AUTOMATED GRID GENERATION FOR FLOOD PREDICTION USING LiDAR DATA RYOTA TSUBAKI 1, ICHIRO FUJITA 2 and GEORGE V. LUGOMELA3 1

Ph.D student, Graduate School of Science and Technology, Kobe University, Rokkodai, Nada-Ku, Kobe, 657-8501, Japan (e-mail: [email protected],) 2 Professor, Department of Architecture and Civil Engineering, Kobe University, Rokkodai, Nada-Ku, Kobe, 657-8501, Japan (Tel: 81-78-803-6439, Fax: 81-78-803-6439, e-mail: [email protected]) 3 Ph.D candidate, Graduate School of Science and Technology, Kobe University (e-mail: [email protected] )

Abstract Due to the complicated topography, it is difficult to predict flood flow around urbanized areas. Recently, airborne laser mapping system has become available which is capable of giving very detailed topographical information. In this study, we developed a high resolution inundation analysis system using airborne laser survey results. The system enables an automatic grid generation while detecting location of buildings and streets. With this system, it would become possible to construct high accurate hazard maps in urban areas. Keywords: Inundation analysis; LiDAR; Mesh generation; Unstructured grid; Tsunami. 1. INTRODUCTION 1.1. BACKGROUND Many cities have experienced flood disasters and tsunami waves in the past and suffered from the resulting damages. Despite the measures taken to minimize the loss from the catastrophes, still these cities and others that haven’t experienced could suffer from the damages. There are two approaches to control inundation catastrophes, namely, hardware measures and software measures. Japanese flood control policy weighted heavily on hardware measures such as to straighten rivers, to fix riverbeds, to make drainage system and to construct underground rivers. But these hardware approaches have space and cost limitations. The former is particularly sensitive in urban areas and the latter in rural areas. In addition, urbanization itself increases runoff discharge. Therefore, software usages have become more and more practical nowadays because hardware measures are not always successful in containing large floods or tsunamis. One of the software measures is an accurate and reliable flood prediction which should be performed for the following tasks.

#1. Estimating human loss, suffering and financial damage in advance. #2. Provide data which support decision making procedure in terms of which areas should be protected and which ones should be alarmed. #3. Detailed analysis related to provision of information for evacuation planning. But it is difficult or it takes extra efforts to predict flood flow around urbanized area for topographical reasons. The slope of each street varies, and complex arrangement of buildings, vegetation and road network has much effect on flood flow behavior. Recently, airborne laser mapping system (LiDAR; Light Detection And Ranging), which can give very detailed topographical information, has become available. In this system, measurement is done intrusively by an instrument set on an airplane so it is possible to obtain topography of a large area accurately in a short period. 1.2. OBJECTIVE The objective of this study is to develop not only accurate but also automated and labor saving procedure for converting LiDAR data into information which is used for flood prediction calculations. An aerial photograph and field survey are conventionally used to make a city planning map, but LiDAR is more useful for this task because of its convenience and efficiency. Therefore more areas will be mapped using LiDAR and the data will easily and cheaply be available in the near future. A conventional line drawn map has insufficient information about both height distribution and vegetation although both conditions affect inundation flow significantly, but such information would become available by using LiDAR. 1.3. PREVIOUS WORKS For post processing of the LiDAR measurement, a filter is applied to extract bare-earth, buildings, trees and so on. A number of filtering algorithms have been developed for this purpose. Vosselman et al. (2004) reviewed different techniques for the extraction of surface from point clouds obtained by both airborne and terrestrial laser scanning. Sithole and Vosselman (2003) described a scan line segmentation method, used especially to extract connection of roads including over headed roads such as bridges and ramps. Rottensteiner et al. (2003) presented an automated detection method of buildings especially in residential areas from LiDAR data whose resolution is comparatively low by using multi-spectral images. In this study, LiDAR data is processed by a statistical filtering procedure for three dimensional space distribution within the template that is developed by Asia Air Survey Co., ltd. There are several models to estimate the inundation flow through urban areas. A number of commercial software packages are used in practice, and several models are developed based on Japanese conditions and features. Kawaike et al. (2000) developed a two-dimensional flow model with unstructured meshes. Takeda et al. (2004) also developed a PRE to POST processing system, which includes models such as rivers, channels, sewers and flood flow

past an urban area. In this model, the Cartesian grid is used for flooding flow prediction. Fig. 1 shows modeling strategies according to the scale factor. In this study, unstructured grid is generated to establish a complicated topography with comparatively fewer grid number of the calculating mesh compared to the one using Cartesian grid. The size of each grid is supposed to be about 5-30m, which is associated with the width of road networks and the spacing between the buildings. 2. LiDAR MEASUREMENT In the LiDAR measurement, the airborne laser mapping (LiDAR), inertial navigation system, GPS and laser scanning are used together as shown in Fig. 2. Its accuracy is 10-30 cm in vertical direction and 1m in horizontal direction. Laser scanning is done intermittently by shooting pulses of the laser. The direction of laser shots is changed transversely as the airplane goes forward, as a result the topography is measured as a points-cloud of location data along the path of the airplane. In the raw data, the spacing between each point is irregular (cloud of points) hence interpolation to Cartesian grid is commonly performed. Raw data itself contains only surface height (height of the bare-earth, the roofs of buildings, the leaves of trees and power lines). The surface height is interpolated to Cartesian grid and it is called the digital surface model (DSM). The filtering procedure is applied to the raw data to extract height of bare-earth only, and the digital elevation model (DEM; The height distribution of ground) is obtained by interpolation, because ground height under the objects isn’t measured in LiDAR directly. In this study, DSM and DEM data with a spacing of 1m is used to generate topographical information for a flood analysis. 3. WORKFLOW OF GRID GENERATION A workflow of our method is depicted in Fig. 3. The method is described as follows, i) Extract buildings from DSM and DEM, and make mask image. ii) Extract outlines of each building. iii) Outline grid is generated and inner area is filled up by using advancing front method. iv) Elevation of each grid is configured by DEM information. A detailed account of the procedure is described in the subsequent sections.

4. BUILDING EXTRACTION The difference of the elevation between DSM and DEM indicates a height of non bareearth objects such as buildings, structures and trees. To distinguish the bare-earth and the other objects, a threshold (1m in this study) is used for the difference of height between the DSM and DEM, then the mask image (M(x,y):M=1,0) is generated where M=0 means the presence of bare-earth and M=1 indicates the other objects. Each consecutive region of M=1

(candidate for the building) is identified and the area which is small (15m2 or less, in this study) is neglected and considered to be bare earth (M=0). Such small area may mean a car, poles and/or road side small trees, and may have small influence on the flow, additionally, to represent it, too small grid spacing is needed. The spacing of the mask image is 1m (same as that of DSM). 5. OUTLINE EXTRACTION To present a road network precisely with a smaller number of grids, it is important to generate the grid system that represents the location of all edges of the buildings. The mask image M(x,y), obtained above, indicates only where objects like buildings exist or not, so the outline of each building is extracted and each edge of this outline is picked up by the following procedure. First, the outline of each building of the mask image is obtained by clockwise (or anti clockwise) tracing of the outline (boundary between M=0 and M=1). The outline is treated by a series of the location points (xi , yi : i=0.. imax). Second, to abstract information about the shape of each building by using the series of locations, the presence of the edge is identified by outline arrows (see Figs. 5 a and b). Then, the location of the edge is assumed by extrapolation as shown in Fig. 5 c. Although information about the edge cannot represent rounded form of some buildings, the locations of the intermediate points of the curved outline (not edge) are considered at some interval. The size of the interval is tuned to the spacing of objected mesh size and 20m is used in this study. 6. UNSTRUCTURED GRID GENERATION By using the outline information extracted above, an unstructured grid is generated. There are two strategies, which are commonly used to generate unstructured grid, namely, the Delaunay triangulation and the advancing front method (AFM). In the mesh generation based on Delaunay triangulation strategy, grid points are allocated first, then using these points, triangles are formed using the Delaunay strategy. In AFM strategy, a domain is filled from the boundary to inner area by adding grid points and forming triangles simultaneously. The AFM makes more equilateral triangles around the boundary area as opposed to the Delaunay triangulation method. We think that representation of the flow along the road network is needed, AFM is suitable because it forms equilateral triangles near the boundary area and so AFM is chosen for use in this study. In AFM, the boundary area is defined first (see Fig. 6 a). Along the boundary (initial front), the new triangle (new front) is generated by adding a new grid point as shown in fig. 6 b. The triangles are formed one after another (Fig. 6 c). In some cases the grid points, which are already set, are used to form other triangles (Fig. 6 d). Triangles are formed with and without adding a new point (front advancing to inner area), and finally the entire domain is filled up by triangles. In AFM, equilateral triangles are formed around the boundary area, but irregular

triangles may be formed in the region where the front is merging. Therefore, to refine the mesh, the locations of the grid points which compose irregular triangles are shifted to achieve equilateral triangles or near ones. 7. APPLICATION 7.1. EXTRACTION OF OUTLINE Fig.7 shows a city planning map of the test area, sized 250 m x 250 m. In this area, small residential houses, school buildings and commercial facilities are densely packed. Fig.8 shows extracted outline of the buildings overlaid on the map. Pale red region indicates the road and bare-earth area where flow goes through. As it can be seen in the figure, the layout of each block and road connection is represented well. Fig.9 shows the outline on the aerial photograph. In the aerial photograph, the vegetation can be seen in the areas pointed by green arrows and that is also represented well on the outline information, which is difficult to recognize only by using the City planning map. 7.2. GRID GENERATION Eight calculating meshes with different grid spacing, from 5 m to 30 m, are generated (as shown in table 1). Here, grid spacing is defined as the typical length of a side of the triangle. The target area is extended to the surroundings of the test area and has size of 700 m x 700 m. Fig. 10 shows 4 samples of generated grids. As it can be seen, coarser mesh (25m) cannot express detailed boundary shape and small houses are omitted in the grid. On the other hand, finer meshes (10 m and 5 m) represent an outline configuration well and 8 m mesh (figure not included) also shows similar result. Consequently, with respect to the clarity of the boundary shape, 10m or finer meshes express boundary shapes clearly. As depicted in Fig.11, generating time increases approximately linearly with the number of triangles. Fig.12 shows the relationship between number of triangles and grid size. Evidently, a grid spacing smaller by half needs four times the number of total triangles, and it makes not only a slower grid generation process but also increases calculating time and memory size exponentially. 8. CONCLUSION We developed a high resolution grid generating system for an inundation analysis. In this system, unstructured grid can be generated automatically by using LiDAR data directly, so detailed simulation of inundation through urbanized area can be easily executed with this system. Out future plans are #1, In this study, we processed already filtered DSM and DEM data. We plan to extract information about vegetation, small channel and river by processing raw data of LiDAR measurement directly, and obtain not only the information of road network but also the

vegetation and network of small channels and rivers in a flooding area. #2, A Cartesian grid system can also represent complex roads and buildings locations well when small grid spacing is used, so we need to evaluate the difference of the results which depends on grid spacing and grid system, and determine which is the most appropriate grid system considering accuracy, required memory size and calculating time. ACKNOWLEDGMENT The authors wish to acknowledge Asian Air Survey Co., ltd. for processing LiDAR data. We would like to thank Prof. Takeshi Okabe of the University of Tokushima for providing measurement data. REFERENCES Kawaike, K. et al., 2000. Inundation flow modeling in urban area based on the unstructured meshes. Hydrosoft 2000, WIT Press, pp.457-466. Rottensteiner, F. et al., 2003. Building Detection Using LIDAT DATA and Multi-spectral Images. Proc. VIIth Digital Image Computing: Techniques and Applications, Sydney, pp.673-682. Sithole, G.., Vosselman, G.., 2003. Automatic Structure Detection in a Point Cloud of an Urban Landscape. Proc. 2nd GRSS/ISPRS Joint Workshop on Remote Sensing and Data Fusion over Urban Areas, URBAN2003, Berlin, pp.67-71. Takeda, M. et al., 2005. The Development of Inundation Analysis System in Urban Area. Proc. Monitoring, Prediction and Mitigation of Water-Related Disasters, Kyoto, pp.125-130. Vosselman, G. et al., 2004. Recognising Structure in Laser Scanner Point Clouds. International Archives of Photogrammetry. Remote Sensing and Spatial Information Sciences, vol.46, part 8/W2, Germany, pp.33-38. Table 1 Meshes with different grid spacing. Grid spacing (m) 5 8 10 13 16 20 25 30

Triangles 25200 11500 8000 5500 4000 2900 2100 1500

Grid points Time (min.) 17900 17 9000 4 6500 2 4600 1 3400 less than 1 2500 1800 1300 -

Fig. 1 Inundation modeling strategies.

INS

GPS

Laser scanning

Fig.2 Schematic image of LiDAR measurement. The topographical features shown in this figure was made using actual measured values.

Fig.3 Workflow of grid generation.

a) Mask image (Vegetation exists in pointed area.)

b) City planning map

Fig. 4 Mask image and conventional map.

a)

Searching the edge

b) Nominated point of the edge

c) Estimation of the location of the edge

Fig. 5 Edge detection from a series of points.

Fig. 6 Advancing front method.

250m

Fig. 7 City planning map.

Fig. 8 Extracted outline overlaid on city planning map

Fig. 9 Extracted outline overlaid on aerial photograph.

Fig. 10 Generated meshes with different grid spacing. 30000 Number of triangles

20

Minute

15 10 5 0 0

10000 20000 Number of triangles

30000

Fig.11 Processing time to generate grids

25000 20000 15000 10000 5000 0 0

10 20 30 Grid spacing (m)

40

Fig.12 Total number of the triangles depends on grid spacing