3. Image Registration
In contrast to the commonly applied image-to-map registration, we perform the image-to-image registration to identify changes between two images. The reason of image-to-image registration is that the major earthquake significantly changes terrain features in many mountain areas, and selecting GCPs using maps becomes very difficult. The ordinary kriging approach of image-to-image registration is as follows.
- Select a set of GCPs using images acquired before and after the earthquake.
- Perform the first order polynomial trend mapping to map the two images.
- Consider the registration errors in E-W and N-S directions as two residual random fields and separately perform the anisotropic variogram modeling for each random field.
- Apply the ordinary kriging estimation to the residual random fields using anisotropic variograms established in step (3).
- Final result of the image-to-image registration is obtained by adding ordinary kriging estimates of the residual fields to the trend field estimated by PTM.
The ASM approach has the advantage of capturing local terrain variation and eliminating or reducing the distortion caused by terrain variation. Readers are referred to Cheng et al. (2000) for a complete description of the ASM approach for image registration.
An example of image-to-image registration by the ASM approach is shown in Figures 1 and 2.
4. Change Detection
Change detection in this study involves three steps: spectral (or band) ratioing, temporal image differencing, and determination of change-no change threshold. Since images acquired before and after the earthquake have different atmospheric conditions, band ratioing analysis is performed to effectively compensate for brightness variation caused by different atmospheric conditions. Figures 3 shows that brightness of the IR image before the earthquake is lower that that of the postearthquake image. After the IR/R band-ratioing, brightness values of the two images are very close. The second step of change detection is to determine the absolute difference between the two pre- and post-earthquake (IR/R) images. Figure 5 illustrates the difference image. By comparing histograms of the two (IR/R) band-ratio images, we determine the areal percentage of changes. Finally, the change-no change threshold is determined by locating the grey level, at which exceeding probability equals the areal percentage of landuse changes, from the histogram of the difference image. Figure 6 shows an example of detected changes in the study area. DTM data are used to generates a slope image of the study area, and landslide identification is restricted only in the area with slope steeper than 18o. Up to the present, many of the identified landslide areas have been verified to by field investigations.
5. Conclusions
In this study we demonstrate that the anisotropic spatial modeling approach of image-to-image registration yields high registration accuracy in mountainous and rugged terrain areas. The band-ratioing technique yields close brightness values for two SPOT images acquired prior to and after the Chi-Chi earthquake. The multi-temporal differencing technique yields a difference image using two (IR/R) images, and a threshold grey-level is determined for successful change detection.
