Urban Classification of Metro Manila for Seismic Risk Assessment using Satellite Images Fumio YAMAZAKI/ yamazaki@edm.bosai.go.jp Hajime MITOMI/ mitomi@edm.bosai.go.jp Yalkun YUSUF/ yalkun@edm.bosai.go.jp Masashi MATSUOKA/ matsuoka@edm.bosai.go.jp Earthquake Disaster Mitigation Research Center, NIED, Hyogo, Japan Abstract To evaluate seismic risk in an urban area, building inventory is necessary. However, a large amount of time and effort is required to develop building inventory by field surveys. Thus an easier method to develop building inventory is being sought with the aid of remote sensing technologies. For the first step to evaluate seismic vulnerability of Metro Manila, land use classification was carried out and the expansion of urban areas was studied using time series satellite imagery from Landsat. The densely built-up areas estimated from Landsat were compared with the GIS data, based on the aerial photographs of 1986. A high-resolution satellite image from IKONOS was also employed for microscopic urban modeling of Metro Mania. Using the normalized vegetation index and the texture of the image, a detailed classification of urban areas with respect to the density and height of buildings is being sought. Although this study is still preliminary, some conclusions on possible applications and limitation of satellite remote sensing will be drawn within EQTAP Phase II. Key Words: building inventory, built environment, remote sensing, Metro Manila, Landsat, IKONOS Background and Objectives Seismic risk in an urban area is closely related to the structure, material and dimension of buildings and their spatial distribution. Hence it is important to study these characteristics of buildings in order to evaluate seismic vulnerability of an area, for disaster mitigation planning and pre- and post-event damage assessments. Building inventory in an urban area can be obtained by a field survey. However, a large amount of time and efforts is required. Thus an easier method to develop building inventory is being sought. Satellite remote sensing, which can easily monitor a large area, could provide effective information to develop building inventory, if it is able to capture the built environment in urban areas. Multi-spectral characteristics show the difference of reflectance from the materials on the earth surface. Many researchers have already proposed algorithms to classify the features on the earth surface in the fields of natural environment mapping, such as for forests and agricultural lands. However, few studies are found in built environment mapping, although the technology was used to evaluate the thermal environments in urban areas 1). In this study, the classification for understanding seismic vulnerability in Metro Manila was attempted using Landsat and IKONOS satellite images. Satellite Imagery and GIS data for Metro Manila Satellite images from Landsat-5 2) and IKONOS 3) were used in this study. Multi-spectral sensor of Landsat-5, Thematic Mapper (TM), has seven bands between visible and thermal infrared. Spatial resolution of its image except for the thermal infrared band is 30m on ground. Three Landsat 1
images, acquired on January 25, 1989, January 26, 1992 and January 16, 2000, were employed in the land use classification analysis. IKONOS, the first high-resolution commercial satellite, can take multi-spectral images of visible and near infrared regions with ground resolution of 4m. The ground resolution of its panchromatic band is 1m. The pansharpened IKONOS image with ground resolution of 1m, which was made by combining the multi-spectral and panchromatic images, was used in this study for Metro Manila. The acquisition date of the IKONOS image is September 28, 2001. The GIS data for Metro Manila, produced in 1986 based on aerial photographs for land use and building information, were also used in the study. Parameters Used for Land-Cover Classification Land cover classification was carried out by using NDVI (Normalized Differential Vegetation Index 4) ) defined by NIR R NDVI = (1) NIR + R where R and NIR represent the digital numbers of band 3 (visible) and band 4 (near infrared) of Landsat-TM, respectively. Vegetated areas yield high values in NDVI because they have relatively high near-ir reflectance and low visible reflectance. To remove the influence from the sunlight and atmosphere, the modification of digital numbers was carried out using the mean value and standard deviation of each image. After this correction, the NDVI was calculated for the three Landsat images. The urban areas in Metro Manila were classified by the level slice method 4) using two indices, NDVI and the texture for the uniformity of digital numbers in a local area. The angular second moment 4), Ta in equation (2) derived from a co-occurrence matrix 4), was used as the texture for the uniformity. m 1 m 1 { P( k, l) } Ta = (2) m= 0 m= 0 2 An co-occurrence probability 4) P(k, l) means the probability that a pixel value l appears in a relative position δ=(r, θ) from the reference pixel value k, where r and θ of δ are the relative distance and direction from the reference pixel, respectively. This matrix is called the cooccurrence matrix, because the column k and row l in the matrix represents the co-occurrence probability of the pixel values. m is a grade of an object image used in calculating the matrix. Before the matrix was calculated, the object image was converted to 4-bits. Therefore, m is equal to 16. The angular second moment, Ta, was calculated for the condition of r=1, which indicates neighboring eight pixels around a reference pixel and four directions of 0 or 180, 45 or 225, 90 or 270, and 135 or 315 degrees. The maximum value for the directions was defined as the representative value of the texture. A window size used for the texture analysis on the uniformity in a local area was 15x15 pixels for the Landsat image and 51x51 pixels for the IKONOS image. If the texture in a local area is uniform, the angular second moment has a relatively large value. Macroscopic Land Use Classification using Landsat Images Using the Landsat-TM images of Metro Manila for the three time instants, the distribution of NDVI was calculated as shown in Figure 1. In the figure, the areas with NDVI larger than 0.06 is considered as vegetation and equal to or less than 0.06 is considered as urbanized areas. A rapid expansion of urbanized area is clearly observed in this 11-year period. It is noticed that along the Valley (Marikina) Fault system, rapid urbanization is dominant. In this area, the farmlands have been converted to the residential areas. Considering the very short distance to the fault system, the seismic risk of the newly developed areas is considered to be high. It is observed from this figure that the total area of urbanization in Metro Manila is almost equal to that of vegetation in 2000. It 2
is expected that the urban land-use will top vegetation very soon in Metro Manila considering the recent rapid urbanization. As the next step of macroscopic analysis using Landsat data, the urbanization areas were further classified into congested (densely built-up) areas and non-congested areas. Using all the seven bands of Landsat data and employing six land cover classes (congested, non-congested, vegetation, bare ground, water, cloud cover), the maximum likelihood classification 4) was carried out and the result is shown in Figure 2. In the figure, red pixels represent the highly built-up areas. It is considered that seismic vulnerability of urbanized areas is dependent of the density of the areas, and that high-density areas might be the most important target of disaster mitigation planning. Microscopic Urban Classification using IKONOS Image Using IKONOS data, the classification of built environment in Old Manila was carried out. Each training data of 4,000 pixels was randomly selected from the typical area designated as an inscribed circle, such as a dense area with low-rise buildings, an upscale residential area, an area with midheight buildings (such as a southern part of Chinatown), and a highly urbanized area with high-rise buildings, and so on. The edge intensity was derived from the Prewitt filter 4), after the image was fabricated from the method to obtain the brightness signal for NTSC, which is one of the image transmitting systems used for television. Then, the cumulative relative frequency of the edge intensity for the dense area with low-rise small buildings was converted to 4-bits data, by means of dividing an accumulated ratio into 16 equal parts. The angular second moment, Ta, was calculated from the 4-bits edge intensity. As a result, it is found that a high-density area with low-rise buildings has small Ta value, representing the degree of built-up density. The classification category for the IKONOS image was determined as shown in Figure 3 (a) based on a scatter diagram for NDVI and Ta for training data. Then the level slice classification was carried out and the result is shown in Figure 4. It is seen in the figure that many high-density areas with low-rise buildings exist in the east and southeast of Old Manila, such as Pandacan and San Andres. Commercial areas with moderate size buildings locate in the south of Chinatown. The houses in upscale residential areas are surrounded by vegetation. On the whole, the result is relatively in good agreement with the actual built environment, except for the areas affected by clouds. In particular, there are very dense areas with low-rise buildings in the southeast of Old Manila, such as Tondo. In order to get the complete result of the classification for the entire Metro Manila, the latest Landsat data was also employed for the urban classification. First, the principal component analysis 4) for six bands excluding the thermal infrared band was performed. Then, an image of the first principal component (converted to 4-bits) was used to calculate the co-occurrence matrix and Ta. For the level slice classification, the category shown in Figure 3 (b) was determined based on scatter diagram of all the training data selected by 20x20 pixels. The Ta values for the densely built-up areas are found to be large because the texture in the 4-bits image of the first principal component is relatively uniform. Upscale residential areas and highly urbanized areas have small Ta values. The Ta values for commercial areas such as Chinatown have an intermediate value between that for dense low-rise building areas and that of high-rise building areas. The result of the urban classification using the Landsat image is shown in Figure 5. Very dense areas with low-rise buildings are mainly found along Pasig River, such as Old Manila, Caloocan City etc. The result for upscale residential areas and highly urbanized areas such as Makati and Mandaluyong, seems to be relatively in good agreement with the actual situation. Comparing Figure 3 (a) and (b), the value of angular second moment representing the texture changes with the spatial resolution of satellite images. The IKONOS image with 1 meter resolution can capture small texture thus high density areas have small Ta values while Landsat image with 30 meter resolution cannot capture this small texture thus the high density areas have large Ta values. Development of Digital Tools for Remote Sensing and GIS Analysis of remote sensing imagery requires the knowledge on remote sensing and some skills of using remote sensing software, e.g. ENVI. Even you can use such software, the software will not 3
provide you intermediate analysis results in a visible form. The results of remote sensing data processing must often be superposed on existing GIS maps and other spatial data. However, usual remote sensing software is not very convenient to do such tasks. Hence in this EQTAP project, a new digital tool, handling both remote sensing images and GIS data consistently, is being developed. This RS/GIS tool can be considered as a part of EQTAP Tool Box, which will available to EQTAP members through Internet in the near future. Figure 6 shows the overlapped image of the GIS information (land-use) with the Landsat imagery of 1989. By overlapping the satellite image with the GIS data, users can recognize changes by visual inspection in macroscopic manner. By this RS/GIS tool, the users can also obtain spatial information from an existing GIS database. Figure 7 shows the distribution of building story in Chinatown area of Metro Manila. As the current stage of Manila case study, we have analyzed the land use and building density distribution using the Landsat and IKONOS images. It is important to verify the results of classification. As a next step, a correlation analysis will be carried out between the classification results and GIS data on the density and structure of buildings in Metro Manila, such as shown Figure 8. The building inventory thus developed will be used for seismic vulnerability assessment of Metro Manila. Conclusions The macroscopic and microscopic land use classifications were attempted for understanding the urban structure of Metro Manila. According to the macroscopic analysis using Landsat images of three time periods, it is found out that Metro Manila has sprawled out in a high speed during the period between 1989 and 2000. It is noted that new developments are seen along the high seismic risk zone along the Valley (Marikina) Fault system. The land cover classification in a smaller scale was further conducted using IKONOS and Landsat images. Based on this analysis, the built environment in Old Manila was classified into four categories: very densely built-up areas, upscale residential areas with vegetation, areas with moderate size buildings, and highly urbanized areas with high-rise buildings. Relatively good agreement with the actual condition was observed, comparing the result with GIS data and field observations. An application tool to handle both GIS data and remote sensing data was developed for its use in risk management studies. Although this study for Metro Manila is still preliminary, some conclusions on possible applications and limitation of satellite remote sensing will be drawn within EQTAP Phase II. References 1) Ohmachi, T. and Roman, R. E.: Metro Manila, in search of a sustainable future, University of the Philippines Press, 388p, 2002. 2) Homepage of NASA for Landsat: http://landsat.gsfc.nasa.gov/ 3) Homepage of Spaceimaging: http://www.spaceimaging.com/ 4) Takagi, M. and Shimoda, H.: Handbook of image analysis, University of Tokyo Press, 775p, 1991 (in Japanese). 4
Figure 1. Expansion of urban areas identified by Landsat images in 1989, 1992 and 2000 using NDVI. Figure 2. Distribution of densely built-up areas (red) for Metro Manila estimated from Landsat images in 1989, 1992 and 2000. 5
Angular Second Moment (51x51 0.020 0.015 0.010 Non-Categorized Area Buildings with Large Size Mid-Height Buildings Highly Dense (Low-Rise Buildings) Very Highly Dense (Low-Rise Buildings) 0.005-0.5 0.0 0.5 1.0 NDVI Vegetation (a) Category for IKONOS image Angular Second Moment (15x15 0.5 0.4 0.3 0.2 0.1 Non-Categorized Area 0.0-0.6-0.4-0.2 0.0 0.2 0.4 0.6 Highly Urbanized Area Highly Dense Area with Low-Rise Buildings Dense Area with Low-Rise Buildings Mid-Height Buildings Upscale Residential Area NDVI Others (including Vegetation) (b) Category for Landsat image Figure 3. Urban classification in terms of NDVI and texture used for level slice method Figure 4. Result of microscopic classification in Old Manila using pansharpened IKONOS image. Orange and yellow: dense areas with low-rise buildings, light blue: mid-height buildings (mainly commercial areas); blue: areas with large-size buildings and others; green: vegetation area. 6
Figure 5. Result of urban classification for Metro Manila using Landsat image; deep and light orange: dense areas with low-rise buildings; yellow: low-rise and mid-height buildings (mixture of commercial and residential areas); blue: highly urbanized areas with high-rise buildings; light blue: upscale residential area; green: others including vegetation area. Figure 6. GIS information (land-use) overlapped with the Landsat imagery (1989). 7
Figure 7. Number of stories of buildings in Chinatown area, Old Manila. Figure 8. Areas to compare the classification results with the GIS data 8