First it is necessary to establish the approximate layout of the four images. They may be previewed with ImageWorks if available, or GCPWorks. On a piece of paper draw the approximate location and orientation of each image relative to the others. This forms a mockup of what the output image will look like.
Based on this picture and the sizes of the input files, it is possible to determine the required size of the merged image file to hold the resulting mosaic. Assuming the four TM scenes form the four corners of the mosaic, the resulting mosaic might be about 12000 pixels by 12000 lines. This may be adjusted if it is desirable to resample the image to a larger or smaller pixel size. This final size should also be adjusted if only a subarea of the mosaic is desired.
The output mosaic file may then be created using the ``New'' option of the ``File'' menu in GCPWorks. Now the four input scenes may be registered into the final file one at a time, though all can be registered in one GCPWorks session. The Geocoded Image steps will be used with Full Processing and Polynomial model. The empty output file will serve as the georeferenced database. Because the output file is not geocoded, you CANNOT use ``Mosaic Only''; Mosaic Only mosaics geocoded images.
If possible, it is best to copy the most ``central'' image into the output file first. The amount of distortion introduced will generally be proportional to the distance from the ``anchor'' image.
The first image may be copied in by selecting one GCP on the input image, and one GCP in the output file. This will result in a simple translation without scaling when copying to the output file. If it is necessary to rotate the first scene while copying, or if it is desired to alter the resolution of the image during the mosaic, then it will be necessary to select three points to perform rotation and scaling. Since there is nothing to visually tie down to, it may be best to type pixel/lines coordinates in the input and output files to ensure minimal undesired distortion or scaling.
The initial copy should be done with first order polynomials, and no cut line or colour balancing.
The remaining three images should be mosaicked into the output file by tying them down to the existing imagery in the output file. In order to mosaic the images successfully it is important that there be a large degree of image overlap. This overlap allows a reasonable distribution of GCPs to be collected. If there is not a great deal of overlap, it is probably best to use only a first or second order model. High order models multiply any errors greatly in areas with few or no GCPs.
As each additional scene is registered into the output file, a cut line should be created in order to control the portion of the overlap region being replaced with new data. The existence of a cut line also enables the use of image blending, or feathering which can help hide image seamlines. It is also important to set the background value to zero, to avoid copying any missing data areas from the TM scenes over valid data already in the master file.
It is likely that colour balancing, and cut line blending will be necessary to merge the scenes so that they form a seamless final result.