Method 4: A Hybrid Object-oriented Method

Most of the conventional automatic classification methods for landslide detection are mainly based on spectral features of remotely-sensed images other than topographic features. Because the spectral features of buildings and roads are similar to those of landslides, therefore serious mis-interpretation took place.

Moreover, limitations are due to the spatial and spectral resolutions of the images. More than 50% of the rainfall-induced landslides in Taiwan are less than 50.0 m in length (Liu et al., 2012). Landslides of this scale are not readily

45

identifiable using images of a pixel-size larger than 10.0 m. By pixel-wise classification, landslides can occupy only individual or just a few pixels without forming an outer shape of landslides (Liu et al., 2009). Moreover, commission and omission errors of pixel-based classification can further complicate the situation. The pixel-based methods are then required to be replaced with approaches based on objects or segments (Kerle and Martha, 2010). Therefore, The main objective of this research is to combine both of an unsupervised of region-based image segmentation and a supervised classification method with SVM classifier using standard products from airborne LiDAR survey, as shown in the scheme of landslide detection in Figure 3.1.

(B) Method

Basic task of segmentation algorithms is the merge of image elements based on homogeneity parameters or on the differentiation to neighboring regions, respectively. Thus, segmentation methods follow the two strongly correlated principles of neighborhood and similarity of pixel values. The region-based approaches start in image space where the available elements either pixels or already existing regions are tested for similarity against other elements. Concerning the definition of the initial segmentation the procedures of region growing (i.e. bottom-up, i.e. starting with a seed pixel) and region splitting (i.e. top-down, i.e. starting with the entire scene) are distinguished. One disadvantage of the splitting method is that it tends to be over-segmented because a splitting always produces a fixed number of sub-regions (normally: 4)

46

although two or three of them might actually be homogeneous with respect to each other. As a consequence, one can apply an integration of the various methods. Thus, it leads to the split-and-merge algorithm that after a split process.

If neighboring regions are similar, they should be remerged again (Chang et al., 2010).

To strengthen the automation of segmentation, clustering is adopted for region-based segmentation. The ISOCLUST is an iterative self-organizing unsupervised classifier based on a concept similar to ISODATA routine of Ball and Hall (1965) and cluster routines such as the H-means and K-means procedures (Jain and Dubes, 1988). Object-oriented analysis (OOA) is inherently more suitable, as it can address the phenomena under study such as landslides in this case, as that they are “objects”, not “pixels” that have spectral, spatial and contextual characteristics. Thus in this study, the unsupervised classification method ISODATA are applied to the nDSM image to find out an optimal range of data values to present landslide distribution (Research System, Inc., 2006).

After the segmentation, a supervised classification of the segments with SVM classifier is applied to obtain landslide class.

SVM (Support Vector Machine) is a relatively new classifier and is based on strong foundations from the broad area of statistical learning theory. Since its inception in early 90s, it has found applications in a wide range of pattern recognition problems, image classification, financial time series prediction, face

47

detection, biomedical signal analysis, medical diagnostics, and data mining (Witten & Frank, 2000; Hwang and Chiang, 2010; Burges, 1998; Chapelle et al., 1999). Under the basic assumption of the SVM approach, the training sample is expressed as ， ， ， ，_… ， where xi ∈ represents an input mode, yi ∈ ±1 . The optimal decision-making formula is as follows:

+ = 0 (3.1) The weighing vectors w and b is deemed satisfactory once converged.

+ ≥ 1 − "# (3.2)

The value "_# is a loose variable existing in a linear, undividable condition.

It describes the degree of module deviation under the ideal linear circumstances.

The goal of the SVM is to identify a decision support phase where the average error of the training samples is minimized. The optimization equation is therefore derived as follows:

φ w, ε = (⁾w + c ∑ ε^,_#- ^# (3.3) Where C is a positive parameter assigned by the end user. It serves as a penalty for the correctness of the SVM. The C value is used to leverage the probable mis-interpretation percentage and the complexity of the algorithm. A converged optimization equation can be derived adopting the Lagrange Multiplication Method:

. / = ∑ / − ⁰_- ∑ ∑ /⁰_- ⁰_1- /_{1 1} 2 , ₁ (3.4)

48

Where α# #-, is the Lagrange multiplier while Eq. (4) fulfilling the following criteria.

∑ /⁰_- = 0, 0 ≤ / ≤ 5, 6 = 1,2,3 … , 9 (3.5)

K ( , ₁ ) is a core function. There are four types of core functions included in the Mercer Theorem:

1.Linear： 2 , 1 = : 1 (3.6) 2. Polynomial: 2 , ₁ = ;: ₁ + < ^d,;>0^d  (3.7) 3. radial basis function (RBF): 2 ₁ = = > ?−;@ − ₁@ A , ; > 0

(3.8) 4. Sigmoid: 2 , ₁ = Cℎ ;: ₁ + < , ; > 0 (3.9) Here, ;, r, and d are kernel parameters (Burges, 1998).

The data flow is shown in Figure 3.10 where the feature space for landslide classification includes only the standard LiDAR survey products, namely ortho image, DEM and DSM. Especially, the derivatives of greenness, slope and nDSM are used for input for the segmentation (Figures 3.11).

Accuracy validation will be made against the results obtained by visual interpretation using all available images derived from the same datasets (Figure 3.12). In addition, a pixel-based classification with the same datasets is also carried out for comparison.

49

Figure 3.10 Flowchart of OOA data processing

(A) Greenness (B) Slope (C) nDSM

Figure 3.11 The three derivatives of othophoto, DEM and DSM for data entry

50

(C) Testing site and materials

A test area is selected in I-Lan County of eastern Taiwan, the same as the one for the test of Method with indices of point cloud density in Paragraph 3.1.3 (C).

The attributes of the data sets is shown in Table 3.2. For this test, only a subset of the area is used (Figure 3.12).

Figure 3.12 Study Area and Ground Truth for OOA Test

(D) Results and discussion

Figure 3.13 shows the results of segmentation and vector classification.

The results generated by pixel-based classification are shown in Figure 3.14.

The comparison between the results of the hybrid OOA method, pixel-based SVM method and those of ground truth is shown in Table 3.3. The Producer accuracy of landslide class by object-based method is 85.68% whereas that by pixel-based method is 72.01%. The user accuracy of landslide class by object-based method is 80.41% whereas that by pixel-based method is 76.2%.

51

Kappa coefficient and overall accuracy of object-based method are 0. 817 and 93.4%, respectively. It is concluded that the hybrid OOA method proposed in this study is an effective method which is better than pixel-based method.

(A) segmentation (B) Training samples (C) OOA result Figure 3.13 Results of the hybrid OOA method

52

(A) training areas on orthophoto (B) Landslide map Figure 3.14 The results generated by pixel-based SVM classification

Table 3.3 Confusion table of OOA classification and pixel-based classification

Object-based PA％ UA％ Pixel-based PA％ UA％

Landslide 85.68 80.41 Landslide 72.01 76.2

River bed 83.21 87.81 River bed 70.88 76.2

Vegetation 95.72 96.59 Vegetation 94.55 93.53 Kappa Coefficient = 0.817 Kappa Coefficient = 0.704 Overall Accuracy = 93.4% Overall Accuracy = 89.64%

3.2 Detection of Deep-seated Landslides