CEE 615: Digital Image Processing Topic 4: Geometric Correction 4-1

W. Philpot, Cornell University, January, 01

4. GEOMETRIC CORRECTION

Sources of geometric distortion:

sensor characteristics
- optical distortion
- aspect ratio
- non-linear mirror velocity
- detector geometry & scanning sequence

viewing geometry
- panoramic effect
- earth curvature

motions of the aircraft/satellite or target

- attitude changes (pitch, roll, yaw)
- position variations (altitude, slew)
- earth rotation

Distortions appear as:

- changes of scale over the image
- irregularities in the angular relationships among the image elements
- displacement of objects in an image
- occlusion of one image element by another

Correcting the distortions is often costly:

- computer & operator time
- affects spatial and radiometric resolution

The nature of the "correction" depends upon the ultimate use of the data:
- area measurements ==> equal area projection
- shape measurements ==>projection which preserves the angular relationships of the
scene
CEE 615: Digital Image Processing Topic 4: Geometric Correction 4-2
W. Philpot, Cornell University, January, 01

Sources of Distortion
A. Sensor characteristics
1. Optical distortion: not usually a problem. Most significant in imaging systems that use a
large portion of the lens: vidicons, photographic systems

(a) (b) (c)

a) planimetric map
b) negative optical distortion (barrel distortion)
c) positive optical distortion (pincushion distortion)

2. Aspect ratio: The aspect ratio is the ratio between scales in the vertical direction to that in
the horizontal. Since distances are often measured by counting pixels, the shape of
the pixel must be defined. A 10-pixel distance may mean one thing in the horizontal
direction and something else entirely in the vertical direction.

3. Non-linear mirror velocity: If the velocity of the scanning mirror is constant (the ideal
case) the sampling interval is well-defined and the size of the resolution cell, while
variable over the scanline, is easily calculated. In many scanning systems, the mirror
does not maintain a precise, constant angular velocity. This is notably the case with
Landsat MSS which uses an oscillating rather than a rotating mirror. Along-scan
distortions are introduced causing pixel compression or pixel stretching at various
points along the scan line. This distortion is difficult to correct since it is difficult to
measure and may change gradually with time.

4. Detector geometry & scanning sequence: Any multispectral scanner attempts to sample
several spectral bands simultaneously, and many sense several scan lines during one
sweep of the scanning mirror. Ideally, all bands fir a single pixel would sense exactly
the same point on the earth's surface and adjacent scan lines would line up exactly,
pixel for pixel. However, the physical displacement of detectors in the image plane
and the fact that only one detector can be sampled at a time introduce some
misregistration of different bands for the same pixels, and displacement of adjacent
scan lines.
CEE 615: Digital Image Processing Topic 4: Geometric Correction 4-3
W. Philpot, Cornell University, January, 01

B. Viewing Geometry
Panoramic effect: a scale distortion (an increase in the resolution cell size) for observations
away from nadir due to an increase in the area viewed by the sensor.
a. panoramic effect (flat-earth)

= observation angle
= IFOV
h = altitude
d = ground distance from nadir
to the sampling point

p = size of a pixel perpendicular

h sec to the observation angle.
p = projected length of a pixel
h along the ground

2
p = 2 h sec

p0
p 0 sec2
d

pc

p
Given a square pixel (at nadir):
p0 = 2h tan = 2h d = h tan
p = 2 h sec = p0 sec
p0 = p sec = p0 sec2
pc = p0 sec
CEE 615: Digital Image Processing Topic 4: Geometric Correction 4-4
W. Philpot, Cornell University, January, 01

b. panoramic effect (curved earth)

h 2 a

+
h
p
p

p = 2a tan

R a = ( h + h) sec
R
p = p sec ( + )
[ h + h ] 2 tan
=
cos cos ( + )
Note: as 0, cos( + ) cos
h 0

2h tan
p
cos2

p0 = 2h tan
R+h
= sin 1
sin p h + h 1
R =
po h cos cos( + )
CEE 615: Digital Image Processing Topic 4: Geometric Correction 4-5
W. Philpot, Cornell University, January, 01

C. Motion of the platform (aircraft/satellite) or target

The motion and orientation of the platform may introduce distortions. This distortion is
independent of the distortions caused by the sensor system itself.
1. Attitude changes: changes in the platform that are significant over the time required to
scan a full scene.
a. Pitch - a rotation of the aircraft (or satellite, or ship) in the direction of motion. Pitch
will cause changes in the spacing of the scan lines.
platform

scan lines

b. Roll - the rotation about the velocity vector. Roll causes lateral shifts in the scan line
position and scale changes in the line direction.

flight path
nadir ground path

c. Yaw - rotation of the satellite or aircraft about the local zenith vector, causes rotation
and skew distortions.
platform

non-parallel scan lines

2. Position variations: This category includes gross deviations in the position and or
direction of flight of the remote sensing platform. The simplest is a deviation in altitude
which simply causes a change in scale.
3. Slew: Motion of the platform perpendicular to the intended direction of motion. The
problem can occur when there is a cross wind over the flight path. The aircraft is
oriented parallel to the intended flight path but the actual motion is at a slight angle to the
intended direction. Thus, the scan lines will be perpendicular to the intended path but
consecutive scan lines will be shifted in the direction perpendicular to the path.

wind
orientation
of aircraft

actual direction
of flight
CEE 615: Digital Image Processing Topic 4: Geometric Correction 4-6
W. Philpot, Cornell University, January, 01

D. Target effects: Distortions may be introduced into the image by target characteristics,
especially by deviations of the target surface from a plane surface and by motions of the
target.

Target motion: The type of distortion introduced by

motion of the target will depend entirely on the nature of
that motion relative to the sampling rate & sequence of
the imaging system. For instance, when the motion of the
scene is of the same order as the sampling rate. For
example, a photo of a nearby building taken from the side
window of a car moving at 50 mph when the shutter
speed is 1/60 second, the image will be blurred. At the
other extreme, the rotation of the earth is slow relative to
the sampling speed of the Landsat MSS (about 0.4 s)
and it is even slow relative to the scan rate (33 ms/scan).
Thus, there is no obvious blur in the final image.
However, between the time that the first scan of a
Landsat MSS image and the time of the last scan, the
earth will have rotated a significant distance relative to
the size of a resolution element. The resulting distortions
for two different cases of earth rotation relative to
satellite motion are illustrated

Geometric Correction
The general process of geometric correction consists of several steps:
1. Select the appropriate projection or reference map (or image).
a. registration:simple point-too-point match of an image to another image or map,
b. rectification:correcting an image to a specific map projection.
2. Select a regular grid which fits the desired projection (i.e., determine the spacing and
position of the grid points.)
3. Select a set of "ground control points" (GCP's) -- pixels whose locations in the base map
can be determined accurately.
4. Define the transformation and compute the positions of the reference grid points in the
image coordinate system. The transformation should deform the grid in such a way that
the average distance between GCP pixels and their map locations is minimized.
5. Resample the image data in order to assign gray values to each grid point.
Steps 1 and 2 depend on individual applications.
CEE 615: Digital Image Processing Topic 4: Geometric Correction 4-7
W. Philpot, Cornell University, January, 01

Map Projections
There is no perfect map projection. That is, there is no projection which preserves all the
characteristics of a three dimensional surface. Projections have been developed which preserve
either area or shape or direction or distance, but never all.
A. Conformal (orthomorphic) projections: Projections that represent shape correctly within
certain limits. Relatively small portions of the earth's surface are properly represented in
shape on a conformal projection. Actually, the conformal principle applies only to each
infinitesimal element of the map. Angles at each point are correct, and consequently the local
scale in every direction around any one point is constant, so the map user can measure
distance and direction between near points with a minimum of difficulty
Ex: Mercator projection: Lines of constant compass bearings are plotted as straight
lines.
Transverse Mercator: Lines of constant scale are straight lines parallel to the
central meridian for the sphere. State Plane coordinates
and Universal Transverse Mercator UTM) coordinates are
special cases of this projection.
Space Oblique Mercator: Projection developed (1978) specifically for the continuous
mapping of imaging from satellites. The ground track for
the satellite is held true to scale, and mapping is made
basically conformal. (Because of the relative motion of
Earth and satellite, the ground track is curved.)
B. Equal area projections: projections which show area correctly. These projections generally
distort shape badly.
Ex: Albers Equal-Area Conic: No area distortion. Along one or two standard parallels
there is no scale or angular distortion.
C. Azimuthal (Zenithal) projections: Projections which correctly portray directions from a
specified point.
Ex: Polar Stereographic: The reference point is the Earth's pole. Lines of constant
scale are circles centered at the pole.
D. Equidistant projections: Projections which portray distances correctly from a stated point.
Ex: Azimuthal Equidistant: Shows both distances and directions correctly from the
chosen center of the map. Shape and area are distorted.
There are also projections which preserve none of the four characteristics (area, shape, direction,
or distance), but which are useful for other reasons. For small areas of the earth's surface nearly
any standard projection will accurately represent the surface accurately. Locally, a simple
rectangular coordinate system will often be entirely adequate.
In satellite imagery, the size of the area imaged is generally large enough that the earth's
curvature cannot be ignored. The most natural projection in most cases is some form of oblique
Mercator projection. For Landsat MSS data, the usual choices include Space Oblique Mercator
(SOM), and Hotline Oblique Mercator (HOM), which are both relatively distortion-free for
single scenes.
CEE 615: Digital Image Processing Topic 4: Geometric Correction 4-8
W. Philpot, Cornell University, January, 01

Selection of Ground Control Points

Ground control points are pixels whose positions can be determined accurately on the base
map. These are usually easily recognizable features or landmarks in the image.
Examples: - crossroads
- land/water boundaries
- cultural features (airstrips, buildings, etc.)
Geometric Transforms
Having chosen the control points, one must determine a transform that will map the base map
coordinates to the image data. That is, find a function which relates the base map coordinates,
(x, y), to the image coordinates (x', y'):
x' = f(x,y); y'= f(x,y)
The procedure is as follows:
1. Select a transformation (linear, 2nd order, 3rd order, ...)
2. Determine the coefficients for that transformation that will minimize the error in
transforming the ground control points. (Usually a least-squares fit procedure.)
3. Analyze the residual errors. (How badly does the transform misplace data points?)
4. If only a few of the control points account for most of the error, then check the outliers to
be certain that the coordinates and image positions are correct.
a. If errors are found in the specification of the control points, rerun the routine that
determines the coefficients for the transformation (step 2)
b. If there is sufficient uncertainty in the location of a ground control point, discard it
and rerun the routine (step 2).
4. If the residual errors remain unacceptable, try a different transformation (Usually, this
means using a higher order equation.)

The simplest type of transformation is linear, involving translation, rotation, and magnification.
Minimally, one needs only two control points to define a linear transform, although many more
would be desirable.

1. Linear Transforms: Can be done using only two control points, although that's not a good
idea)

a. Translation: x' = x + xo y' = y + yo

b. Rotation: x' = y sin + x cos y' = y cos - x sin
c. Magnification x' = mx y' = ny
d. General linear transform: x' = ao + a1x + a2y y' = bo + b1x + b2y

2. Non-linear transformations
x' = ao + a1x + a2y + a3xy + a4x2 + a5y2 + . . .
y' = bo + b1x + b2y + b3xy + b4x2 + b5y2 + . . .
CEE 615: Digital Image Processing Topic 4: Geometric Correction 4-9
W. Philpot, Cornell University, January, 01

Resampling
In resampling, each data point (pixel) in the base map coordinate system is assigned a value
(intensity, gray value, etc.) based on the gray values of local image pixels. Consider an example
in which an image of a checkerboard pattern is geometrically corrected.

pixels sites in the

original image data
(checkerboard pattern)

sites of the pixel

coordinates of the desired
base map (plotted in the
image coordinate system.)

1. Nearest-neighbor resampling
The gray value of the image pixel is assigned to the nearest base map coordinate:
- the base map grid values correspond to actual measured values.
- the value assigned to the grid point will probably not be the same as that which would
have been measured at that point.
- fastest and cheapest method of resampling.
- does the least radiometric damage to the data.
- tends to result in a blocky appearance at sharp boundaries.

Nearest-neighbor resampling
CEE 615: Digital Image Processing Topic 4: Geometric Correction 4-10
W. Philpot, Cornell University, January, 01

2. Bilinear interpolation
Bilinear interpolation is a two-dimensional version of linear interpolation. It involves
finding the four image pixels closest to the grid coordinate and obtaining the value for the
grid coordinate by linear interpolation. In effect the variation in gray values is modeled as a
series of planar surfaces (Assumes that the original scene varied linearly in the vicinity of the
four sample pixels.)

Bilinear Interpolation

Only three points are needed for a 2-dimensional linear interpolation since three non-
collinear points uniquely define a plane. However, bilinear interpolation the 4 nearest
neighbors. The approach is to use a sequence of one-dimensional interpolations between
pairs of points rather than a single two-dimensional fit. The bilinear procedure is somewhat
more efficient than a fully 2-dimensional procedure.

In this example, the resampled data are undersampled, i.e., there are fewer grid
coordinates than image pixels. The dark squares indicate adjacent sets of four pixels for
which there is no corresponding grid coordinate. Interpolation is illustrated for the points
within the circle.

Note that the farther away a sample pixel is from the grid coordinate, the less the effect it
will have on the value assigned to the coordinate. The image resulting from this resampling will
have a smoother, less blocky appearance than one created by nearest-neighbor interpolation. At
the same time there will be some loss of radiometric information since the value assigned to the
grid pixel represents a weighted average of the four nearest points.