The SSM/I image was acquired at 03:13 on February 4, 1996 (96035).
The AVHRR overpass schedule was approximately:
Coregistration of SSM/I and AVHRR is inherently difficult due to the large dimensions of the SSM/I pixels (25 and 12.5 km vrs 1.1 km for AVHRR). Coregistration was achieved here by registering each image to
There are two basic approaches to coregistration. The first would be to register each image using pixel latitude/longitude coordinates predicted by satellite ephemeris data. The second is coregistration using ground control points matched between the two images. Method 1 was tried with limited success. The prediction of cell centers based on orbital ephemeris data is not accurate enough for this application.
Eight coastal ground control points were selected on a three band (85v, 85h, 37h) SSM/I image and matched with corresponding points on the AVHRR image. Lake Superior is somewhat ideal with respect to SSM/I registration in that it has identifiable coastline features such as bays, peninulas and islands. Even so, only eight points were located with enough certainty to warrent inclusion in the warping procedure.
A third degree polynomial was used in the warping, and a nearest neighbor algorithm to assign Tbs to the new pixels. Coregistration accuracy was accessed by displaying various multiband images containing both data sets. Toggling each image plane on and off is useful, as well as the ENVI function, 'Dynamic Overlay', a sophisticated version of image plane toggling.
The SSMI data were warped to fit the map-registered AVHRR image in an image-to-image registration. In the process, a nearest neighbor algorithm was applied to assign DN to the newly created pixels. The following images were generated to evaluate the spectral distortion (if any) produced by the nearest neighbor resampling.
It looks like the original Tbs are adequately represented in the warped image. This subjective assesment is based on comparing the two Tb profiles, in part by animating the jpgs using xanim. The 'stair step' profile in the warped image is a manifistation of the expansion of the number of pixels from 58x38 to 402x333. The images display a checkerboard appearance for the same reason.
It would worth testing to see if resizing the SSMI image prior to
registration reduces spectral resampling distortion.
unwarped_profile1_location_image
unwarped_tb_profile1
warped_profile1_location_image
warped_tb_profile1
A single SSM/I image (f10_96035_04A) was registered to an AVHRR image (same date, 4 Feb. 1996). Each image has latitude/longitude for each pixel, calculated from satellite ephemeris. Ground control points (gcp's) is selected by identifying a pixel on the SSM/I image (85 GHz, H pol) and writing down the latitude/longitude for the pixel (e.g., 47.1600, -89.4700). Then, the equivalent pixel is located in the AVHRR image by moving the cursor until the AVHRR latitude/longitude best matches the SSM/I lat/lon (e.g., 47.5851, 89.2112) (the AVHRR lat/lon have 2 more digits than the SSM/I). The SSM/I point is selected first because of the finer spatial resolution of theAVHRR data.
The registration software used here (ENVI) provides three warping procedures: RST ('rotation, scaling and translation'), polynomial and Delaunay Triangulation. Each was tried; polynomial warping is favored and used in this analysis because of its relatively straightforward nature. In general, nearest neighbor resampling was used to assign digital numbers to the new pixels for images used in mathematical/statistical analysis, and bilinear or cubic convolution resampling was used to generate smoother images for visual analysis.
Various degrees of polynomial warp were tried. First through fourth produced the best results. A sixth order warp introduced artifacts in the data, a doubling of many rows in the middle of the warped image. Data plots (x plot) (y plot) of AVHRR versus SSM/I gcps reveal them to be linearly related, suggesting that a first order polynomial may provide an appropriate warping function.
