1. Tae Young Kim and Myung jin Choi
Image Registration for Cloudy KOMPSAT-2 Imagery Using Disparity Clustering
Tae Young Kim and Myung jin Choi
Korea Aerospace Research Institute
Korean Journal of Remote Sensing, Vol.25, No.3, 2009, pp.287~294

2. 1. Introduction
High-spatial, low-spectral resolution panchromatic(PAN) image
Low-spatial, high-spectral resolution multispectral(MS) image
Reasons that most satellites don’t collect MS image:
Incoming radiation energy to the sensor
data volume collected by the sensor

3. In order to obtain a high-resolution MS image from bundle(PAN+MS)images, pan-sharpening technique is needed.
The accuracy of image registration directly affects the quality of pan-sharpening.

4. It determines the best spatial fit between two or more images.
Image registration
It determines the best spatial fit between two or more images.
A lot of images registration techniques have been developed in remote sensing.
Four steps of images registration method
Feature detection
Feature matching
Transform model estimation
Image re-sampling and transformation

5. like IKONOS, QuickBird, and KOMPSAT-2 has the time and angle difference in an acquisition of the PAN and MS image because the satellite imaging system based on Charge Coupled Device(CCD) line sensors have the offset of the CCD combination in the focal plane.
It causes a displacement if a scene has high altitude objects or moving objects.

6. The purpose of this paper is to enhance the accuracy of image registration at the ground region around cloud boundary through preventing to select the matching points at cloud boundary.
We attempt to use the distance disparity.
On the distance disparity distribution, we can see that the group of cloud matching points is separated from that of ground matching points.

7. In order to classify several matching point groups on the distance disparity distribution, we utilize the unsupervised classification.
We calculate the ratio of cloud matching point in the group using enhanced cloud mask to decide cloud matching point groups from the classification result.
To validate our method, we used several KOMPSAT-2 cloudy images.

8. 2. Characteristic of KOMPSAT-2 Sensors
KOMPSAT-2 has the time and angle difference in an acquisition of the PAN and MS image.
Multi-Spectral Camera (MSC) on the KOMPSAT-2 has five CCD panels.

9. Different view direction we can find
Different view direction we can find. So a mismatch occurs in ground height (H1,H2).
If PAN sensor acquires images on t2 at P1 ground position, MS sensor acquires images on t2 at same position.
PAN and MS have different position of moving object.

10. 3. Methods Image Registration
Generally, an images registration method consists of the following steps,
Feature Detection
Feature matching
Transform model estimation
Image re-sampling and Transformation

11. Additional four steps below
We propose a new image registration method which removes cloud matching points.
Additional four steps below
Cloud detection
Make cloud mask
Classify cloud matching points
Discard cloud matching points

12. Cloud Detection & Make Cloud Mask
In classifying cloud matching points step, the proposed method uses a binary cloud mask which is made by automatic thresholding algorithm to decide cloud matching points groups.
We use Otsu algorithm which is simple, fast and effective to calculate threshold value automatically.

13. The result binary images of Otsu method. (Fig.6)
White parts covers cloud region.(a) However white region does not cover perfectly boundary of the cloud.(b)
Divide the images into blocks
Decide whether a block is cloud or not
Extend the cloud block to cover sufficiently the boundary of the cloud

14. Classify and Discard Cloud Matching Points
The proposed method executes the following five steps;
Calculate the distance disparity of the matching points.
Clustering the matching points
Measure the cloud matching points ratio of the clusters
Determine the cloud clusters
Discard cloud matching points

15. 4. Experimental Results We choose the cloudy KOMPSAT-2 scene.
The PAN and MS images was taken simultaneously.
Fig.8 shows the overlaid image which displays the MS1 band image, the cloud mask and matching points.

16. Fig.9-a is the result of the former method which doesn’t consider cloud objects. Although cloud objects have no misregistration, the distortion arises at ground regions nearby cloud objects because the cloud matching points affect the ground regions matching.
Fig.9-b is the proposed method result. There is no distortion at the ground regions because the proposed method uses only the ground matching points.

17. 5. Conclusion
In this paper, we propose a new approach for the image registration of cloudy high resolution image.
It shows how to discard the matching points of cloud objects which cause a mismatch.
The experimental results show the accuracy of the proposed method is high at ground region around cloud objects. In addition, the proposed method may be able to be applied to other high resolution satellite data.
We will concentrate on studying more suitable automatic thresholding technique and classification technique

18. Panchromatic (1m Pan)
Multispectral (4m MS)

19. Bundle (1m Pan & 4m MS separated)
Pan-sharpened (1m colour in 4 bands: R,G,B,NIR)