Method 2: Method with Indices of Point Cloud Density

Basic products of airborne LiDAR include all points, ground points, digital

31

elevation model (DEM), and digital surface model (DSM). The former two are vectors of discrete points and the later two are interpolated raster grids of the discrete points of the former two. The DEM and DSM grids are commonly used for applications whereas point clouds are rarely used. LiDAR discrete points are worthy of a further study due to the fruitful information adhered with the attributes of individual points. Point density has been used as an important indicator of DEM/DSM quality (Shih and Huang, 2006; Liu et al., 2007; Puetz et al., 2009; Raber, 2003). An understanding of the forest closure and crown density can be obtained by inspection of the point-density distribution of point clouds (Dubayah and Blair, 2000; Means et al., 2000; Naesset, 2002). Therefore, point density derived from specific properties of point clouds can be used to explore the possibility of extracting landslide information from point clouds.

Visual interpretation of shaded-relief image derived from DEM is usually adopted by geologists whereas other LiDAR products have not been commonly applied. In this paper, possible derived indices from point clouds are discussed first and then experiments of selected indices are made to find out the most descriptive ones for landslide detection.

(B) Deriving a point density map

The attributes of individual points of LiDAR point clouds are recorded in a LAS format. The format contains binary data consisting of a header block, variable length records, and point data (ASPRS, 2008&2009). Each point data record

includes the XYZ coordinates, intensity, return number, number of returns, scan direction, and classification of the point. These attributes of point cloud are closely related to the geometry of laser scanning configuration and thus relevant to the point density of unit ground area. The spatial distribution of point density implies the properties of the land surface.

determined by the relationship between the location of laser head, scanning angle, location of object and shape of the ground surface, as

shown in Figure 3.4. In other words, factors for point density include all these flight design parameters and the effects of ground conditions:

look angle; (3) flight height (6)strip overlap; (7) terrain relief

Figure 3.4 Schematic diagram showing the geometry of airborne LiDAR

32

includes the XYZ coordinates, intensity, return number, number of returns, scan direction, and classification of the point. These attributes of point cloud are closely related to the geometry of laser scanning configuration and thus relevant to the point density of unit ground area. The spatial distribution of point density implies the properties of the land surface. Exact coordinates of each point are determined by the relationship between the location of laser head, scanning

of object and shape of the ground surface, as

. In other words, factors for point density include all these flight design parameters and the effects of ground conditions:

(3) flight height; (4) plane attitude (roll, yaw, pitch);

(7) terrain relief; and (8) above-ground objects

Schematic diagram showing the geometry of airborne LiDAR scanning

includes the XYZ coordinates, intensity, return number, number of returns, scan direction, and classification of the point. These attributes of point cloud are closely related to the geometry of laser scanning configuration and thus relevant to the point density of unit ground area. The spatial distribution of point density Exact coordinates of each point are determined by the relationship between the location of laser head, scanning of object and shape of the ground surface, as schematically . In other words, factors for point density include all these flight design parameters and the effects of ground conditions: (1) pulse rate; (2) 4) plane attitude (roll, yaw, pitch); (5) flight speed;

ground objects.

Schematic diagram showing the geometry of airborne LiDAR

33

For discriminating landslide and non-landslide lands, the types of point clouds for point density enumeration can be categorized as follows: (1) all points, (2) ground points, (3) single-echoes points or only-echoes points, (4) multiple-echoes points, i.e. (first + intermediate + last) returns, (5) first-return points, (6) intermediate-return points, and (7) last-return points. Secondary indices can also be created by combining two or more types of point clouds, for example, penetration rate can be derived by the ratio of ground and all points denoting the fraction of points hitting the bare ground. For exploring the capability of point clouds for the detection of landslides, four types of point density with five searching radii are used in this study, including point density type of all points, ground points, only-echoes points, and multiple-echoes points.

Comparison of them will be made to search for most descriptive ones for landslides.

Point density can be measured by various approaches (Shih and Huang, 2006).

In this study, point density is measured by subdividing the surveyed area into grid cells, then computing the unit density of the number of points in a circle with certain searching radius centered at the cell center. A software application is implemented in this study to cater for the output grid size, searching radius, and type of points. This method is comparable to that by Crosby (2007). In this dedicated software application, the function of reading Terrascan PTC file for point class definition is also implemented so that various type of point density

34

can be designated. Point density distribution of ground points is affected by the criteria and procedures of both automated and manual editing process.

Nevertheless, point density except that of ground points is mainly decided by flight operation parameters including pulse rate, look angle, flight height, aircraft attitude, flight speed, strip overlap, terrain relief, and above-ground objects. Because the average ground points of the selected study area is 0.75 pts/m², 1 m is selected for grid spacing. To cater for the effects of the uniformity of point distribution, ground surface undulation and land-cover types, five searching radii are used, i.e. 0.707m, 1.414m, 3.0m, 5.0m, and 10 m.

(C) Testing site and materials

The study area is located in I-Lan County of northeastern Taiwan, on the track of the Typhoon Kalmaegi attacked Taiwan on July 16th~18th, 2008, about nine month after Typhoon Krosa on October 4th, 2007 in this area. The dataset for the experiment was taken on 4th November, 2008 after Typhoon Kalmaegi. In general, the accuracy of bare grounds checked in the field is about 0.15 m. An area covering 2 km by 2 km is selected for the experiment. The overall point density of the study area is 2.75 points/m² with ground point density of 0.75 points/m² (Table 3.2).

35

Table 3.2 Attributes of the LiDAR data used in point density study Type of points All points Ground

points

points 12,142,434 3,320,615 5,789,148 6,353,286 Average point searching radii, findings are as follows:

(1) Striping noise of point density map is obviously affected by flight speed and strip source as shown in Figure 3.5.

(2) Point density map of multiple-echoes point gives better contrast between landslide and non-landslide areas than any maps derived from other three types of point density, as shown in Fig. 3.6 (B).

(3) For output of 1m grid spacing, point density map with a searching radius of 1.414 m shows best result among all radii including 0.707, 1.414, 3, 5, and 10 m.

This result is subjected to the density of all points and ground points. A larger radius cannot give better enhancement of landslides.

(4) Although the overall point density of only-echoes points of the whole study area is similar to that of multiple-echoes points as shown in Table 3.2, a conspicuous contrast of landslide area is observed on the density map of

36

multiple-echoes points other than that of only-echoes points. This is due to a high concentration of multiple-echoes points in forested land and most of bare grounds are covered by only-echoes points.

(5) Landslide feature is conspicuous in some part of the density map of ground points whereas it is vague in other parts. This is a consequence of the factors of penetration rate in different part of the area and the filtering process of non-ground points with both automated algorithm and manual editing. However, on the map of all points overlaid by ground points, landslides features can be enhanced for visualization. Nevertheless, commission errors of landslide interpretation can be serious on this map, especially in those bare lands which are not landslides. These errors might be eliminated by slope gradient of ground surface in a later step.

37

Figure 3.5 Selected results of four types of point density and their distribution under various searching radii with 1m grid.

38

(A) (B)

Figure 3.6 (A) Point cloud distribution with attribute of flight strip source ID.

(B) Density map of multiple-return echoes with r=1.414m, grid spacing = 1m.

3.1.4 Method 3: Method of nDSM slicing