Starting the GOES-8 satellite, the GOES pointing stabilization went from a spin-stabilized system to a gyroscope stabilized system. Built in to the new gyroscope stabilized system is a feedback mechanism for correcting for pointing inaccuracies due to mainly to gravity anomalies and platform motion caused mechanical motion of systems on the platform. This feedback mechanism requires the uploading from Earth of pointing correction factors. NOAA had originally hoped that an automatic phase correlation approach implemented in the NOAA Product Monitor would be sufficient. Early experimentation, however, showed that the automatic phase correlation was not robust enough for routine use (primarily due to the confounding effects of even a minor amount of clouds). NOAA then had to used its fall-back option of "point" manual landmarking (described below) in order to provide the required correction factors.

This paper briefly describes and then compares the effectiveness of five image registration approaches for GOES visible band imagery. The techniques compared are (i) NOAA "point" manual landmarking, (ii) manual landmarking on extant feature (whole island, lake, etc.) - which is considered to be the "true" registration, and (iii) a modified version of the NOAA automatic phase correlation. Two new approaches are also described and compared to the earlier approaches: (i) automatic spatial correlation on edge features, and (ii) automatic spatial correlation of region boundaries derived from image segmentation.

Comparative results are given for varied sets of GOES visible imagery.

NOAA "point" manual landmarking: The NOAA operator displays a 128x128 pixel section of the GOES-8 imagery data on the Product Monitor and compares the image to a Landmark map. The operator finds the line and pixel location of a particular land/water feature (peninsula, bay, etc.) and records the difference between the expected line and pixel location and the line a pixel location actually found.

Manual landmarking on extant feature (whole island, lake, etc.): Performed manual registration of GOES-8 imagery data from a particular day and Landmark by flickering between successive images using the Khoros function "animate." This result was used as a baseline "best result" for comparison to the other techniques. This version of manual registration differs from the NOAA landmarking in that is uses spatial extant features (e.g. whole island, large section of coastline, whole lake, etc.).

Automatic phase correlation: Phase correlation is a mathematical technique that was developed to register images to one another. It works best in cases in which the misregistration is only a translation (which is the case for GOES data). The technique can be described as follows (from [1]):

Given a reference image, g_R, and a sensed image, g_S, with two-dimensional Fourier transforms G_R and G_S, respectively, the cross-power spectrum of the two images is defined as G_RG_S*, and the phase of that spectrum is defined as

The phase correlation function, d, is then given by

where F-1 denotes the inverse Fourier transform.

The spatial location of the peak value of the phase correlation function, d, corresponds to the translation misregistration between g_R and g_S.

The innovation in the implementation demonstrated here is in the finding of the peak of the correlation function, d. Instead to looking for an interpolated peak of d (as in the implementation installed - and not used - on the NOAA Product Monitor), the center of mass of the peak of d is found. We have found that this gives a more robust result than searching explicitly for the peak.

Automatic spatial correlation on edge features: Automatic edge detection is followed by spatial correlation (multiply reference image times the shifted input image). Automatic edge detection was performed using the "vdrf" function from Khoros - which is an implementation of an edge detection algorithm developed by Shen and Castan [2,3].

Automatic spatial correlation of region boundaries derived from image segmentation: Automatic image segmentation is followed by spatial correlation (multiply reference image times the shifted input image). Automatic image segmentation is performed by Iterative Parallel Region Growing developed by Tilton [4].

One problem was encountered when spatially correlating the reference image and input image. Mainly due to varying lighting conditions, the input image region outline obtained may be expanded or shrunk as compared to the reference image region outline. In tests reported here, this was compensated by slightly smearing the reference image region outline (region boundary pixels were given the value "3," while pixels touching the region boundary pixels were given the value "2," and the pixels touching the pixels touching the region boundary pixels were given the value "1." Results using both the unsmeared and smeared reference image region outlines are given.

Registration results were obtained for three widely different GOES 8 visible scenes, examples of which are given in Figures 1, 2 and 3. Complete results for the scene over Isla Ángel de la Guarda off of Baja California, Mexico (Figure 1) are given in Table 1, using the non-smeared reference image for the region boundary registration case. Summary results for all three images are given in Table 2. The NOAA point manual landmarking shows an overall bias shift versus the other techniques. This is probably due to a problem with the navigation program used to locate the image segments - it is probably not an indication of an overall bias in the NOAA point manual landmarking itself. The standard deviation results indicate that the spatial correlation on edge features approach was most consistent in matching the manual landmarking on extant features (assumed as reference). The results also show that "smearing" the reference edge map did improve the results from the spatial correlation on region boundaries approach for the BAJ case, but not enough to match the results produced by the spatial correlation on edge features approach.

[1] C. D. Kuglin, A. F. Blumenthal, J. J. Pearson, "Map-matching techniques for terminal guidance using Fourier phase information," SPIE Vol. 186 Digital Processing of Aerial Images, pp. 21-29, May 22-24, 1979.

[2] J. Shen and S. Castan, "An optimal linear operator for edge detection," Proceedings of CVPR '86, Miami Beach, FL, pp. 109-114, 1986.

[3] J. Shen and S. Castan, "An optimal linear operator for step edge detection," Graphical Models and Image Processing, Vol. 54, No. 2, pp. 112-133, 1992.

[4] J. C. Tilton, "A refinement of Iterative Parallel Region Growing and application to remotely sensed imagery data," to be submitted to IEEE Transactions on Geoscience and Remote Sensing.

Figure 1. GOES 8 visible image from 1815 UTC on June 12, 1996, from over Isla Ángel de la Guarda off of Baja California, Mexico. Selected as reference image.

Figure 2. GOES 8 visible image from 1915 UTC on June 13, 1996, from over Puerto Vallarta, Mexico. Selected as reference image.

Figure 3. GOES 8 visible image from 1445 UTC on June 14, 1996, from over Lake Okeechobee, Florida, U.S.A.. Selected as reference image.

Table 1. Shifts west and east from the manual registration result for Phase Correlation, Edge Feature Correlation, Region Boundary Correlation, and NOAA Point Manual Landmarking for the GOES 8 scene over Isla Ángel de la Guarda off of Baja California, Mexico (Figure 1).

Table 2. Mean and standard deviation of the shifts west and east from the manual registration result for Phase Correlation, Edge Feature Correlation, Region Boundary Correlation, and NOAA Point Manual Landmarking for all three GOES 8 image data sets. BAJ corresponds to the data from over Isla Ángel de la Guarda off of Baja California, Mexico (figure 1). PUV corresponds to the data from over Puerto Vallarta, Mexico (figure 2). OKE corresponds to the data from over Lake Okeechobee, Florida, U.S.A. (figure 3).