GEOREFERENCING

Georeferencing Satellite Images—Getting Data Ready for Use in

Change Detection

Introduction
Change Detection from Multi-Date Satellite Imagery
To detect changes on the Earth’s surface, multiple images that were acquired on
different dates are necessary. Ideally, these images were acquired with the same
sensor and overlap perfectly. Many different types of changes can be observed, ranging
from seasonal changes in the vegetation to changes caused by drought, fire and
flooding. Human-caused changes such as large-scale deforestation or urban
development are among the most dramatic changes that have been documented using
satellite imagery.

Change detection based on satellite imagery can be very difficult if multiple factors and
multiple kinds of changes act together. For example, changes in forest cover as
determined from a Landsat TM Image can be caused by natural fires, commercial
deforestation or simply by changes in season between the dates for which satellite
imagery is available.

Methods used to quantify change also may affect the outcome. There are many
different techniques:

• Postclassification comparison – Two or more image dates are classified

independently from each other using the same image classification scheme.
Once the classification is completed the two images are compared and
differences in landcover are recorded as change.
• Multitemporal classification – Two or more image dates are combined into one
multitemporal data set and classified on the basis of their multispectral
characteristics. This can be accomplished using supervised or unsupervised
classification procedures. If the spectral characteristics of the changed areas are
significantly different from the spectral characteristics of the unchanged areas,
this technique works well.
• Temporal image differencing – The images are simply subtracted from each
other.
• Temporal image rationing – The ratio of the two images is computed. Areas with
no change approach 1, while areas with change will have higher or lower ratio
values.
• Other techniques include Principal Components Analysis and Change Vector
Analysis.

Research in improving methods for change detection is continuing and which of these
techniques produces the best results is hotly debated. We believe that each of these
techniques has value and you have to choose a technique that is most appropriate for
your data, your imagery and your analysis capabilities.

As in any experiment, change detection will be most successful if you can keep many of
your variables constant. Thus, you should strive to have imagery from the same sensor,
acquired during the same season, and taken during similar atmospheric conditions.

1
GEOREFERENCING

Anniversary dates can be used to assure that there is similarity in sun angles and
illumination. However, most importantly the images must overlap perfectly. Variation in
locations obviously will bias your change detection results dramatically. The next
section deals with the image preprocessing procedures that can be used to assure
overlap in your imagery.

Image Preprocessing
Image preprocessing describes mathematical procedures used to correct distortions and
degradations of data in the image in order to produce a more accurate image. The three
most important procedures involve geometric corrections, radiometric calibration, and
the elimination of noise in the data. While radiometric corrections and elimination of
noise are not always required, geometric correction is absolutely essential to turn
satellite imagery into useful information.

Raw, remotely-sensed image data, gathered by a sensor onboard a satellite or air

craft, are representations of the irregular surface of the Earth. Images are distorted by
the curvature of the Earth and characteristics of the sensor being used. Geometric
correction deals with adjusting an image so that it can be used as a map and can be
combined with other data sources.

Rectification is the process of projecting data onto a plane and making it conform to a
map projection system. In rectification, data from one grid system is transformed into
another grid system using mathematical equations. Since pixels do not always align
after transformation sometimes they must be resampled. In resampling, data values for
the pixels in the new grid are extrapolated from values of the old grid.

Orthorectification is necessary because variations in the terrain cause geometric

distortions of distances and angles in the photo when compared to the reality. To
orthorectify a satellite image or aerial photograph, the terrain distortions are corrected by
applying ground control data, triangulation results and elevation models to the image.

Frequently, digital data from different sensors are combined in remote sensing. This
requires that pixels in one image grid correspond to pixels in the other image grid.
Images do not need to be rectified, you can just register image A to image B, by
assigning each pixel of image A its corresponding pixel in image B. This can be done
using mathematical equations similar to those used in rectification. If no particular
projection system is used, this process is called Image-to-Image Registration.

Georeferencing refers to the process of assigning map coordinates to image data. The
image may already be projected onto the desired plane, but not yet referenced to the
proper coordinate system. Rectification involves georeferencing by definition.

Most image data that you will purchase will already be geocoded. Geocoded data are
images that have been rectified to a particular map projection and pixel size, and usually
have had radiometric corrections applied. Geocoded data should be rectified only if they
must conform to a different projection system or be registered to other rectified data.

2
 GEOREFERENCING

NASA Data Levels

NASA has developed a system to identify levels of image processing. It may be worth
the extra expense to have the corrections and geocoding done by the data distributing
center rather than in your own lab.

Level 0 – Raw data with elimination of transmission errors and artifacts.

Level 1 – This data includes:

Geometric correction – referencing the image to Lat/Long on the Earth’s
surface, based on the satellite’s position and viewing angle at the time of
acquisition. Uses either a spheroid model of Earth or a detailed terrain
model; the latter enables higher precision in hills/mountains. Requires
Ground Control Points (GCPs: points in the image which can be accurately
located on Earth) for high precision, referenced to coordinates of the Earth.

Radiometric correction – compensating for known characteristics of the

sensor.

Atmospheric correction – compensating for the distortion (lens effect) of the

atmosphere.

Level 2 – derived geophysical variables at the same resolution and location as the
Level 1 data.

Level 3 – data has been geocoded and projected. An image is mapped to a

Projection of the Earth, and in some cases also composited (i.e. several
images are mosaiced to show a larger scene).

There are many reasons for rectifying data:

• image-to-image comparisons of individual pixels (as in change detection)

• identifying training samples according to map coordinates, prior to classification

• overlaying images with vector data such as data from Arc/Info or ArcView

• extracting accurate distance and area measurements

• performing any analyses that require precise geographic locations

Research has demonstrated that rectification tends to degrade the spectral properties of
an image and as a consequence reduces classification accuracy. Many researchers
recommend classifying an image prior to rectification. However, if you need to
determine precise geographic positions for your training samples you must rectify the
image prior to classification.

3
GEOREFERENCING

Exercise 1 Image-to-Image Registration and Georeferencing

In this exercise you will learn how to:

ƒ link two images for comparison.

ƒ rectify one Image by another Image using ground control points.
ƒ resample using the GCP Correction Tool.

1. Use the Viewer icon in the ERDAS Imagine icon panel to open two Viewers.

2. Select Session, Tile Viewers to position them side by side. In Viewer #1, browse for
raster layer saskatchwan_unref.img under c:\course\ch3_georef. This Image has not
been rectified.

3. Click the Raster Options tab. After Display as: select True Color. Under Layers to
Colors: select Red =4, Green = 5, Blue = 3. Click on Fit to Frame and click OK.

4. In Viewer #2 open raster layer saskatchwan_utm.img. Use the same Raster Options
parameters as the previous image. This file is a Landsat image that has been
georeferenced to Universal Transverse Mercator (UTM) map projection.

5. Right click the mouse in Viewer #1 and highlight Geo. Link/ Unlink... The Link/Unlink
Instructions box pops up. Try to link the unreferenced Image in Viewer #1, with the
referenced Image in Viewer #2.

Because the two Images are not currently spatially referenced to each other, they cannot
be linked. We need to Georeference the Image in Viewer #1 first.

6. You will start the Geometric Correction Tool from Viewer #1. Select Raster,
Geometric Correction to open the dialogue.

7. In the Set Geometric Model dialog, select Polynomial and then click OK. Close the
Polynomial Model Properties dialog.

4
 GEOREFERENCING

8. The GCP Tool Reference Setup dialog opens. Click OK to accept the default of
Existing Viewer.

9. The GCP Tool Reference Setup dialog closes and a Viewer Selection Instructions box
opens. Click on Viewer #2. This Image provides the reference coordinates. The
Reference Map Info will show the projection of the Image. Click OK to close it.

The Chip Extraction Viewer #3 and #4, link boxes, and the GCP Tool open. You may
want to resize and move the link boxes so that they are easier to see. In this exercise
you will rectify the Image in Viewer #1 to the Image in Viewer #2.

You will be using 2 sets of GCPs, input GCPs from the source Image and reference
GCPs, which are known coordinates of the corresponding input GCPs. With the GCP
Tool and its Automatic Transform Calculation mode turned on you can interactively edit
GCPs and calculate and edit the transformation matrix until the RMS error is acceptably
low. GCPs can be recorded from a Viewer, digitizing tablet, or from an ASCII file.

GCPs that are digitized in a Viewer are stored with the raster data that are listed in the
Viewer. However, GCPs can also be recorded from a digitizing tablet or straight from
the keyboard. In these cases, they are stored alone in ERDAS Imagine files with the
extension *.gcc.

10. Select Viewer #1 and zoom in on an area that looks like a good GCP area. Good
GCPs are features that are easily identifiable on both Images.

Ground Control Points

Accurate ground control points are essential for an accurate registration. Select
many GCPs throughout the scene. The more dispersed the GCPs are, the more
reliable the registration. GCPs to look for are the intersection of roads, airport
runways, utility corridors, towers, bridges, buildings or another feature that can be
selected precisely and seen easily on both images. Landmarks that vary should
not be used (for example, lake edges, other water bodies or vegetation). However,
in remote areas or undeveloped regions, natural features may be all you can use
for your image registration.

There are a few ways to obtain GCPs. They may be known beforehand and
entered by keyboard. GCPs may be digitized from a hard copy map on a digitizing
tablet. In this exercise we will use the mouse to select coordinating pixels from
both Images, from the source Image and the reference Image. When entering
GCPs with the mouse, it is best to try to match resolution of the imagery as much
as possible rather than using coarse imagery with finer resolution imagery. GCPs
entered with the mouse are stored in the *.img file, and those entered at the
keyboard or digitized are stored in a separate file with the extension *.gcc.

11. To make the GCP easier to see, right-hold in the Color column to the right of GCP
#1 in the GCP Tool Cell Array and select the color Yellow.

12. In the GCP Tool, click the Create GCP icon . In Viewer #1 or Viewer #3 (chip
viewer) click on the spot you have chosen. The GCP can be dragged inside Viewer #3.

5
GEOREFERENCING

The point you have selected is marked as GCP #1 in the Viewer and its X and Y
references are listed in the GCP Tool Cell Array.

13. Locate the same point in the reference Image in Viewer #2. In the GCP Tool there
should now be X and Y coordinates for both the Input and Ref columns.

14. This process should be repeated for as many points as you feel confident in
digitizing. The more good points you choose the better the accuracy of your rectification.
After you have identified three or four sets of GCP’s, ERDAS will usually begin to
automatically select a reference GCP for every subsequent input GCP you select, based
on the locations of all preceding points. You may accept the reference GCP in its
automatically selected location, or alternatively, you can click and drag it to a more
accurate location.

Observe your RMS error, which will be reported in the GCP editor window once you
have entered three sets of GCPs, and which automatically revises with the addition of
each subsequent set. A total of less than 1 pixel error would make it a reasonable
resampling. You can see how much error is contributed by each set of points. For the
points with higher error you may want to go back and move them. To do this just place
the chip viewer over the area with the GCP you want to move. You will then be able to
click and drag it to a new spot.

To select GCPs in the Cell Array, click in the Point # column, or use any of the Cell Array
selection. The chip viewer will jump to whatever GCP you select.

To delete a GCP, select the GCP in the Cell Array in the GCP Tool and right-hold in the
Point # column to select Delete Selection.

15. For the GCP Correction Tool, resampling is the process of calculating the file values
for the rectified Image and creating the new file. All of the raster data layers in the
source Viewer will be resampled. This can be done with 3 different algorithms: Nearest
Neighbor, Bilinear Interpolation, or Cubic Convolution. Click on the Resample symbol on

the Geo Correction Tool dialogue .

16. Name the output file saskatchwan_georef.img. Under Resample Method choose
Nearest Neighbor. Click Ignore Zero in Stats so that pixels with zero file values will be
excluded when stats are calculated for the output file.

17. Click OK. When the job is complete open a new Viewer and the output file you just
created. Check the new file with the reference file to see if they conform.

18. Compare the 3 Images. Close Viewer#1 without saving changes. Right-hold Geo.
Link/Unlink and link the viewers.

19. Right-hold the Inquire Cursor. Drag the inquire cursor around to verify that it is in
approximately the same place in both Viewers. Notice that as the cursor is moved the
data in the Inquire Cursor dialog are updated. After inspection close the Inquire Cursor
dialog.

6
 GEOREFERENCING

Other Image Corrections

There are several processes for correcting distorted or degraded image data in order
to create a better representation of the original scene. This may involve correcting
geometric distortions, calibrating data radiometrically, or eliminating noise in the data.
The sources of geometric distortion can result from the atmospheric reflectance,
curvature of the Earth, relief of the Earth’s surface, problems with the sensor, and
other factors involved in obtaining the data.

Radiometric correction is needed to correct for variance in scene illumination,

atmospheric conditions, viewing geometry, and instrument response characteristics.
The importance of these corrections depends on the sensor and the purpose of the
data. Scattered path radiance may cause “haze” in the Image and reduce image
contrast. Remember that scattering varies depending on wavelength, with shorter
wavelengths scattering more. One way of compensating for haze is to observe the
radiance recorded over target areas of zero reflectance. Reflectance of deep clear
water is virtually zero in the near-infrared region. Any signal observed there
represents the path radiance and can be subtracted from all pixels in that band.
These can be applied throughout a scene, but this may not always be valid. If
extreme viewing angles are involved you may also need to compensate for the
varying atmospheric path length. This can be accomplished by normalizing off-nadir
pixels to their nadir equivalents.

Image noise is any disruption or disturbance in image data resulting from sensor
malfunction, signal digitization, or data recording process. The results may be
striping or banding, which results from variations in the response of individual
detectors used within each band. A similar problem is a line drop, which occurs
when a number of adjacent pixels or sometimes an entire line contains incorrect pixel
values. Both of these errors can be corrected by normalizing the incorrect pixels with
respect to their neighboring pixels. Random noise such as bit errors may cause a
salt and pepper or snowy appearance. To correct these you must first use moving
“windows” to detect these errors. You can determine what the difference between
neighboring pixels should be. If a pixel exceeds this threshold it can be replaced by
the average of its neighboring values.

Other correction processes include sun elevation correction, earth-sun distance

correction, blackbody correction, and de-skewing.

7
GEOREFERENCING

GCP Tool Bar Options

Toggle Click to toggle the fully automatic GCP editing mode.

Calculate Click to solve the geometric model with control points.

Automatic Click to set automatic transformation calculation. When this is

set, the transformation will be updated automatically as points
change.
Compute Error Click to compute the error for checkpoints.

Select Click to activate the GCP selection tool. This tool allows you to
select a GCP symbol displayed in a Viewer by clicking on it. This
is the default tool.
Create Click to create a GCP in a Viewer.

Lock Click to lock and unlock the current tool for repeated use.

Find in Input Click to locate the selected points in the input Image. This option
finds the point selected in the GCP Tool Cell Array and highlights
it in the input Viewer.
Find in Reference Click to locate the selected points in the reference Image. This
option finds the point selected in the GCP Tool Cell Array and
highlights it in the reference Viewer.

Exercise 2 Reprojecting Images

Projections determine where and how your image is displayed on the Earth’s surface,
and provides an algorithm to flatten the globe for your specific area. Chapter 6 will
provide a more indepth introduction to projections. ArcMap can be used to reproject
Images and Grids from one projection into another, however ERDAS Imagine and
Arc/Info offer a nicer set of tools that allow you to reproject spatial data sets.

In this exercise, you will learn how to:

ƒ Use the ERDAS Imagine Layer Information dialog to learn about the projection of
your images;
ƒ Use the Image Geometric Correction dialog to reproject an image.

1. Open ERDAS Imagine and the Image file p133r48_442000etm.img in Viewer #1.
This is an Image from Burma projected in UTM. Display Layer Information by clicking
the Layer Information icon on your Viewer toolbar. Write down the Datum name and
pixel size. Close the layer information window.

2. Click the Imagine DataPrep icon and select Image Geometric Correction. When
the Set Geo Correction Input File window appears, choose Select Viewer and then click
your cursor anywhere inside your Viewer #1 image. This selects your
p133r48_442000etm Burma file as the file you want to reproject. Once you have

8
 GEOREFERENCING

selected the Image in Viewer #1, the Set Geometric Model window pops up. Highlight
Reproject and select OK.

3. The next step is to change the projection of your Burma file from UTM to Lambert.
Select the Projection Page in your Reproject Model Properties dialog box, and click Add/
Change Projection.

4. On the Custom page of your Projection Chooser, change your projection type to
Lambert Azimuthal Equal-area. Change the Spheroid Name to the Datum you recorded
from the Layer Information of your original file. Enter 100:00:00.000000 E for the
longitude center of projection and 45:00:00.000000 N for the latitude center of projection
and click OK. This last step tells the program where to center your image on the
coordinate system.

A warning window will pop up. Ignore it and press OK.

Also close the Reproject Model Properties window.

5. Now you must resample your data, by utilizing the Geo Corrections Resample Tool
. The reprojected file will be your output file.

6. Save your reprojected file as p133r048_reproj.img in your c:/course/ch3_georef

directory. Accept the Nearest Neighbor resampling default and make sure your
output cells are the same as the pixel size you recorded from the Layer Info page
(p133r48_442000utm.img). Check Ignore Zero in Stats and click OK. It would take
approximately 5 minutes to resample.

9
GEOREFERENCING

7. Because we do not want to wait 5 minutes for the reprojected Image, kill the process
by selecting Session/Active Process List.... In the Process List dialog, highlight the JOB
resample process, then click on the Kill button followed by the Close button. This will
terminate the resampling and reprojecting. You may also have to close the Job Window
by clicking OK.

8. Exit your Geo Corrections Tools window without saving changes to the current
geometric model. A reprojected file, called p133r48_442000lamb.img, has already been
created for you – open this file in Viewer #2. You can close your Viewer #1. At the
bottom of your Viewer #2 window, it should tell you that your new Image is in Lambert
Azimuthal Equal-area/ WGS 84.

Exercise 3: Checking your Reprojected Image with a Raster Layer

In this exercise you will learn how to:

ƒ test your reprojected Image with a raster layer in the same projection
ƒ use the swipe tool to examine your reprojected Image and the raster layer

10
 GEOREFERENCING

1. In the same Viewer, open the grid file glcc_asia. You will need to change Files of
Type to GRID. On your Raster page, make sure that the Clear Display box is not
checked. Click OK. Note: If you are unable to locate the grid file glcc_asia, look for a
grid titled ch3_georef. On some computers, this grid takes the name of the folder it is
located under depending on where you were in the hierarchy of folders when you
changed your files of type to GRID.

2. Go to View, Arrange Layers and drag the glcc_asia layer below the
p133r48_442000lamb.img and click Apply, Close. Zoom in on the Burma layer.
You can see the small, Burma layer in the center of your glcc_asia Image. The
Burma layer will be in color.

3. Go to Raster/ Pixel Transparency.... and check Transparent Background Click OK.

This eliminated the black box surrounding your Burma image so that it can be easily
compared to the Southeast Asia grid.

4. Maximize your Viewer so that you can compare the raster layer with the Image layer.
Go to Utility, Swipe.

11
GEOREFERENCING

Using the Swipe Tool allows you to see how well your reprojected Image lines up with
your raster layer of Southeast Asia. Zoom into the land borders of Burma, using your
Swipe Tool to reveal the raster layer underneath your Burma Image.

Question 1: How well do your two Images overlay?

Question 2: What other features could you use to double-check the reprojection of your
Images?

Zoom in closely to the primary river tributary of your Burma image. Compare the shape
of the river with the grid layer underneath.

Question 3: What do you notice about the area the grid pixels cover versus the area
represented in the Image file? What might account for any discrepancies between the
two? Consider the river, for example.

Right click and highlight Fit Image to Window. You should now see the full Southeast
Asia Image.

Question 4: Why is the Image distorted in the marginal regions?

12