslides in Canada. With the use of differential interferometry SAR (DInSAR) and interferometry SAR (InSAR) techniques, small movements of the terrain can be detected with centimeter to millimeter accuracy. Jebur et al. [46] identified the landslide that occurred in the Gunung Pass area, Malaysia, by using InSAR created from Advanced Land Observing Satellite Phased Array-type L-band SAR repeat pass data. In another study conducted in [47], the slope failure in Cameron Highlands, Malaysia, was mapped by using the DInSAR technique. The problem with these techniques is related to their complex procedure and the need for specific software. Moreover, two or more precise coregistered SAR data are needed, which is expensive and increases the workload. Most researchers found this technique difficult to understand. Furthermore, interpretation of the acquired final map requires adequate experience in SAR data analysis.

As revealed by the literature review, a set of criteria and a detailed analysis are required to construct accurate landslide inventory maps. The accessibility of high-spatial-resolution satellite images and improvements in image processing technologies provided the opportunity to conduct a more sophisticated analysis through the fusion of data from multiple sources. The novelty of the current research is that very high-resolution satellite images were fused with LiDAR-derived DEMs to gain a better view of the topography, which can be classified or visually interpreted to recognize landslides. Data fusion enhances urban surface features. Therefore, this method is valuable for many remote sensing applications. Most satellites, such as SPOT, IKONOS, QuickBird, and OrbView, produce multispectral images with low and high spatial resolutions. QuickBird satellite is rapidly becoming one of the most popular choices for high-resolution mapping with the use of satellite images. Thus, in the current research, we conducted data fusion

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 4

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

Fig. 2. Stepwise scheme for rule-based detection of landslide using object-oriented classification by segment optimization.

for QuickBird imagery with a LiDAR-derived DEM to map the landslide locations in Bukit Antarabangsa, Ulu Klang, Malaysia. A wavelet transform method was used for data fusion. Moreover, advanced methods, such as Taguchi and rulebased classification, are used instead of traditional classification techniques for segmenting and categorizing the fused data. II. S TUDY A REA Bukit Antarabangsa, Ulu Klang, Malaysia was selected as the study area for the analysis. This area is located at the zone of 3◦ 11 N to 3◦ 12 N latitude and 101◦45 E to 101◦ 46 E longitude. Area coverage is 53 km2 , covering a small part of Selangor state (see Fig. 1). From April to June, the highest temperature in the study area is between 29 ◦ C and 32 ◦ C. The average rainfall is between 58 and 240 mm/month. The frequent occurrence of slope failure in this area is attributed to the heavy monsoon precipitation and unplanned development work. A destructive landslide occurred in Bukit Antarabangsa, in an upper middle-class neighborhood in Ulu Klang, Selangor, on December 6, 2008. Many families were forced to evacuate their houses. Therefore, this region was selected as a pilot study because of frequent landslide occurrences. This study focused on the landslides which occurred in 2006. The predominant land use types in the study area are vegetation and urban areas [49]. The soil has a weak structure, and the main soil types are loam and clay [50]. III. DATA U SED Two data sources, namely, LiDAR-derived DEM and QuickBird imagery, were utilized in conducting the data fusion techniques. Airborne LiDAR is a popular remote sensing method used for the digital presentation of the topographic

surface of regions with small to large coverage [25]. This method uses a laser sensor, which is positioned on an airplane, to record the distance from the device to various points on the earth. In each square, 100 points can be recorded on the basis of some conditions, such as elevation, speed, and type of sensor [51]. Moreover, the condition of the terrain is an important factor. In this research, LiDAR vector point data were gathered over ∼3.46 km2 of Bukit Antarabangsa on August 3, 2007. The data were recorded with approximately 14 732 461 data points. As previously mentioned, QuickBird imagery using high-resolution satellites has become a popular choice for mapping. The satellite contains panchromatic (61–72-cm spatial resolution) and multispectral sensors (2.44–2.88-m spatial resolution) and was captured on February 8, 2007. The sensor covers 16.5–19 km in the across-track direction. The significant advantage of these data over conventional aerial photography is the capability to provide multitemporal landslide maps with a revisiting rate of one day [52]. IV. M ETHODOLOGY The current research methodology is constructed using various steps and is shown in Fig. 2. QuickBird imagery with 2.6-m spatial resolution is fused with the LiDAR-derived DEM. Such fusion facilitates the recognition of landslide events in the study area. Data fusion enhances urban surface features and provides a better view of the topography. Here is the brief description of the methodology employed in the current research. First, a DEM was extracted from the LAS (LASer) data with 1-m spatial resolution. Second, a wavelet transform algorithm was utilized to conduct data fusion for QuickBird and the DEM. Subsequently, the result of the fusion was moved to the segmentation stage. A multiresolution segmentation method was used to implement segmentation.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. PRADHAN et al.: DATA FUSION USING WAVELET TRANSFORM AND TAGUCHI METHODS FOR DETECTION

Numerous features of the objects can be used in multiresolution segmentation. In this paper, three parameters, namely, size, compactness, and shape, were chosen in the analysis. Various combinations of these segmentation parameters can be chosen, and the selection of each significantly affects the outcome of the analysis [53]. The Taguchi method was used to select the best combination of these parameters [54]. Most available optimization techniques are only capable of optimizing the scale but not the combination of parameters [54], [55]. The Taguchi method is an efficient optimization technique that can overcome this limitation and recognize the optimum combination of segmentation parameters [56]. The optimum combination of parameters was specified and examined by using the plateau objective function (POF) [53]. The output of this stage was used to construct the segmented images. Subsequently, the rule-based classification technique was implemented to set the rules and differentiate between the classes. The slope factor was derived from LiDAR data for generating rules to increase the precision of classification. Three categories were defined: urban, vegetation, and landslide. A classification map was prepared. A confusion matrix was used to examine the efficiency and reliability of the classification map. With the use of the confusion matrix, four accuracies were measured, namely, overall accuracy, user accuracy, producer accuracy, and kappa coefficient. In the next step, the result was exported to be overlaid with the slope and aspect factors, which were derived from LiDAR imagery, to detect the direction of the landslide. The details for each step are explained in the next section. In order to represent the impact of fusion in analysis, a rule-based classification method was also applied on original QuickBird data which have not been fused. A. Data Fusion Image fusion is a technique that compounds information from various images captured in the same region. These images can be recorded by using different devices, at different periods, or with different spatial and spectral features. The fusion outcome is a new image that preserves the desired information and features of each input image [57]. With the accessibility to multisensory data in various domains, image fusion became popular for an extensive range of applications. Image fusion can be implemented by using numerous techniques [51]. The most popular image fusion techniques are intensity–hue–saturation (IHS) transform-based fusion [58], principal component analysis (PCA)-based fusion [59], and Brovey transform [60]. IHS, PCA, and Brovey transform are categorized as standard image fusion techniques applied under the spatial domain. The drawback of these fusion techniques is that, in some cases, they create spectral degradation [61]. This condition is more evident when images were captured at different times. Thus, outcomes are poorer than the expected results. The wavelet transform method, which is a type of multiresolution analysis, can overcome the drawback of standard techniques [62]. The efficiency and applicability of the wavelet method in numerous fields have been examined in [63]. Previous research revealed that wavelet transform can provide the fused image with high spatial and spectral quality, partic-

5

ularly in terms of minimizing color distortion. This method was used for the first time in the 1980s for the purpose of signal processing. Within the last few years, the efficiency of wavelet transform in image processing has been improved [62]. Determining the important theory behind wavelet-based fusion can be difficult for readers because of different and extensive mathematical explanations. The details for waveletbased fusion calculation can be seen in the study in [62]. In this study, wavelet transform was applied for the fusion of LiDARderived DEM and QuickBird imagery using the MATLAB software. B. Object-Oriented Classification Object-oriented methods have two main steps, namely, segmentation and classification [64]. In segmentation, the image is partitioned into homogeneous regions (objects). This step reduces the inherent noise in pixel-based analysis and enables multiscale analysis [65]. The rule-based and nearest neighbor techniques are implemented in object-oriented classification [66], [67]. The nearest neighbor classifier works on the basis of user-defined samples. However, rule-based classification provides the opportunity for the user to define specific rules for each object and to have control over the entire classification process. Thus, rule-based classification was utilized in this research. Rule-based classification considers the spectral, textural, and contextual characteristics of each segment. Segmentation quality significantly influences classification. Thus, pixel-based approaches first classify pixels, considering them as independent areas to describe object restrictions, whereas the object-oriented technique works more accurately by first determining and then categorizing objects [68], [69]. In this research, eCognition software was used to implement object-oriented classification. Three categories, namely, urban, vegetation, and landslide, were visually detected and classified in this study. 1) Segmentation: As previously mentioned, segmentation is a technique that divides the image into nonoverlapping areas or segments. This technique is an initial and important step in object-oriented classification. The precision and quality of segmentation directly affect the accuracy of the generated classification map [70]. Two main groups of segmentation algorithms are available: edge and region based [71]. Edgebased algorithms recognize the image edges by thresholding the image gradient gained from a differentiation filter. By contrast, region-based algorithms start by dividing the entire image and then merging the homogeneous areas [71]. In this study, a multiresolution algorithm was utilized for segmentation. This algorithm can be classified as region based. This algorithm follows various steps, is initiated with one pixel, and continues until it covers all the criteria that were specified by the user [72]. Various parameters are used in multiresolution segmentation, namely, size, compactness, and shape. The Taguchi method was utilized to determine the optimum combination of the previously mentioned parameters, which is discussed in the following section (i.e., Section IV-B2). 2) Taguchi-Based Optimization of Segmentation Parameters: Determination of the optimum combination of the segmentation

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 6

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

parameters by trial and error is time consuming and requires extensive work. Thus, optimization techniques can be a suitable solution to facilitate and reduce the time needed for the selection of parameters. The influence of various parameters on the performance attribute in a decreased set of choices can be tested by using the orthogonal array experimental design of Taguchi. An orthogonal array is related to the case where the columns for the independent variables are “orthogonal” to one another. These tables provide an easy and consistent design of experiments. This technique is often used in cases in which various levels of parameters are available. For instance, in this research, the three available parameters can produce 243 combinations, which all require a significant amount of time when tested. However, with the use of the Taguchi technique, the choices will be considerably reduced to 25 experiments. To apply the general steps in the Taguchi method, the following steps were performed. a) The process objective was determined in the beginning. This step entails defining the possible values of a particular parameter for the process. b) The parameters that can influence the process were then defined. These parameters exhibit variable values that can affect performance; thus, the level should be defined by the user, depending on the parameter’s effect on the process. For instance, a scale value can vary from 0.1 to 1.0. When the level increases, the number of experiments to be conducted will be increased as well. c) An orthogonal array was created to design the condition and determine the number of experiments. Selecting the orthogonal array depends on the number of levels and the number of parameters. d) The experiments were then applied after the appropriate array had been selected. The effect of each parameter on the performance was then measured. The loss function can be calculated as follows [56]: l(y) = kc (y − T )2

(1)

where τ is the difference between the target value of the performance characteristic of a process, and the measured value y as a loss function. kc is the constant in the loss function that can be calculated by considering the acceptable interval as follows: kc =

C . Δ2

(2)

When whole parameters that affect the process are defined, the level of each parameter should also be defined. The level refers to the probable value of each parameter, in terms of maximum, minimum, and current value. In the case of a big gap between the minimum and the maximum value of a specific parameter, additional levels are also added to that parameter. The proper array was selected after the number of parameters and levels have been defined. A constant array was found for the Taguchi method. Each array can be selected depending on the parameters and levels. Table VI (shown later) was created using an algorithm of Taguchi. In the case of three parameters

and two levels, the L9 array was selected. The array assumes that the number of levels is equal for each parameter. Otherwise, the assumption will be based on the highest value. In the next step, the POF was measured for each test to examine the quality of segmentation, by using each testing combination, and to determine the optimum segmentation parameters. POF is the combination of a spatial autocorrelation index and a variance indicator. In the next step, signal-to-noise (S/N ) ratio was measured to assess the testing segmentation outcomes. A higher S/N ratio represents higher segmentation accuracy. Equation (3) was used to calculate the S/N ratio, as follows:    1 1 (3) SNR = −10 log10 n yi2 where n is the number of repetitions under similar test situations (n = 1 in this study), and y denotes the POF values obtained from each segmentation test. The S/N ratio table was then achieved, and the optimum condition was determined. 3) Defining the Rules for Classification: Once objects were detected and segmented by using the Taguchi method, eCognition calculates various parameters, such as size, compactness, and shape of each object. These parameters will be used as class discriminators in the object-oriented classification. These attributes of the objects are called object features. Setting the rules is the final step in classifying the fused image. Table I lists some statistical information for each band of the image before and after fusion. This information is also used in the creation of rules, as shown in Table I. Table II shows the created rules and their category for data before and after fusion. For instance, for data after fusion, the normalized difference vegetation index was used as an important index to differentiate between forest and agricultural crop lands from other classes. During the creation of the rules for QuickBird data before fusion, no any additional information, such as DEM and slope, were used. Various rules were examined, and the best result was achieved by using the final rules shown in Table II. With the use of the created rules, the fused image was classified, and homogeneous groups were recognized. C. Accuracy Assessment The quality of a landslide inventory map depends on its accuracy and on the type and certainty of the information represented in the map. The accuracy of the generated map can be described on the basis of the completeness of the map and the thematic precision of the information illustrated on the map [15]. The confusion matrix was used to examine the efficiency of the classification map and the detected landslide locations. The confusion matrix is a cross-tabulation of the classified and actual class labels for the study area [73]. This matrix is a square array of dimension r × r, where r is the number of categories. The confusion matrix represents the correlation among two samples of measurements taken from the classified region. With the use of the confusion matrix, the overall accuracy, user accuracy, producer accuracy, and kappa coefficient can be

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. PRADHAN et al.: DATA FUSION USING WAVELET TRANSFORM AND TAGUCHI METHODS FOR DETECTION

7

TABLE I S TATISTICAL I NFORMATION FOR E ACH BAND OF THE I MAGE B EFORE AND A FTER F USION

TABLE II C REATED RULES FOR I DENTIFICATION OF THE L ANDSLIDE L OCATIONS

measured. The overall accuracy is determined by dividing the aggregate of the main diagonal entries of the confusion matrix by the entire number of samples. The kappa coefficient was measured by using ˆ = θ1 − θ2 K 1 − θ2

(4)

where θ1 is the ratio of correctly classified areas, and θ2 is the proportion of agreement that is expected by chance. V. R ESULTS AND D ISCUSSION A. Fusion Result Fusion was implemented using LiDAR-derived DEM and QuickBird imagery. As shown in Fig. 3, objects can be easily identified in the fused image, which has the information of both input data points. This characteristic facilitates class discrimination and classification. B. Optimization Outcomes and Segmentation Prior to the implementation of the Taguchi method, the choices for the three parameters, namely, size, compactness, and shape, were defined and listed in five levels, as shown in Table III. With the use of these levels, 243 combinations can be created for segmentation, and the analysis of these combinations requires a significant amount of time. However, the use of the Taguchi optimization method reduced the number of combinations to 25 experiments. As shown in Table IV, with the use of the Taguchi method, 25 combinations of the parameters for segmentation were selected for further analysis. Therefore, all 25 segmentation

prototypes were examined according to the Taguchi orthogonal array. Furthermore, the measured POF and S/N ratio for each combination are shown in the last two columns of Table IV. The highest POF and S/N ratio showed the optimum combination for segmentation. As shown in Table IV, the optimum conditions for segmentation are as follows: 1) size of 80; 2) compactness parameter equal to 0.9; and 3) shape parameter of 0.5 (experiment number = 20). Evidently, a phenomenon such as a landslide has various sizes. Therefore, multiresolution segmentation is the appropriate method to segment these types of features. The use of the multiresolution segmentation technique enables the selection of various choices of parameters. Thus, instead of choosing one scale factor with the highest POF, two other scales with the highest POF values within the examined segmentations were also selected. The optimum combination of parameters was selected, and the entire image was segmented through the multiresolution technique. Fig. 4 shows the resultant segmentation map. In the segmented image (see Fig. 4), we observed that the boundary of the objects was detected more accurately. A few samples were selected to represent the effect of various segmentation parameters. These samples are shown in Fig. 5. We observed that the use of different segmentation properties changed the defined boundaries. Visually, the segmentation with a scale of 80 can detect the exact boundary of the landslide. Such boundary is denoted in blue [see Fig. 5(d)]. However, other segmentations included vegetation in the landslide class. C. Classification and Accuracy Assessment Outcomes The defined rules were used to classify the entire images (before and after fusion), after which the classification maps were produced. Fig. 6 shows the classification maps with the

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 8

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

TABLE IV L25 O RTHOGONAL A RRAY, POF, AND S/N R ATIO FOR S EGMENTATION P ROCESS

Fig. 3. New image constructed by the fusion of LiDAR-derived DEM and QuickBird imagery. TABLE III D EFINED L EVELS FOR S EGMENTATION PARAMETERS

landslide inventory maps before and after fusion. By overlaying the LiDAR-derived slope and aspect factors, the direction of the landslides was identified. The map in Fig. 6(b) shows the location of the landslides with their direction. From the map, one can observe that landslides mostly occurred in the nonvegetated areas. This finding indicates the importance and positive effect of forests and vegetation on landslide mitigation. Land use changes and deforestation, as well as heavy rainfalls, are the landslide-triggering factors in the study area. Surprisingly, there are several misclassifications, which are shown in Fig. 6(c). Most of the urban area is classified as landslide, which shows exaggeration in the results. A 3-D view of the landslides was also used to better represent the location of the landslides. Fig. 7 shows the landslides and their direction. Table V shows the results of the confusion matrix for objectoriented results. For the fused data, the achieved overall accuracy and kappa coefficient were 90.06% and 0.84, respectively. The landslide class had 95.86% and 95.32% producer and user accuracies, respectively. This result reveals that the locations that have been detected as landslide areas from fused data are significantly precise.

Fig. 4. Segmented image.

Based on the accuracy assessment outcomes, the achieved classification result can be considered a reliable and accurate source for further analysis. As previously explained, the

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. PRADHAN et al.: DATA FUSION USING WAVELET TRANSFORM AND TAGUCHI METHODS FOR DETECTION

9

Fig. 5. Spatial structure of various segments for one landslide. (a) Segmentation with a scale size of 20, (b) segmentation with a scale size of 40, (c) segmentation with a scale size of 60, (d) segmentation with a scale size of 80, and (e) segmentation with a scale size of 100.

Fig. 6. Rule-based classification results. (a) The classified map after fusion, (b) the landslide inventory map, and (c) the classified map before fusion.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 10

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

Fig. 7. Landslide locations in 3-D perspective views. TABLE V E RROR M ATRIX R ESULTS FOR THE C LASSIFICATION P ERFORMED W ITH F USED I MAGE Fig. 8. Spectral diagram of the landslides (a) before and (b) after the data fusion.

inventory map is the basis for susceptibility, hazard, and risk analysis. Therefore, the accuracy of the inventory map directly affects future results. However, for QuickBird data before fusion, the achieved overall accuracy and kappa coefficient were 65.65% and 0.59, respectively. The reason is that DEM and slope were not involved in creation of the rules and fusion has not been done for the data at this stage. Hence, many misclassifications can be seen in the results. Compared to the fused classified data, most of the urban area is classified as landslides, which is obviously not proper to be used as an inventory map for further analysis. Object-oriented classification considers the geometric properties of the objects in addition to their spectral properties. Therefore, such approach effectively discriminated between objects and produced reliable maps. The use of objects instead of pixels helps reduce the spectral variance in each category, particularly in high-resolution imagery. Pixel-based classification methods have some weak points that the object-based classification method can overcome. For instance, spectral heterogeneity in each category and spectral similarities between different objects can be observed in pixel-based methods. However, object-based methods, which use contextual information such as geometric features associated with objects, along with spectral characteristics and auxiliary data, could address the aforementioned drawback. Another reason for the acquisition

Fig. 9. Field photographs of some of the landslides in the study area.

of precise outcomes is related to the use of data fusion in landslide detection. Data fusion enhances the urban surface features and produces a better view of the topography. Spectral diagrams of all landslides in the study area were plotted before and after applying fusion to show the effect of fusion on the spectral characteristic of the features, particularly those of landslides. In Fig. 8(a), each landslide has its own direction and format. However, all landslides produced a spectral diagram in the same direction and format after fusion [see Fig. 8(b)]. Thus, fusion can considerably facilitate the segmentation and generation of the rules for classification. The validation results and the acquired maps reveal that the proposed method can be considered a reliable technique for the recognition and mapping of landslide locations.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. PRADHAN et al.: DATA FUSION USING WAVELET TRANSFORM AND TAGUCHI METHODS FOR DETECTION

D. Field Verification In addition to accuracy assessment, field verification was conducted to examine the correctness of the landslide inventory map and to evaluate the efficiency of the proposed method. The location of the landslides was detected by using the GeoExplorer 6000 handheld Global Positioning System (GPS) (see Fig. 9). With the use of the differential method, the accuracy of the collected points was scaled to centimeter accuracy in the real-time method. The information obtained from in situ measurements proved the reliability of the produced inventory map. The detected landslides from the fused image are actual landslides, which were validated in the field (see Fig. 9).

VI. C ONCLUSION This research aimed to produce an accurate landslide inventory map with the use of efficient techniques of data fusion, rule-based object-oriented image classification, and Taguchi optimization. The combination of the proposed methods and the methodology used were applied for the first time to detect landslide locations in a tropical urban area. Data fusion was conducted using LiDAR-derived DEM and very high resolution QuickBird imagery. Data fusion enhanced the visual appearance of the features and created a better view of the topography. Therefore, data fusion facilitated and enhanced rule generation and classification performance. Although object-oriented classifications require more time for processing than pixelbased methods, object-oriented classifications are capable of overcoming the drawbacks of the pixel-based methods. The object-based method considers the spatial, spectral, and textural characteristics of the features in the process of classification, which distinguishes it from other classification methods. Rulebased object-oriented classification was selected to discriminate between classes. Segmentation, which is the first step in object-oriented classification, was implemented by using the multiresolution technique. Optimization of the segmentation parameters and creation of the rules for each feature type are the main challenges in the generation of accurate object-based classification for rapid landslide recognition. The Taguchi optimization technique was applied to determine the optimum combination of the segmentation parameters. The Taguchi technique has been proven to be an efficient method for such applications. Orthogonal arrays, which were produced by the Taguchi technique, facilitated the recognition of the optimum combination for segmentation by implementing a limited number of examinations. Consequently, the appropriate rules were defined and applied to the entire fused image. The classification map was produced, and validation was conducted using the confusion matrix method. Moreover, in order to represent the impact of fusion in analysis, the rule-based classification method was also applied on original QuickBird data which have not been fused. The high overall accuracy of the classification map showed that it can be used as a valid inventory map, which is highly valuable in planning and disaster management policies in urban areas. Moreover, field verification was also applied to ensure the reliability of the method employed and of the inventory map produced

11

through the use of a GeoExplorer 6000 handheld GPS. The results showed that the detected landslides from the fused image are actual landslides that happened in the past. Therefore, the achievements of the current research confirm that the proposed methodology can detect the location of landslides and produce a reliable inventory map. ACKNOWLEDGMENT The authors would like to thank the National Mapping Agency (JUPEM) and the Department of Mineral and Geosciences (JMG), Malaysia, for providing the various data sets used in this paper, and two anonymous reviewers and Editorial comments by Prof. A. Plaza, which helped us to revise the manuscript. R EFERENCES [1] F. Guzzetti et al., “Distribution of landslides in the Upper Tiber River basin, central Italy,” Geomorphology, vol. 96, no. 1/2, pp. 105–122, Apr. 2008. [2] B. Pradhan, M. H. Abokharima, M. N. Jebur, and M. S. Tehrany, “Land subsidence susceptibility mapping at Kinta Valley (Malaysia) using the evidential belief function model in GIS,” Nat. Hazards, vol. 73, no. 2, pp. 1019–1042, Sep. 2014. [3] R. N. Parker et al., “Mass wasting triggered by the 2008 Wenchuan earthquake is greater than orogenic growth,” Nat. Geosci., vol. 4, no. 7, pp. 449–452, Jul. 2011. [4] Z. Umar, B. Pradhan, A. Ahmad, M. N. Jebur, and M. S. Tehrany, “Earthquake induced landslide susceptibility mapping using an integrated ensemble frequency ratio and logistic regression models in West Sumatera Province, Indonesia,” Catena, vol. 118, pp. 124–135, Jul. 2014. [5] B. Pradhan, S. Mansor, S. Pirasteh, and M. F. Buchroithner, “Landslide hazard and risk analyses at a landslide prone catchment area using statistical based geospatial model,” Int. J. Remote Sens., vol. 32, no. 14, pp. 4075–4087, Jul. 2011. [6] R. W. Fleming, D. J. Varnes, and R. L. Schuster, “Landslide hazards and their reduction,” J. Amer. Plan. Assoc., vol. 45, no. 4, pp. 428–439, Oct. 1979. [7] F. Guzzetti, A. Carrara, M. Cardinali, and P. Reichenbach, “Landslide hazard evaluation: A review of current techniques and their application in a multi-scale study, Central Italy,” Geomorphology, vol. 31, no. 1–4, pp. 181–216, Dec. 1999. [8] F. Guzzetti, M. Cardinali, P. Reichenbach, and A. Carrara, “Comparing landslide maps: A case study in the upper Tiber River Basin, central Italy,” Environ. Manage., vol. 25, no. 3, pp. 247–263, Mar. 2000. [9] C. J. van Westen, E. Castellanos, and S. L. Kuriakose, “Spatial data for landslide susceptibility, hazard, and vulnerability assessment: An overview,” Eng. Geol., vol. 102, no. 3/4, pp. 112–131, Dec. 2008. [10] F. Guzzetti et al., “Landslide hazard assessment in the Collazzone area, Umbria, Central Italy,” Nat. Hazards Earth Syst. Sci., vol. 6, no. 1, pp. 115–131, Jan. 2006. [11] S. Siyahghalati, A. K. Saraf, B. Pradhan, M. N. Jebur, and M. S. Tehrany, “Rule-based semi-automated approach for the detection of landslides induced by 18 September 2011 Sikkim, Himalaya, earthquake using IRS LISS3 satellite images,” Geomatics, Nat. Hazards Risk, to be published. [12] F. Guzzetti et al., “Landslide inventory maps: New tools for an old problem,” Earth-Sci. Rev., vol. 112, no. 1/2, pp. 42–66, Apr. 2012. [13] D. Tien Bui, B. Pradhan, O. Lofman, I. Revhaug, and O. B. Dick, “Landslide susceptibility mapping at Hoa Binh province (Vietnam) using an adaptive neuro-fuzzy inference system and GIS,” Comput. Geosci., vol. 45, pp. 199–211, Aug. 2012. [14] D. L. Turcotte, B. D. Malamud, F. Guzzetti, and P. Reichenbach, “Selforganization, the cascade model, and natural hazards,” Proc. Nat. Acad. Sci. USA, vol. 99, no. S1, pp. 2530–2537, Feb. 2002. [15] M. Galli, F. Ardizzone, M. Cardinali, F. Guzzetti, and P. Reichenbach, “Comparing landslide inventory maps,” Geomorphology, vol. 94, no. 3/4, pp. 268–289, Feb. 2008. [16] F. Brardinoni, O. Slaymaker, and M. A. Hassan, “Landslide inventory in a rugged forested watershed: A comparison between air-photo and field survey data,” Geomorphology, vol. 54, no. 3/4, pp. 179–196, Sep. 2003.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 12

[17] M. Santangelo et al., “Remote landslide mapping using a laser rangefinder binocular and GPS,” Nat. Hazards Earth Syst. Sci., vol. 10, no. 12, pp. 2539–2546, Dec. 2010. [18] P. E. Miller et al., “A remote sensing approach for landslide hazard assessment on engineered slopes,” IEEE Trans. Geosci. Remote Sens., vol. 50, no. 4, pp. 1048–1056, Apr. 2012. [19] F. Fiorucci et al., “Seasonal landslide mapping and estimation of landslide mobilization rates using aerial and satellite images,” Geomorphology, vol. 129, no. 1/2, pp. 59–70, Jun. 2011. [20] S. Lee, K. Y. Song, H. J. Oh, and J. Choi, “Detection of landslides using web-based aerial photographs and landslide susceptibility mapping using geospatial analysis,” Int. J. Remote Sens., vol. 33, no. 16, pp. 4937–4966, Aug. 2012. [21] J. E. Nichol, A. Shaker, and M.-S. Wong, “Application of highresolution stereo satellite images to detailed landslide hazard assessment,” Geomorphology, vol. 76, no. 1/2, pp. 68–75, Jun. 2006. [22] B. D. Malamud, D. L. Turcotte, F. Guzzetti, and P. Reichenbach, “Landslide inventories and their statistical properties,” Earth Surf. Process. Landforms, vol. 29, no. 6, pp. 687–711, Jun. 2004. [23] G. Antonini et al., “Surface deposits and landslide inventory map of the area affected by the 1997 Umbria-Marche earthquakes,” Boll. Della Soc. Geol. Italiana, vol. 121, no. 1, pp. 843–853, 2002. [24] F. V. De Blasio, “Landslides in Valles Marineris (Mars): A possible role of basal lubrication by sub-surface ice,” Planet. Space Sci., vol. 59, no. 13, pp. 1384–1392, Oct. 2011. [25] J.-Y. Rau, J.-P. Jhan, and R.-J. Rau, “Semiautomatic object-oriented landslide recognition scheme from multisensor optical imagery and DEM,” IEEE Trans. Geosci. Remote Sens., vol. 52, no. 2, pp. 1336–1349, Feb. 2014. [26] T. R. Martha, N. Kerle, C. J. van Westen, V. Jetten, and K. V. Kumar, “Segment optimization and data-driven thresholding for knowledge-based landslide detection by object-based image analysis,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 12, pp. 4928–4943, Dec. 2011. [27] M. H. Derron and M. Jaboyedoff, “Preface ‘LIDAR and DEM techniques for landslides monitoring and characterization’,” Nat. Hazards Earth Syst. Sci., vol. 10, no. 9, pp. 1877–1879, Sep. 2010. [28] K. C. Slatton, W. E. Carter, R. L. Shrestha, and W. Dietrich, “Airborne laser swath mapping: Achieving the resolution and accuracy required for geosurficial research,” Geophys. Res. Lett., vol. 34, no. 23, Dec. 2007, Art. ID L23S10. [29] K. A. Razak, M. W. Straatsma, C. J. Van Westen, J. P. Malet, and S. M. De Jong, “Airborne laser scanning of forested landslides characterization: Terrain model quality and visualization,” Geomorphology, vol. 126, no. 1/2, pp. 186–200, Mar. 2011. [30] W. C. Haneberg, W. F. Cole, and G. Kasali, “High-resolution lidarbased landslide hazard mapping and modeling, UCSF Parnassus Campus, San Francisco, USA,” Bull. Eng. Geol. Environ., vol. 68, no. 2, pp. 263–276, May 2009. [31] F. Ardizzone, M. Cardinali, M. Galli, F. Guzzetti, and P. Reichenbach, “Identification and mapping of recent rainfall-induced landslides using elevation data collected by airborne Lidar,” Nat. Hazards Earth Syst. Sci., vol. 7, no. 6, pp. 637–650, Nov. 2007. [32] U. Mann, B. Pradhan, N. Prechtel, and M. F. Buchroithner, “An automated approach for detection of shallow landslides from LiDAR derived DEM using geomorphological indicators in a tropical forest,” in Terrigenous Mass Movements, P. Biswajeet and B. Manfred, Eds. Berlin, Germany: Springer-Verlag, 2012, pp. 1–22. [33] A. Abellán, J. Vilaplana, and J. Martínez, “Application of a long-range Terrestrial Laser Scanner to a detailed rockfall study at Vall de Núria (Eastern Pyrenees, Spain),” Eng. Geol., vol. 88, no. 3/4, pp. 136–148, Dec. 2006. [34] G. Teza, A. Galgaro, N. Zaltron, and R. Genevois, “Terrestrial laser scanner to detect landslide displacement fields: A new approach,” Int. J. Remote Sens., vol. 28, no. 16, pp. 3425–3446, Aug. 2007. [35] A. Bauer, G. Paar, and A. Kaltenböck, “Mass movement monitoring using terrestrial laser scanner for rock fall management,” in Geo-Information for Disaster Management, V. O. Peter, Z. Siyka, and M. F. Elfriede, Eds. Berlin, Germany: Springer-Verlag, 2005, pp. 393–406. [36] J. Nichol and M. Wong, “Detection and interpretation of landslides using satellite images,” Land Degradation Develop., vol. 16, no. 3, pp. 243–255, May/Jun. 2005. [37] M. Koga and A. Iwasaki, “Improving the measurement accuracy of three-dimensional topography changes using optical satellite stereo image data,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 8, pp. 2918–2923, Aug. 2011. [38] A. Mondini et al., “Semi-automatic recognition and mapping of rainfall induced shallow landslides using optical satellite images,” Remote Sens. Environ., vol. 115, no. 7, pp. 1743–1757, Jul. 2011.

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

[39] K. R. Czuchlewski, J. K. Weissel, and Y. Kim, “Polarimetric synthetic aperture radar study of the Tsaoling landslide generated by the 1999 ChiChi earthquake, Taiwan,” J. Geophys. Res., Earth Surf., vol. 108, no. F1, pp. 7.1–7.11, Dec. 2003. [40] J. Gao and J. Maro, “Topographic controls on evolution of shallow landslides in pastoral Wairarapa, New Zealand, 1979–2003,” Geomorphology, vol. 114, no. 3, pp. 373–381, Jan. 2010. [41] P. Rosin and J. Hervas, “Remote sensing image thresholding methods for determining landslide activity,” Int. J. Remote Sens., vol. 26, no. 6, pp. 1075–1092, Mar. 2005. [42] A. Borghuis, K. Chang, and H. Lee, “Comparison between automated and manual mapping of typhoon-triggered landslides from SPOT-5 imagery,” Int. J. Remote Sens., vol. 28, no. 8, pp. 1843–1856, Apr. 2007. [43] X. Yang and L. Chen, “Using multi-temporal remote sensor imagery to detect earthquake-triggered landslides,” Int. J. Appl. Earth Observ. Geoinf., vol. 12, no. 6, pp. 487–495, Dec. 2010. [44] A. Stumpf and N. Kerle, “Object-oriented mapping of landslides using Random Forests,” Remote Sens. Environ., vol. 115, no. 10, pp. 2564–2577, Oct. 2011. [45] V. Singhroy, K. Mattar, and A. Gray, “Landslide characterisation in Canada using interferometric SAR and combined SAR and TM images,” Adv. Space Res., vol. 21, no. 3, pp. 465–476, 1998. [46] M. N. Jebur, B. Pradhan, and M. S. Tehrany, “Detection of vertical slope movement in highly vegetated tropical area of Gunung pass landslide, Malaysia, using L-band InSAR technique,” Geosci. J., vol. 18, no. 1, pp. 61–68, Mar. 2013. [47] M. N. Jebur, B. Pradhan, and M. S. Tehrany, “Using ALOS PALSAR derived high-resolution DInSAR to detect slow-moving landslides in tropical forest: Cameron Highlands, Malaysia,” Geomatics, Nat. Hazards Risk, to be published. [48] L. Cascini, G. Fornaro, and D. Peduto, “Analysis at medium scale of lowresolution DInSAR data in slow-moving landslide-affected areas,” ISPRS J. Photogramm. Remote Sens., vol. 64, no. 6, pp. 598–611, Nov. 2009. [49] M. N. Jebur, B. Pradhan, and M. S. Tehrany, “Optimization of landslide conditioning factors using very high-resolution airborne laser scanning (LiDAR) data at catchment scale,” Remote Sens. Environ., vol. 152, pp. 150–165, Sep. 2014. [50] A. A. Hassaballa, O. F. Althuwaynee, and B. Pradhan, “Extraction of soil moisture from RADARSAT-1 and its role in the formation of the 6 December 2008 landslide at Bukit Antarabangsa, Kuala Lumpur,” Arabian J. Geosci., vol. 7, no. 7, pp. 2831–2840, Jul. 2014. [51] V. Saeidi, B. Pradhan, M. Idrees, and Z. Latif, “Fusion of airborne LiDAR with multispectral SPOT 5 image for enhancement of feature extraction using Dempster-Shafer theory,” IEEE Trans. Geosci. Remote Sens., vol. 52, no. 10, pp. 6017–6025, Oct. 2014. [52] K. Cheng, C. Wei, and S. Chang, “Locating landslides using multitemporal satellite images,” Adv. Space Res., vol. 33, no. 3, pp. 296–301, 2004. [53] V. Moosavi, A. Talebi, and B. Shirmohammadi, “Producing a landslide inventory map using pixel-based and object-oriented approaches optimized by Taguchi method,” Geomorphology, vol. 204, pp. 646–656, Jan. 2014. [54] L. Dr˘agu¸t and T. Blaschke, “Automated classification of landform elements using object-based image analysis,” Geomorphology, vol. 81, no. 3/4, pp. 330–344, Nov. 2006. [55] N. Kerle and J. de Leeuw, “Reviving legacy population maps with objectoriented image processing techniques,” IEEE Trans. Geosci. Remote Sens., vol. 47, no. 7, pp. 2392–2402, Jul. 2009. [56] G. Taguchi, Introduction to Quality Engineering. New York, NY, USA: McGraw-Hill, 1990, p. 191. [57] V. S. Petrovic and C. S. Xydeas, “Gradient-based multiresolution image fusion,” IEEE Trans. Image Process., vol. 13, no. 2, pp. 228–237, Feb. 2004. [58] T. M. Tu, P. S. Huang, C. L. Hung, and C. P. Chang, “A fast intensityhue-saturation fusion technique with spectral adjustment for IKONOS imagery,” IEEE Geosci. Remote Sens. Lett., vol. 1, no. 4, pp. 309–312, Oct. 2004. [59] M. González-Audícana, J. L. Saleta, R. G. Catalán, and R. García, “Fusion of multispectral and panchromatic images using improved IHS and PCA mergers based on wavelet decomposition,” IEEE Trans. Geosci. Remote Sens., vol. 42, no. 6, pp. 1291–1299, Jun. 2004. [60] J. G. Liu, “Evaluation of Landsat-7 ETM+ panchromatic band for image fusion with multispectral bands,” Nat. Resour. Res., vol. 9, no. 4, pp. 269–276, Dec. 2000. [61] J. Nunez et al., “Multiresolution-based image fusion with additive wavelet decomposition,” IEEE Trans. Geosci. Remote Sens., vol. 37, no. 3, pp. 1204–1211, May 1999.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. PRADHAN et al.: DATA FUSION USING WAVELET TRANSFORM AND TAGUCHI METHODS FOR DETECTION

[62] K. Amolins, Y. Zhang, and P. Dare, “Wavelet based image fusion techniques—An introduction, review and comparison,” ISPRS J. Photogramm. Remote Sens., vol. 62, no. 4, pp. 249–263, Sep. 2007. [63] T. Hilker et al., “Generation of dense time series synthetic Landsat data through data blending with MODIS using a spatial and temporal adaptive reflectance fusion model,” Remote Sens. Environ., vol. 113, no. 9, pp. 1988–1999, Sep. 2009. [64] M. N. Jebur, H. Z. Mohd Shafri, B. Pradhan, and M. S. Tehrany, “Perpixel and object-oriented classification methods for mapping urban land cover extraction using SPOT 5 imagery,” Geocarto Int., vol. 29, no. 7, pp. 792–806, Oct. 2013. [65] T. Blaschke, “Object based image analysis for remote sensing,” ISPRS J. Photogramm. Remote Sens., vol. 65, no. 1, pp. 2–16, Jan. 2010. [66] A. S. Laliberte et al., “Object-oriented image analysis for mapping shrub encroachment from 1937 to 2003 in southern New Mexico,” Remote Sens. Environ., vol. 93, no. 1/2, pp. 198–210, Oct. 2004. [67] M. S. Tehrany, B. Pradhan, and M. N. Jebu, “A comparative assessment between object and pixel-based classification approaches for land use/land cover mapping using SPOT 5 imagery,” Geocarto Int., vol. 29, no. 4, pp. 351–359, May 2013. [68] L. Dr˘agut¸t and C. Eisank, “Object representations at multiple scales from digital elevation models,” Geomorphology, vol. 129, no. 3/4, pp. 183–189, Jun. 2011. [69] M. S. Tehrany, B. Pradhan, and M. N. Jebur, “Remote sensing data reveals eco-environmental changes in urban areas of Klang Valley, Malaysia: Contribution from object based analysis,” J. Indian Soc. Remote Sens., vol. 41, no. 4, pp. 981–991, Dec. 2013. [70] A. S. Laliberte and A. Rango, “Texture and scale in object-based analysis of subdecimeter resolution Unmanned Aerial Vehicle (UAV) imagery,” IEEE Trans. Geosci. Remote Sens., vol. 47, no. 3, pp. 761–770, Mar. 2009. [71] M. Möller, L. Lymburner, and M. Volk, “The comparison index: A tool for assessing the accuracy of image segmentation,” Int. J. Appl. Earth Observ. Geoinf., vol. 9, no. 3, pp. 311–321, Aug. 2007. [72] U. C. Benz, P. Hofmann, G. Willhauck, I. Lingenfelder, and M. Heynen, “Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready information,” ISPRS J. Photogramm. Remote Sens., vol. 58, no. 3/4, pp. 239–258, Jan. 2004. [73] G. M. Foody, “Thematic map comparison: Evaluating the statistical significance of differences in classification accuracy,” Photogramm. Eng. Remote Sens., vol. 70, no. 5, pp. 627–634, May 2004.

Biswajeet Pradhan received the B.Sc. degree with honors from Berhampur University, Brahmapur, India; the M.Sc. degree from the Indian Institute of Technology (IIT), Bombay, India; the M.Tech. degree in civil engineering from the IIT, Kanpur, India; the Habilitation in “Remote Sensing” from the Dresden University of Technology, Dresden, Germany, in 2011; and the Ph.D. degree in GIS and geomatics engineering from the Universiti Putra Malaysia, Serdang, Malaysia. He is currently a Faculty Member of the Department of Civil Engineering, Universiti Putra Malaysia. He has more than 16 years of teaching, research, consultancy, and industrial experience. Out of his more than 330 articles, more than 230 have been published in science citation index (SCI/SCIE) technical journals. He has written two books in GIS data compression and disaster management and edited three volumes, and he has written 12 book chapters. He specializes in remote sensing, GIS application, and soft-computing techniques in natural hazard and environmental problems. Prof. Pradhan has been a Humboldt Ambassador Scientist for the Alexander Von Humboldt Foundation, Germany, since March 2015. He was a recipient of the Alexander von Humboldt Research Fellowship from Germany during 2008–2010. He was also a recipient of the prestigious German Academic Exchange Research (DAAD) Fellowship Award, the Saxony State Fellowship from 1999 to 2002, the Keith Aurtherton Research Award, and the Georg Forster Research Award from the German Government. He is on the editorial board of many Institute for Scientific Information (ISI) journals.

13

Mustafa Neamah Jebur was born in Misan, Iraq, on January 11, 1988. He received the B.Sc. degree in environmental engineering from the Allameh Mohaddes Nouri University, Nur County, Iran, in 2007 and the M.Sc. degree in civil engineering and the Ph.D. degree in GIS and geomatics engineering from the Universiti Putra Malaysia, Serdang, Malaysia, in 2013 and 2015, respectively. He is currently with the Department of Civil Engineering, Geospatial Information Science Research Center, Faculty of Engineering, Universiti Putra Malaysia. His areas of interest include flooding, landslides, and ensemble modeling applications.

Helmi Zulhaidi Mohd Shafri received the bachelor’s degree (first-class honors) in surveying from the Royal Melbourne Institute of Technology, University, Melbourne, Australia, in 1998 and the Ph.D. degree in remote sensing from The University of Nottingham, Nottingham, U.K., in 2003. He is currently the Coordinator of the Remote Sensing and GIS Program at the Faculty of Engineering, Universiti Putra Malaysia, Serdang, Malaysia. He is actively involved in research related to algorithm development and new applications of remote sensing, particularly in the areas of urban engineering and environmental informatics. He has more than 12 years of teaching, research, administrative, and consultancy experience, with more than 80 papers in refereed technical journals.

Mahyat Shafapour Tehrany was born in Tehran, Iran, on July 29, 1985. She received the B.Sc. degree in environmental engineering from the Allameh Mohaddes Nouri University, Nur County, Iran, in 2007 and the M.Sc. degree in civil engineering and the Ph.D. degree in GIS and geomatics engineering from the Universiti Putra Malaysia, Serdang, Malaysia, in 2013 and 2015, respectively. She is currently with the Department of Civil Engineering, Geospatial Information Science Research Center, Faculty of Engineering, Universiti Putra Malaysia. Her areas of interest include flood, landslides, and ensemble modeling applications.