Merge images first, or georeference first?

Question

I'm wanting to georeference and vectorise some features from scans of the OS 1896 Town Plans that I have purchased from the NLS (http://maps.nls.uk/os/london-1890s/info.html). The scans provided are not georeferenced and when I try to crop the borders they are not quite square so will need rotating first. The problem is the features lie at the intersection of 4 separate sheets and I'm uncertain of the best / most accurate approach to take to merge them:

1) Treat each image separately: Rotate and crop each image in Gimp, georeference in QGIS, then "nudge" the individual GCP points to try get them to line up properly, then merge with gdal_merge?

2) Merge first: Rotate, crop and merge the images together in Gimp, then georeference the whole image in QGIS

3) Some magic automated solution I'm not aware of :-)

Thanks!

Answer 1

Well, QGIS is not that good in creating of continuous rasters if that's what you want. Nevertheless it all depends on precision you wish to have in the end. If you wish to trade precision for swiftness - merge them all in GIMP and then georeference resulting image. But be aware that every modification of the raster increases the errors. So if you want to have more precision - georeference each image first, then crop them inside QGIS (Raster-> Extraction->Clipper) if you need. I would just try to make borders transparent in layer properties if it is possible. There is actually no need in merging georeferenced images - if you need to treat them as one you may just create a virtual raster (Raster->Miscellaneous->Build Virtual raster).

Answer 2

Perhaps a personal opinion here, but when trying to georeference data, I tend to try and get some CGP's as far apart as possible, as with points that lie close to eachother, a small error in georeference can accumulate to fairly large errors on the more outlying points of your data.

That in mind, I'd merge the images first, then georeference on the outlying points to shoot for minimal error in the center parts.

Answer 3

I suggest to clip the data in GIMP if necessary, then georeference individually to vrt's.

If you use QGIS Georeferencer, use File -> Generate GDAL Script for that, and change the output to -of VRT.

The VRT file still contains your GCP in a editable file. If you are not satisfied with the precision at the borders, you can insert or move as many GCP as you wish. Thin Plate Spline is the best transformation method for that.