[Proj] old german topo

Maciek Sieczka werchowyna at pf.pl
Mon Jul 26 07:54:49 EDT 2004

From: "John Smart" <john.smart at caris.com>

> Maciek
> Please see below...

> > I tried it and appears allright. I registered the Mestichblatter map
> > using Grass 5.3 into the projection above with all the grid crossings
> > (121 points), exported into tiff/tfw, reprojected into the Polish
> > "System 92" (tmerc on GRS80) using gdalwarp and imported it into
> > Grass mapset with a "System 92" 1:10000 topo.When the
> > Mestichblatter is displayed over the "System 92" map it looks quite
> > in sync but there are differences of about 3-15m in the location of even
> > churches and few other buildings which I don't suppose to have
> > moved. I can't see a clear tendency in these differences, they are in
> > different directions. Do you think the projection used may not be the
> > proper one? Or maybe I expect too much? I mean the map was
> > prepared in '30s and it's only 1:25000 so it can't be any more
> > accurate propably (?).

> I think you have to check the results step by step. The projection I
> suggested may or may not be the right one!

> First, I am curious as to how you did the registration. 121 points you
> say, but how was the image transformed? Although I'm not familiar with
> Grass, one typical approach would be to have a best-fit to those control
> points, in which case you don't expect the image of any grid intersection
> to match exactly to its nominal position (since error is distributed). A
> second approach would be to use a rubber-sheeting transformation,
> where the images of your grid intersections will be mapped exactly to
> the nominal positions, the areas between grid intersections being
> stretched or shrunk. Usually, I prefer the first approach, with an affine
> transformation, especially if I think I know the projection.

I think I used the best method available in Grass - i.rectify with "3d order
polynomial transformation matrix". With those 121 points the RMS was 6.24.
Is there any better way? Or maybe some other software? There is no such
thing like rubber-sheeting in Grass AFAIK.

> Having done this registration, how good is it? You need to ask at this
> point, because if you are getting significant differences at this stage,
> then clearly you won't get any better after you reproject and do the datum
> transformation.

> Suggestions: point (in Grass or whatever) to a grid intersection that's
> shown on the map; now what is the typical size of difference that you get
> between the image of the intersection, and the nominal position? Do the
> errors tend to be in any particular direction (i.e. systematic error), or
> are the errors random? Also do the same test on some of your map's edge
> ticks which show the nominal geographic positions. (If you are getting
> different results between the grid intersections and the geographic values
> then that suggests the projection is not what we are assuming).

So I did.

Top left corner on the Mestichblatter was labeled 15:40N 50:48E. After
registration into the following:

DHDN, Gauss Krueger, Bessel 1841, Potsdam datum (Federal States middle part,
50°20' N ... 52°20'), Zone 5
(+proj=tmerc +ellps=bessel +lat_0=0 +lon_0=15 +k=1 +x_0=5500000 +y_0=0

for this point I got:

EAST:             NORTH:               LON:               LAT:
5546998.675       5629470.325     15:40:00.268758E     50:47:59.986324N

Top right corner: 15:50N 50:48E.
EAST:             NORTH:               LON:               LAT:
5558739.475       5629589.15   15:49:59.890461E   50:47:59.977131N

Looks quite allright...

Few grid crosses within the registered map:
expected:  5548000E 5629000N
measured: 5548002.64166667 5628998.49375

5550000E 5627000N
5549999.3 5626998.76875

5555000E 5623000N
5554998.753125 5622998.6765625

5558000E 5620000N
5557999.97916667 5620000.875

> For the registration quality, my guess is that you would be doing well to
> have most errors within 0.2mm at map scale, i.e. within 5m  on the ground
> for your map.

The difference stays within +/-2m in general on the registered
Mestichblatter as to the grid crossings. But please take a look at the
pictures which I sent directly to your email as they are a bit to big for
sending them here. There is the registered Mestichblatter reprojected to
(with "gdalwarp -order 3 -rcs") and overlayed (violet or blue) on a pretty
new, GRS80 1:10000 Polish topo, with pixel size set to 1m. I'm 100% sure the
backdrop map is not disturbed anyway because it was scanned and registered
by a proffesional geodetic company. And also I have verified it's accuracy
on my own. So, when I compare the location of churches or other buildings
the error is bigger than for the grid crosses after the registration only -
about 15m. And it's not a simple shift in one direction - the error is
distributed in many directions. On the opposite some features are projected
a little better - like the unchanged fragments of the borderline (former
Czech/Germany, now Czech/Poland). Reasumming it seems the registration
cannot be done any better and this is a result of the quality of the German
1:25000 topo itself, not of the registration and reprojection process,
right? I've forgotten to say that the area is very mountainous wich is
propably a factor influencing the Mestichblatter's accuracy. Though I had
a doubt once I'd anyway say now that the projection you suggested would be
ok here. How do you think?

> With the map-drawing process, printing and material aging, results
> can't be expected to be too much better than that. The scan resolution may
> also plays a part: I take it your scanned map has finer resolution than
> the .png you sent out the other day?

Yes, that was about 3x reduced.


More information about the Proj mailing list