International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
156 
Geometric Correction Technique for Landsat Images 
Quarter Scene (South-East) 
Full Scene 
Ground 
Control 
1 st Order 
Polynomial 
2nd Order 
Polynomial 
neglecting 
DEM 
neglecting DEM and 
earth curvature 
with DEM and 
earth curvature 
constant 
height: 40m 
with 
DEM 
without 
DEM 
RMS at 
Control Points 
0.4704 
0.3265 
0.4016 
0.4016 
0.4016 
0.4016 
0.5820 
0.5723 
RMS at 
Check Points 
0.3466 
0.3410 
0.3952 
0.3952 
0.3952 
0.3952 
0.4889 
0.4837 
Table 2. Geocoding RMS errors for one Landsat image. 
4. GEOMETRIC PREPROCESSING 
To use remotely-sensed imagery and their classification 
results in GIS, these images have to be geometrically 
transformed to a reference coordinate system. Using the 
polynomial correction techniques, an image can be registered 
to a map coordinate system allowing its pixels to be 
addressed in terms of map coordinates rather than pixel and 
line numbers (Richards, 1994). Many applications of remote 
sensing image data require more than one scene of the same 
geographic area, acquired at different dates, to be processed 
together. Such a situation arises when, as in most monitoring 
projects, changes are of interest, in which case registered 
images allow a pixel by pixel comparison to be made. There 
are two ways to register two images to each other. Two 
images can be registered separately to a common map 
coordinate system. Alternatively, one image can be chosen as 
a master image to which the other, known as the slave, is to 
be registered. We chose the first method for our study. In our 
case, the reference coordinate system was the one of ATKIS 
(Gauss-Kriiger coordinate system) covering the complete area 
of our interest. The images were geocoded using 2nd degree 
polynomial procedures with a nearest neighbour resampling 
technique. 
The maximum residual error was about 0.5 pixel (15 meters) 
for Landsat images. Table 2 depicts some approaches for 
geocoding. It shows that in our test area, where the total 
difference in height is very low (150 m max.; average 5-40 
m), a 2nd order polynomial approach for geocoding a 
Landsat TM scene provides the best results. In the best case, 
we are using 30 to 40 control points for the computation of 
the transformation matrix. It was found that a digital 
elevation model could not lead to better results. The latter 
coincides with reports of Bahr and Vogtle (1991). They 
pointed out that the theoretical height error (h) is given by 
the following equation: 
where: 
hg = flying altitude 
P = pixel size 
s/2 = swath width 
In our case: Landsat TM: 705km * 30m / 90.25km = 234m, 
i.e., if the difference of terrain heights is more than 234m, we 
introduce an error of approximately 1 pixel, if we don’t use 
an elevation model. 
The assessment of residual errors was made by overlaying 
geometrically correct ATKIS datasets and by measuring 
check points in both data layers. If the registration error is 
greater than 1 pixel, this may result in the identification of 
spurious areas of change between the multitemporal datasets 
(Jensen, 1986). 
5. THE AUTOMATED APPROACH: DESIGNING A 
LANDUSE CLASSIFICATION SYSTEM 
5.1. Methodology 
In Anderson et al. (1973), the authors stated that, there is no 
ideal classification of landuse, and it is unlikely that one will 
be ever developed. Each landuse classification is made to suit 
the needs of a specific user, and few users will be satisfied 
with an inventory that does not meet most of these needs. In 
attempting to develop an operational automated classification 
system for use with remote sensing techniques, certain 
guidelines of criteria must first be addressed. In a monitoring 
approach, there are many repeated tasks. Most of them are 
very time consuming and therefore call for automated 
processing means. Figure 2 gives an overview of a common 
classification scheme. 
The steps of radiometric preprocessing, like atmospheric 
correction, are not content of this paper. More information 
about image processing can be found in Richards (1994), 
Lillesand and Kiefer (1987) and Campbell (1996). The next 
task, training data selection for a classification, is the most 
time consuming process and demands a lot of user expertise. 
For a single application, this workflow is a robust procedure. 
But disadvantages occur, if the user is willing to perform 
change detection analysis over a time frame of interest. In 
this case, normally all processing steps have to be repeated. 
Figure 3 gives an overview of a possible modem workflow of 
training data collection.