International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999
The main characteristics of the GEORIS geometric correction
model are:
• Global correction (using a classical polynomial function)
• Local correction
• Capability to use linear features
2.1. Global correction
A first global model is calculated using an empirical approach
based on tie points measured by the operator in the set of
images. The tie points are used to determine the coefficients of a
polynomial transformation (least squares minimisation). This
model allows performing the main geometric correction.
A polynomial function of order one handles similarity and
affine transformations (rotation, scale, translation). A higher
order polynomial permits modelling of more complex
distortions, e.g. due to attitude variations. However, the use of
order higher than two is dangerous; if the number of tie points
is limited and/or their spatial distribution poor, it usually leads
to wrong corrections.
The quality of the model is usually evaluated through the use of
the residuals at each tie point (error between the location of the
tie point and the location given by the model). When a tie point
has a larger residual than the other ones, it is usually re-visited
by the operator to check whether it is due to a pointing error.
However, in most of the cases, this error does not come from
the tie point measurement but from local deformation (for
instance, due to the relief). This type of problem can usually not
be solved using this classical approach. A larger set of tie points
will minimise the problem locally but will generally introduce a
larger error around this area. To limit this risk, especially
because the operator does not know the mathematical approach
used, control tie points are required. Then, residuals are
computed but the control tie points are not used in the least
squares estimation.
In the GEORIS system, there are in fact four categories of tie
points available:
• Control tie point: not considered in the minimisation
process
• Normal tie point: this is the usual case
• Questionable tie point: the user is not confident in the
location of the tie point (because of difficulties to point it
in the image with a good accuracy)
• Super tie point: when the user is very confident in the
location of the tie point
The last three categories of tie points are introduced in the
minimisation process using different weights. This approach has
a small overhead for the user but permits during this step the
analysis of residual errors and leads to higher accuracy.
At this step, a pixel of co-ordinates (Xj,Yj) is transformed in the
coordinates (x i ’,y i ? ) in the other image using a polynomial
function f .
-> (X i \y i ’)
For each tie point, the coordinates (xi, yi) in the other image are
known. This permits to compute residuals defined by Eq. 1:
dx I = x i -x i ’ (1)
dyi = Yi-yi’
2.2. Local correction
A specific characteristic of GEORIS is the use of a local model.
This model guarantees that each of the selected tie point,
measured by the operator, will be perfectly fitted by the model
(no positioning errors on selected tie points). This local
correction is applied taking into account the tie point
neighbourhood area, whose size depends of the surrounding tie
point context (i.e., road crossing, bridge over a river, shoreline,
etc.). Obviously, this local model permits to correct for local
deformation errors such as introduced by the terrain relief. If a
digital elevation model is also available, the local correction can
be automatically linked to a 3-D function (horizontal position
and altitude). In this case, the local model and correction are
achieved taking into account the third dimension.
When no DEM is available, the pixel coordinates (X^Y;) are
transformed to (x;”, y^ in the other image using successively a
polynomial f plus a function (called local correction):
Xi,Yj —*■» (xr,y^ — l i-> (xj-yfl
The function /. depends on the location of pixel X^Y,.
For a tie point Xj.Yj, it is a translation corresponding to the
residuals (dx^dyO, see Eq. 2.
Xi” = Xi’+ dxi (2)
yi”=yi’+dyi
For another point (X^YO of the image and its transformed
coordinates (x i ’,y i ’), the translation dx^dy, is defined by:
dx i =^Pk^ dx k / Pk (3)
k
<>»= Zp* Z^* 7 Pt
k
where:
k: number of tie points
p k : distance between (x i , ,y i ’) and the tie point (x k ,y k )
dx k : residual value (along x axis) of the tie point k
dy k : residual value (along y axis) of the tie point k
This means, that the translation depends on existing
surrounding tie points, each with a weight depending on its
distance from the currently processed tie point.
This approach has been shown to be interesting in many cases.
Table 1 shows the results of two typical cases where this
approach is especially useful. The first one corresponds to
results obtained at a mountainous area (using SPOT images
(Xi,Yj)
L