Use of Landsat Images for the Identification of Damage
Due to the 1999 Kocaeli, Turkey Earthquake
Identification of the Subsidence
Remote sensing provides a straightforward means to map the extent of water bodies, to inventory an area occupied by open water, and to monitor changes in water bodies over time. Likewise, comparisons of shoreline positions before and after flooding, permits measurement of areas flooded, as well as determination of locations of flooded areas. Such information can be difficult to acquire by conventional means. Determination of the land-water body is usually easiest in the near-infrared region, where land, especially if vegetated, is bright and open water dark. In order to detect the sunken areas we have matched the histogram of the near-infrared band (band 4). By matching the histograms of the two images (pre- and post-event) we will obtain an apparent distribution of brightness as close as possible and will minimize the brightness value variations. Since the reflectance value of the water in the near infrared band is low, we match the histograms in the range 1 to 20 and obtain the images in Fig. 6. Comparing these two images, it can be seen that in the image of the post-event (Fig. 6 (b)) there are some areas that differ from those in the pre-event image. These differences represent the inundated area. Also it is possible to calculate the extension of the flooding, by counting the number of pixels that differ between these two images. In order to count the number of pixels that belong
to water body we have used the 2-D scatter diagram. In this diagram we plot the distribution of reflectance of two bands, in this case band 4 vs. band 5. We define threshold values for the reflectance of the water, based on the values of training areas of water. Finally with the threshold values we make the selection in the 2-D scatter diagram. Finally we count the number of pixels that fall into this 2-D range. The number of pixels that belong to water: Before: 8430 pixels, after: 9413 pixels, difference = 983 pixels. Total affected area (flooding) = 983 x 30 x 30 = 884,700 m 2 , that correspond to red areas in the Figure 6 (b).
Figure 5. Histogram of the near infrared band.
Figure 6. The coastline in Gölcük, after matching the histograms. The intense red in (b)
shows the sunken areas.
Identification of Different Levels of Damage
Comparison in the Visible Range
As a first comparison for the detection and localization of damage we have conducted the comparison of the data in the visible range. For Landsat/TM satellite images bands 1, 2 and 3 cover the visible range. First we calculate de average of these first three bands by the formula:
where: BBk(i, j) is the digital number of the pixel (i, j) of the band k of the pre-event image. BRavg(i, j) is the averaged value of the pixel (i, j) of the pre-event image. This same formula is also applied to the post-event image. Then, we calculate the ratio between these two averaged images by Equation 2, where AR represents the averaged image after the event and R the ratio. Figure 7 shows the image obtained from this equation.
In the center of this image, the damaged area appears brighter and also the sunken area can be distinguished (in dark tones). This image is then compared with the ground truth data (AIJ, 1999) shown in Figure 8. As a result of this comparison we obtain the distribution of DN for the different levels of damage. Figure 9 shows 4 different sections: the sunken area (blue), low level damaged (green and turquoise) the section for level of damage 12.5% - 25% (yellow) and the section for high level of damage (read and orange).
Figure 9. Cumulative probability of the ratio (after/before) of the averaged images.
Principal Component Analysis
The Principal Component Transformation is designed to reduce redundancy in multispectral data. The purpose is to compress all of the information contained in an original n-band data set into fewer n new bands or components. Let L={L1(DKN), L2(DKN)} be a combined set of 2n bands images composed of two different dates images and CL be the variance-covariance matrix of L . Each principal component Xj is expressed as:
where {k, â k} is the normalized eigenvectors of the variance-covariance matrix CL of L. Difference of the surface reflectance between two dates is evaluated by the following principal component:
In this case, by joining the 6 non-thermal bands of the pre- and post-event images we obtain the 2n set. In this study the 3 rd principal component satisfies the condition given by Formula 4. Where the first 6 coefficients are positives and the lasts 6 are negative. It means, this principal component conveys the changes between the pre- and post-event images for this multitemporal set of images. The digital number of the pixels for the 3rd principal component image also follows the same distribution as shown in Figure 9. Based in the distribution of the digital number and using the maximum likelihood classifier we obtain the spatial distribution of the damage for Gölcük as is shown in Figure 10. This figure shows the distribution of the different levels of damage from Level 1 (no damage) up to Level 5 (high level of damage) and the sunken area. Same procedures have been conducted for Adapazari city, however, in this case the classification has been performed in only 3 levels of damage (Figure 11): no damage (blue), extensive (green) and catastrophic (red).
Conclusions
Damage detection due to The 1999 Kocaeli, Turkey earthquake was carried out using Landsat/TM images. Fire and sunken areas were successfully identified using the thermal band and the near infrared band, respectively. Different levels of building damage were recognized properly to some extent. It can be seen that the value of the digital number for the affected area increases with the increasing of damage level. This is due to the spread of debris after the earthquake. The debris has a higher reflectance in the visible region than the buildings in good condition. After using Principal Component Analysis we have found that in case of Landsat/TM images, principal component 3 conveys the changes in the multitemporal 12 non-thermal bands image.
References
- AIJ Reconnaissance Team, Kabeyasawa, T. et. al., 1999. Progress Report on Damage Investigation after Kocaeli Earthquake by Architectural Institute of Japan. Proc. ITU-IAHS International Conference on the Kocaeli Earthquake 17 August 1999.
- Campbell, J. (1996). Introduction to Remote Sensing Second Edition. The Guilford Press USA.
- Matsuoka, M. and Yamazaki, F., 1998. Characteristics of Satellite Images of Damaged Areas due to the 1995 Kobe Earthquake. 2nd Conference on the Applications of Remote Sensing and GIS for Disaster Management.
- Richards, J. and Jia, X., 1999. Remote Sensing Digital Image Analysis An Introduction Third Edition. Springer. Germany.
- Mather, P., 1999. Computer Processing of Remotely-Sensed Images An Introduction Second Edition. John Wiley & Sons. England.
- Yasuoka, Y., 1990. Digital Image Processing for Landcover Change Detection in Environmental Monitoring. National Institute for Environmental Studies. Japan.
