IMAGE RECTIFICATION IN A GIS ENVIRONMENT
T.Y. Shih*, E. Derenyi, D. Fraser
Department of Surveying Engineering
University of New Brunswick
Fredericton, N.B. Canada
ISPRS COMMISSION IV
ABSTRACT
The usual practice in geometric correction and registration of remotely sensed data is to fit the image onto a map base. In a
geographic information system (GIS) with both vector and raster data handling capability, a map to image registration is also
possible. This is the preferred approach when the two data types must only be in registration for the purpose of image analysis but a
permanent georeferenced image file is not required. Both the image-to-map and the map-to image transformation has been
implemented in a GIS at the University of New Brunswick. The design and implementation of this registration method is
discussed.
KEY WORDS: Geometric registration, Rectification, Resampling, Georeferencing, Transformation, GIS.
*Presently with the Department of Civil Engineering, National Chiao-Tung University, Taiwan, R.O.C.
1. INTRODUCTION
Geometric correction and registration are essential elements
of digital image processing and analysis. Without
georeferencing the exact spatial relationship of an image to
the earth's surface is unknown. Without known spatial
position of the information obtained from digital images has
limited value.
Georeferencing of any raster data involves two
transformations, a geometric and a radiometric. The
geometric transformation defines the coordinates of every
pixel in the new reference system. It is performed by one of
the well known two-dimensional coordinate transformation
functions such as the similarity, affine, projective or
polynomial transformation. Radiometric transformation,
which is usually referred to as resampling, means to
determine the radiance value of every pixel in the
transformed raster. It is an interpolation based on the
radiance values in the unregistered scene. The three most
frequently employed resampling techniques are the nearest
neighbour, the bilinear and the bicubic interpolation.
2. EFFECTS OF RESAMPLING
Radiometric transformation is a computationally demanding
task and is not without problems. The nearest neighbour
method is reasonably fast and the new pixel values are
copies of existing values from the input image. Thus the
radiometric characteristic of the image remains unchanged.
On the other hand, it does tend to produce a rather blocky,
disjointed appearance in the output image. Bilinear
interpolation results in a smoother looking image because it
is essentially an averaging process. This, however means
that sharp boundaries in the input image become somewhat
blurred in the output image. This technique introduces new
radiometric values and problems may be encountered in
subsequent spectral pattern recognition analyses of the data.
The computational time required is longer than for the
nearest neighbour method. The bicubic technique requires
the longest computation time, but it tends to give the most
natural-looking image. Again, some loss of high-frequency
information occurs, as this interpolation is essentially a low-
pass filter.
The following example illustrates the effect of resampling
673
Derenyi and Saleh, 1989]. Agricultural fields and water was
classified on a Landsat TM, Band 5 subscene of 342 lines by
313 pixels. First, the classification was performed on the
image in its original state. It was then repeated after a 9° and
35° rotation, followed by a nearest neighbour and bicubic
interpolation resampling. The 9" rotation approximates the
orbital inclination of earth observation satellites, while the
35° represents rotations which could occur when airborne
data are georeferenced.
Table 1 shows the mean and standard deviation obtained in
training areas for the two classes in the original and in each
of the four resampled images. Table 2 shows the pixel count
of each class after the maximum likelihood classification at
95% probability level of the five data sets.
The effect of resampling is demonstrated by the shift of the
mean value, the expanded standard deviation and by the
significant increase in the pixel count of both classes in the
bicubic interpolation resampled image.
3. MAP-TO-IMAGE REGISTRATION
The lengthy processing time and the undesirable effects on
the radiometric characteristic could be avoided if both the
image and the map resides in a GIS with an integrated raster-
image/vector-graphics handling capability. In this case, a
temporary digital map-to-image registration is possible. All
information extracted from the image can later be
georeferenced by an inverse transformation . This map-to-
image transformation scheme has been implemented in the
raster image extension (RIX) of the Computer Aided
Resource Information System (CARIS) GIS in two
versions: For small sub-scenes the registration can be
performed on-line interactively. For large data sets an off-
line batch processing mode is available. Here, various two
or three-dimensional transformation functions are provided
as options.
3.1 On-line Georeferencing
It is not uncommon that large data sets still exhibit noticeable
residual misregistrations in subregions, after
georeferenceing. This may happen when relief displacement
was ignored and when the number, distribution and accuracy
of the ground control points (GCPs) or the transformation
function used are inadequate for properly modelling the
