Geometric Registration Method For 10-Day Composite Avhrr Data For Asian Region
6. Evaluation Of Errors In Registration Accuracy
It is reasonable to measure geometric correction quality for all over image by RMSE obtained from the GCPs different from those used to determine the mapping functions. Therefore we used the estimation approach in which the accuracy of the geometric registration method was evaluated by calculation RMSE and mean deviations of the GCMPs for samples and lines when a geometrically corrected image was overlaid on the reference overlay image. The reference overlay image was generated from the geographical database. For test, we used AVHRR HPRT data acquired from August 1 to August 10 of 1999 at Ulaanbaatar (Mongolia). In the first experiment, RMSE and mean deviations of pixels and lines were estimated from the image that is not corrected by terrain elevation and GCPs. The result is compared with the accuracy of the image corrected by terrain elevation and GCPs (Table 1). The RMSE is represented by
and mean deviation for pixels (X_Dev) and lines (Y_Dev) calculated by
where x
i0 and
y
i0 are pixel and line number of i-th GCMP detected in AVHRR imagery,
x
i1
y
i1and are corresponding pixel and line number in the map overlay image, n is the number of GCMPs.
In the second test, two-stage correction was carried out for the each single scene and daily mosaic images were produced. Then RMSE and mean deviations were estimated from the daily mosaic images (Table 1). Table 1 clearly showed that GCPs matching result and terrain correction give significant improvement in the first stage of geometric correction. However, it can be seen that accuracy of the first stage correction is not sufficient for compositing requirement. Comparison of two-stage correction result shows that fine correction allows getting images with much better accuracy than the images performed by only first-stage correction.
Table 1. RMSE comparison of the
correction steps for daily composite
images
| Day |
Syst. Corr. |
Stage-one |
Stage-two |
| 1 |
17.534 |
3.918 |
1.679 |
| 2 |
16.283 |
4.254 |
1.873 |
| 3 |
15.412 |
5.236 |
2.134 |
| 4 |
15.207 |
4.692 |
1.580 |
| 5 |
14.735 |
5.399 |
2.463 |
| 6 |
15.783 |
3.899 |
2.238 |
| 7 |
17.496 |
4.697 |
1.628 |
| 8 |
16.786 |
5.184 |
1.826 |
| 9 |
14.428 |
5.428 |
2.176 |
| 10 |
14.746 |
5.876 |
1.876 |
Table 2. RMSE comparison of the correction steps for 10 day composite image
| |
Syst. Corr. |
Stage-one |
Stage-two |
| GCP |
25 |
168 |
168 |
| X_DEV |
7.28 |
3.301 |
1.598 |
| Y_DEV |
12.16 |
2.863 |
0.963 |
| RMSE |
14.063 |
4.701 |
1.969 |
In the third experiment, 10-day composite image was produced from the daily mosaic images and the accuracy of the geometric registration was evaluated (Table 2). The composite image of 14400 pixels and 6000 lines covers area between 50.0 and 170.0 degree of longitude and 30.0 and 80.0 degree of latitude. The number of GCPs actually used for the evaluation was 166. The result shows that RMSE was 1.969. Mean deviations for sample and line were 1.598 and 0.963 respectively. The result shows that the correction methodology gives relatively good accuracy in the along-track direction, but the across-track error can not reach to one pixel accuracy. We consider two factors, which decrease RMSE and mean deviation value. The first is that the errors detected in off-nadir pixels of the image or in the end of scenes specially in North Polar region, were larger than other ones. The second factor is that the accuracy of the DCW and water mask data used for reference overlay image is still not high. Because it was visually detectable when the reference image is overlaid on the composite image. The third is that ordinary least square method was applied to determining mapping functions which can not take count local distortions.
Conclusions
A full automatic method of geometric registration used of AVHRR data has been presented. This approach based on two-stage correction algorithm has been used in the geometric registration of composite AVHRR data set for Asian region with satisfactory result. The increase of registration accuracy of AVHRR imagery has been achieved by two improvements. In the first, an accuracy of AVHRR image navigation using PaNDA package is improved by accurate identification of GCPs and terrain correction. In the second, AVHRR image navigation accuracy is estimated and a final correction is carried out by registration of the navigated images to fixed reference image. Realizing this algorithm, we developed the registration software in "C" and implemented it under UNIX operating system in HP workstation. This software has been operationally used for geometric registration of AVHRR data and it is applicable to pre-processed AVHRR data for fine geometric correction without any operator intervention.
The development of the composite data set has been shown that registration of AVHRR data for large area with subpixel accuracy is somewhat difficult in practical realization. The difficulties have arisen due to several problems discussed in previous sections: inaccuracies in navigation model and coast and shoreline data, also some errors in orbital parameters. Specially, these problems are very important to resulting accuracy when operational registration system has been used and they are long-term issues for the increase registration accuracy.
In spite of the existing difficulties, we consider some way to improve present accuracy in the future. The overlay of the coastline image on AVHRR imagery showed that there are non-uniform shifts across the whole navigated image. In this case the use of ordinary polynomial mapping function for image transformation determined by least-squares technique may not enough to get more higher accuracy because a local geometric distortions are equally averaged for all over the image. Therefore, utilization of mapping functions, which is sensitive to local distortions, is one of the possible ways to improve the overall registration accuracy.
References
-
Hashimoto, T. and Murai, C., 1993, geometric correction of NOAA AVHRR Imagery in accordance with the number of GCPs, Journal of the Japan Society of Photogrammetry and Remote Sensing, 13.
- PaNDA User Manual, 1998.
- Yokoyama, R., Lei, L., Purevdorj, Ts. and Tanba, S., 2000, AVHRR 10-day Mosaic Composite Image Data Set for Asian Region, Journal of the Japan Society of Photogrammetry and Remote Sensing, Vol. 39, No.1, 33-38
