1. Geometric Correction of Imagery
Lecture 4
prepared by R. Lathrop 10/99
updated 2/08
Readings:
ERDAS Field Guide 6th Ed. Ch 9, 12
App B Map Projections

2. Where in the World?

3. Learning objectives Remote sensing science concepts Math Concepts
Projection and Coordinate Transformations
Registration vs. Rectification
2D vs. 3D rectification
Resampling
Math Concepts
Polynomial equations
Root mean square error and geometric accuracy measures
Skills
Registration/rectification/resampling process

4. Geometric Correction of Imagery
The intent is to compensate for the distortions introduced by a variety of factors, so that corrected imagery will have the geometric integrity of a planimetric map.
Systematic distortions (e.g., curvature of the earth, the sensor etc.): predictable; corrected by mathematical formulas. These often corrected during preprocessing.
Nonsystematic or random distortions: corrected statistically by comparing with ground control points. Often done by end-user
Geometric correction includes systematic distortions (e.g., distortion resulted from the curvature of the Earth and the sensor be used) and nonsystematic or random distortions (by Rectification techniques to deal with).

5. Some basic terminology --Map projections

6. Map of NJ Pinelands under different projections
What went wrong?

7. Map Projections
A map projection system is any system designed to represent the surface of a sphere or spheroid (such as the Earth) on a plane, i.e., the manner in which the spherical (3d) surface of the Earth is represented on a flat (2-d) surface.
Reprojection: transformation from one projection system to another using a standard series of mathematical formulas

8. Map Projections
Think about lighting up the interior of a transparent globe that has lines and features boundaries drawn on its surface. If the lighted globe is placed near a blank wall, you can see the shadow of the globe’s map on the wall. You will have “projected” a 3d surface to a 2d surface.
All projections involve some form of distortion

9. Map Projections
Think about lighting up the interior of a transparent globe that has lines and features boundaries drawn on its surface. If the lighted globe is placed near a blank wall, you can see the shadow of the globe’s map on the wall. You will have “projected” a 3d surface to a 2d surface.
All projections involve some form of distortion

10. Properties of Map Projections
Conformality: preserves shape of small areas
Equivalence: preserves areas, not shape
Equidistance: scale of distance constant
True direction: preserves directions, angles
There is no single “perfect” map projection, all projections require some form of compromise

11. Families of Map Projections
Cylindrical
Conical
Azimuthal
Distortion is minimized along the line of contact
Tangent: intersect along one line
Secant: intersects along two lines
Graphic from

12. Factors in Choosing a Map Projection
Type of map
Size and configuration of area to be mapped
Location of map on globe: tropics, temperate or polar regions
Special properties that must be preserved
Types of data to be mapped
Map accuracy
Scale

13. Universal Transverse Mercator (UTM)
UTM is a planar coordinate system
Globe subdivided into 60 zones - 6o longitude wide
Referenced as Eastings & Northings, measured in meters
Graphics from:

14. State Plane Coordinate system
State Plane is a planar-rectangular coordinate system
Each state has its own coordinate system oriented to the particular shape of the state. SP based on a conformal map projection.
Referenced as X & Y coordinates, measured in feet
Graphics from :

15. Other Considerations
Spheroid: the Earth isn’t a true sphere. Most projections require the definition of a spheroid (or ellipsoid), which is a model of the earth’s shape. Choose the spheroid appropriate for your geographic region (e.g., Clarke_1866).
Datum: reference elevation, determined by National Geodetic Survey (e.g., NAD83, WGS84)
Minor axis
Major axis

16. Types of Coordinates
Geographical: spherical coordinates based on the network of latitude and longitude lines, measured in degrees
Planar: Cartesian coordinates defined by a row and column position on a planar (X,Y) grid, measured in meters or feet. Examples include: Universal Transverse Mercator (UTM) and State Plane

17. Latitude and Longitude Coordinates
Spherical coordinates are based on the network of latitude (N-S) and longitude lines (E-W), measured in angular degrees
Latitude lines or parallels
Longitude lines or meridians
Graphics from:

18. Coordinate Transformations
Georeferencing refers to the process of assigning map coordinates to image data.
Transformation equation allows the reference coordinates for any data file location to be precisely estimated in a different coordinate system.

19. Transforming from one Coordinate System to Another: What transformations need to be accomplished?
x,y : map X’, Y’ : image
Map
Image

20. What transformations need to be accomplished? Rotate
Map
Image

21. What transformations need to be accomplished? Scale
Map
Image

22. What transformations need to be accomplished? Translate
Map
Image

23. Transformation Equation
To transform from the data file coordinates to a reference coordinate system requires a change in:
Scale
Rotation
Translation

24. Example: Transformation Equation from Jensen 1996 p 133
How to put this transformation into a mathematical equation.
X’ = x y
Y’ = x y
x,y : map X’, Y’ : image

25. Example: Transformation Equation from Jensen 1996 p 133
X’ = x y
Y’ = x y
x,y : map X’, Y’ : image
Given an x, y location, compute the X’, Y’ location.
x = 598,285 E y = 3,627,280 N
X’ = = 190.3
Y’ = = 180

26. Calculating Transformation Equations
The statistical technique of least squares regression is generally used to determine the coefficients for the coordinate transformation equations.
Want to minimize the residual error (RMS (root mean square) error) between the predicted (X’) and observed (X’orig) locations
RMS error = SQRT[(X’ - X’orig)2 + (Y’ - Y’orig)2]

27. Geometric Correction of imagery
-- Registration and Rectification

28. Registration vs. Rectification
Registration: reference one image to another of like geometry (i.e. same scale)
Rectification: process by which the geometry of an image area is made planimetric by referencing to some standard map projection.
Registration and rectification involve similar sets of procedures. Both registration and rectification require some form of coordinate transformation. Many references to rectification may also apply to image-to-image registration.
In many cases, images of one area that are collected from different sources must be used together. To be able to compare separate images pixel by pixel, the pixel grids of each image must conform to the other images in the data base. The tools for rectifying image data are used to transform disparate image to the same coordinate systems. Registration is the process of making an image conform to another image. A map coordinate system is not necessarily involved. For example, image A is not rectified and it is being used with Image B, then image B must be registered to image A so that they conform to each other. In the example, image A is rectified to a particular map projection, so there is no need to rectify image B to a map projection.
Rectification is the process of transforming the data from one grid system into another grid system using geometric techniques. Since the pixels of the new grid may not align with pixels of the original grid, the pixels must be resampled.
Resampling is the process of extrapolating data values for the pixels on the new grid from the values of the sources pixels.

29. When to do rectification
Whenever accurate area, direction and distance measurements are required, rectification is required.
Integration of the image to a GIS: need to be able to georeference an individual’s pixel location to its true location on the Earth’s surface.
Standard map projections: Lat/long, UTM, State Plane
Before rectifying the image data, you must determine the appropriate coordinate system for the data base. Rectification, by definition, involves georeferencing, since all map projection systems are associated with map coordinates.

30. Rectification Process (Registration process is similar)
How?
Locate ground control points
Compute and test a coordinate transformation matrix. Iterative process till required RMS error tolerance is reached.
Create an output image in the new coordinate system. The pixels must be resampled to conform to the new grid.

31. Ground Control Points (GCP)
Ground control points (GCPs): locations that are found both in the image and in the map
The GCP coordinates are recorded from the map and from the image and then used in the least squares regression analysis to determine the coordinate transformation equations
Want to use features that are readily identifiable and not likely to move, e.g., road intersections
DONOT confuse GCP with GRC(ground resolution cell) in the 1st lecture.

32. Ground Control Points (GCP)
Map
Image x y X’ Y’

33. 1st order transformation
Also referred to as an affine transformation
X’ = ao + a1x + a2y X’,Y’ = image Y’ = bo + b1x + b2y x, y = map
ao, bo: for translation
a1, b1: for rotation and scaling in x direction
a2, b2: for rotation and scaling in y direction
6 unknowns, need minimum of 3 GCP’s
difference in scale in X & Y, true shape not maintained but parallel lines remain parallel
A general rule of thumb is to use a first-order affine polynomial whenever possible. Select a higher-order polynomial (second or third order only) when there are serious geometric errors in the dataset. These types of error are often found in imagery obtained from suborbital aerial platforms where roll, pitch, and yaw by the aircraft introduce unsystematic, non-linear distortions.

34. 2nd order polynomial X’ = ao + a1x + a2y + a3x2+ a4y2 + a5xy
Y’ = bo + b1x + b2y + b3x2+ b4y2 + b5xy
12 unknowns, need minimum of 6 GCP’s
nonlinear differences in scale in X & Y, true shape not maintained, parallel lines no longer remain parallel
Higher order polynomials are different than rubber sheeting

35. Minimum of GCP’s
Higher orders of transformation can be used to correct more complicated types of distortion. However, to use a higher order of transformation, more GCP’s are needed. The minimum number of GCP’s required to perform a transformation of order t equals:
( ( t + 1 ) ( t + 2 ) ) 2
t Order Min. # of GCP’s
1 3
2 6

36. Root Mean Square (RMS) Error
RMS Error per GCP, Ri:
Ri = SQRT( XRi2 + YRi2)
where Ri = RMS Error per GCPi
XRi = X residual for GCPi
YRi = Y residual for GCPi
Source GCP
X residual
RMS error
Y residual
Retransformed GCP

37. Total RMS Error X RMS error, Rx Y RMS error, Ry
Rx = SQRT( 1/n (SUM XRi2)) Ry = SQRT( 1/n (SUM YRi2))
where XRi = X residual for GCPi
YRi = Y residual for GCPi
n = # of GCPs
Total RMS, T = SQRT(Rx2 + Ry2)
Error Contribution by Individual GCP, Ei :
Ei = Ri / T

38. RMS Error: Example X Y xi xr Resid Resid2 yi yr Resid Resid2
TotalSum
X RMS Error = RMSx = SQRT(0.59/5) = SQRT(0.118) = 0.34
Y RMS Error = RMSy = SQRT(1.11/5) = SQRT(0.222) = 0.47
Total RMS Error = SQRT( ) = SQRT(0.340) = 0.58

39. RMS Error: Example Ei = Error Contribution by Individual GCPi
Error Contribution per GCP
Rx Ry Ri Ei
GCP (=0.86/0.58)
GCP (=0.86/0.58)
GCP (=0.86/0.58)
GCP (=0.86/0.58)
GCP (=0.86/0.58)
where Ri = RMS Error per GCP
Ei = Error Contribution by Individual GCPi

40. RMS Error Tolerance
Rectification always entails some amount of error. The amount of error that the user is willing to tolerate depends on the application.
RMS error is thought of as a error tolerance radius or window. The re-transformed pixel is within the window formed by the source pixel and a radius buffer
For example: a RMS of 0.5 pixel would still put the transformed coordinate within the source pixel

41. RMS Error Tolerance 2 pixel RMS error tolerance (radius) Source pixel
Retransformed coordinates within this range are considered correct
From ERDAS Imagine Field Guide 5th ed.

42. Geometric Accuracy Measures
RMS generally equivalent to 1 standard error in statistical parlance. Approximately 68% of the residual errors are less than + or - the threshold distance (assuming a normal distribution).
Alternative measure is CE90% or a circular error of 90%, i.e., 90% of the residual errors are less than +- the threshold distance. This would be a more restrictive standard if set at the same distance threshold.

43. Error Tolerance 2 pixel RMS error tolerance (radius) Source pixel
Individual realization of a GCP
CE90 threshold
Using RMSE as a metric: On average, an individual realization of a GCP is within 2 pixels of the predicted source pixel location.
Using CE90% as a metric: 90% of the points are within the threshold distance of the predicted source pixel location

44. Error Tolerance 2 pixel RMS error tolerance (radius) Source pixel
Individual realization of a GCP
CE90 threshold
If RMSE and CE90% were set at the same distance threshold in the above example, then the CE90% criteria would not be met because less than 90% (i.e., 7 out of 10 above) of the samples fell within the set distance threshold.

45. Geometric Accuracy: how close do you need?
Error tolerance
Either as CE 90% or linear RMSE
Mapped pixel vs. “true” ground location
The greater the geometric accuracy the higher the cost
Example from Quickbird Scale CE90% RMSE 1:50, m m
1:5, m M

46. RMS Errors: Bias?
Are the residuals unbiased in the X or Y direction?

47. Geometric Accuracy
RMSE/CE90 assume that the residuals are randomly distributed (positive or negative) and unbiased in particular geographic direction.
To check for bias, plot the residuals on a map and look for spatial patterns.
For example, if the residuals are all negative in the interior of the map but are positive along the edges, you might consider going to a higher order polynomial

48. 3D GIS
With the increasing availability of high spatial resolution imagery, as well as LIDAR, there is a push towards 3D GIS showing buildings and other features as 3D objects
Requires stereo-imagery and the application of digital photogrammetric principles

49. Example of Digital Photogrammetry software: ERDAS IMAGINE StereoAnalyst
Graphic from: Gis.leica-geosystems.com

50. Digital Photogrammetric WorkFlow: Example using ERDAS StereoAnalyst
Define sensor model: describes the internal geometry of the sensor as the image was captured
Measure GCPs in X, Y, Z
Automated Tie Point Collection: points that are visually recognizable in the overlap are of multiple images
Bundle Block Adjustment to establish relationship between the various images in a project, the camera/sensor and the ground
Graphic from: Gis.leica-geosystems.com

51. Digital Photogrammetric WorkFlow: Example using ERDAS StereoAnalyst
Automated DTM (Digital Terrain Models) Extraction to identify and measure the 3D positions of common ground points
Orthorectification: using the sensor model and DTM, errors associated with sensor orientation and topographic relief displacement are removed.
For more info, consult the ERDAS IMAGINE StereoAnalyst User’s Guide

52. Ortho-rectification
Orthorectification: correction of the image for topographic relief distortion (3D).
This is especially important in mountainous areas where there can be distortions (e.g. scale changes) due to the varying elevations in the image scene.
If stereo-imagery is not available, orthorectification can still be accomplished using Z values (elevation from a DEM)

53. Resampling process
The rectification/resampling process at first glance appears to operate backwards in that the transformation equation goes from the map to the image, when you might expect it to go from the image to the map.
The process works as follows: - create output grid backtransform from map to image - resample to fill output grid

54. Resampling: for each output pixel calculate the closest input coordinate
Rectified Output image Original Input image
Map
Image

55. Resampling process
An empty output matrix is created in the map coordinate systems at the desired starting coordinate location and pixel size
Using the transformation equation, the coordinates of each pixel in the output map matrix are transformed to determine their corresponding location in the original input image
This transformed cell location will not directly overlay a pixel in the input matrix. Resampling is used to assign a DN value to the output matrix pixel - determined on the basis of the pixel values which surround its transformed position in the original input image matrix

56. Resampling Methods Nearest-neighbor Bilinear Interpolation
Cubic Convolution

57. Nearest-neighbor
Substitutes in DN value of closest pixel. Transfers original pixel brightness values without averaging them. Keeps extremes.
Suitable for use before classification
Fastest to compute
Results in blocky “stair stepped” diagonal lines

58. Nearest neighbor resampling: chose closest single pixel
Rectified Output image Original Input image
Map
Image

59. Bilinear Interpolation
Distance weighted average of the DN’s of the four closest pixels
More spatially accurate than nearest neighbor; smoother, less blocky
Still fast to compute
Pixel DN’s are averaged, smoothing edges and extreme values

60. Bilinear Interpolation: weighted average of four closest pixels
Rectified Output image Original Input image
Map
Image

61. Cubic Convolution
Evaluates block of 16 nearest pixels, not strictly linear interpolation
Mean and variance of output distribution match input distribution, however data values may be altered
Can both sharpen image and smooth out noise
Computationally intensive

62. Cubic Convolution: weighted average of 16 closest pixels
Rectified Output image Original Input image
Map
Image

63. Summary 1 Projection and Coordinate Transformations
2 Registration and Rectification:
What is image-to-image registration vs. image-to-map rectification?
How to rectify a distorted 2-D image? What is ortho-rectification?
GCP (Ground control points)
Geometry accuracy (RMS error)
Resampling methods: nearest neighbor vs. bilinear vs. cubic convolution

64. Geometric correction: work flow
Selection of GCPs;
Coord. transform;
DN Resampling.
Registration or
Rectification
Rotate
Transform
Scale

65. Homework 1 Homework: Geometric correction;
2 Reading web lecture 4 supplemental;
3 Reading Ch. 7; ERDAS Guide Ch. 10, 13, and App. B.