Lab 6

Goal and Background

          Before a satellite image is analyzed, a number of preprocessing activities must be performed first to prepare the image for accurate information extraction. For Lab 6, students were introduced to the preprocessing exercise known as geometric correction. The two geometric correction processes introduced in this lab are image-to-map rectification and image-to-image rectification. 

Methods

          In Erdas Imagine, display the Chicago USGS 7.5 minutes Digital Raster Graphic and the 2000 Chicago satellite image from the Lab 6 folder in two separate viewers. The satellite image needs to be geometrically corrected based on the accurate DRG in a image-to-map rectification. With the satellite image activated, click on Control Points in the Multispectral tab. Select Polynomial in the Set Geometric window. Accept Image Layer (New View) as the default in the GCP Tool Reference Setup window. Next, navigate to the Lab 6 folder and select the Chicago DRG. In the Polynomial Properties (No File) window, establish a first order polynomial to develop a transformation coefficient and then accept the defaults. In the Multipoint Geometric Correction window, the input image is on the left and the reference image is on the right, each with a zoomed in window and a fit to frame window above. Delete the GCPs that were automatically added in the Multipoint Geometric Correction window. Fit both images in the window. Use the Create GCP tool and place a GCP in the input image. Place another GCP in its corresponding place in the reference image. Change the GCP colors if they're hard to see. Create four GCPs in total, making sure they are evenly dispersed. In a first order polynomial model, three GCPs are the minimum. The model solution will say Model solution is current when the minimum GCP count has been been to run the transformation. Make sure to add only one GCP to one image after the third GCP because the model will automatically place the corresponding GCP. Evaluate the Root Mean Square Error. For the first geometric correction process, try to achieve a total RMS error of 2.0 or lower. Zoom in on each GCP in one of the images and reposition it. As the GCP is moved watch the RMS error. Move the GCP in a direction that lowers the RMS error. After repositioning each GCP, the RMS should have lowered passes 2.0. Once it has, the Multipoint Geometric Correction window should resemble Figure 1. Click Display Resample Image Dialog button to begin to run the model. Save the output to a personal folder. With a Nearest Neighbor resampling method, run the model. Figure 2 is the result. 

          Display the 1991 Sierra Leone image in Erdas Imagine and overlay the Sierra Leone image on top of it. Using the Swipe function, evaluate the distortion of the 1991 image by sliding the Transition extent back and forth. Clear the Viewer Swipe window and the images from the viewer. Bring the images back in two separate viewers. With the distorted image active, click on Control Points in the Multispectral tab. Once again, select Polynomial in the Set Geometric Model window. Click OK in the GCP Tool Reference Setup. Then, navigate to the Lab 6 folder and bring in the Sierra Leone reference image. Hit OK in the Reference Map Information. This time, change the polynomial order to three in the Polynomial Properties dialog. A third order polynomial transformation required a minimum of ten GCPs. As before, add GCPs to the input image and the reference image - ten in total. Then the model will say model solution is current. Add two more GCPs to just one image to prevent a wrap tool error. It's corresponding points will automatically be placed on the other image. Reposition the GCPs until the RMS error reads 1.0 or lower. Less than 0.5 is ideal, so try for that too. The Multipoint Geometric Correction window should resemble Figure 3. Click Display Resample Image Dialog. Save the output to a personal folder, change the resampling method to bilinear interpolation, and run the model. The output image should look like Figure 4. 

Results

           For the first order polynomial transformation, at least three GCPs are required to run the model. However four GCPs were used for the Chicago map-to-image transformation because more GCPs can lead to a better geometric correction (Figure 1). It's important to disperse the GCPs evenly across the image, and place them on man-made or distinguishable features. Water is essentially featureless and vegetation can be extremely variable between images.

Figure 1: Multipoint Geometric Correction window showing the input and reference image of the Chicago area with the GCPs and RMS error.

Figure 2: The geometrically corrected image of the Chicago area.

          For the third order polynomial transformation, at least ten GCPs are required to run the model. In this instance twelve GCPs were used for the Sierra Leone image-to-image transformation (Figure 3). Resampling method was bilinear interpolation. The resulting image is much more diluted than the original image (Figure 4). Also, the Swipe function revealed some distortion in the upper left-hand corner of the output image. This might be because GCPs were more condensed on the right side of the input image than the left.

Figure 3: The Multipoint Geometric Correction window showing the input and reference image of Sierra Leone with GCPs and the RMS error.

Figure 4: The geometrically corrected image of Sierra Leone. 

Sources

Earth Resources Observation and Science Center, United States Geological Survey (1991-2000) Satellite images. Reston, VA.

Illinois Geospatial Data Clearinghouse (n.d.) Digital Raster Graphic. Champaign, IL.