Lab 6 Geometric Correction

The purpose of this lab was to introduce students to the process of geometric correction of satellite images. For this lab I was required to geolocate a Landsat TM satellite image by the use of a USGS topographic map of the corresponding Chicago Metropolitan area. I was also to do a similar process with an image of an area in eastern Sierra Leone.

 Objective one: Image to map rectification

For the first objective of this lab I was tasked with rectifying a satellite image of the Chicago area with Erdas Imagine by the use of Geographic Control Point pairs. I used a first order polynomial for the geometric model which essentially translates the pixels from the image onto a flat plane. While it is not as accurate as using a higher order polynomial to correct the image it did save some time rendering the finished rectified image (accuracy really wasn’t a huge problem anyways, the original image was not terribly distorted).  After specifying to use the USGS topographic map as the reference image I proceeded to add Geographic Control Points (GCP) in places easily identified in both the satellite and the map. Seeing as I was only using a first order polynomial I needed only three GCPs. For good measure I added a forth to help minimize distortion which may have resulted by not distributing the GCPs evenly throughout the map. The below image (Figure 1) shows multipoint geometric correction window I used to place the GCPs. 

Figure 1. Multipoint geometric correction window. Note the similar location of the geographic control point placement in both the satellite image (Left) and USGS topographic map (Right). 

Initially the GCPs were not placed accurately which resulted in a moderately high Root Mean Square Error (RMS), a number derived from the distance formula to help the analyst to accurately place GCPs. After carefully moving the GCPs on the satellite image to better represent the locations selected on the reference map I was able to decrees the RMS from over 4.0 to ~0.31. Once the RMS was decreased to an acceptable level I was tasked to resample the image using nearest neighbor interpolation. The rectified imaged looked essentially the same as the original image with the exception of the pixels being slightly adjusted.

Objective 2: Image to image registration

For the second portion of the lab I was to correct two similar images of a portion of Sierra Leone. Unlike the image used for objective one, this image was noticeably distorted when compared the reference image. So instead of using a fist order polynomial function I used a third order one. Because of this I had to create at least ten GCPs instead of just three. Figure 2 below shows the interface and the GCPs I placed on the distorted and corresponding reference image.

Figure 2. Multipoint geometric correction window showing distorted image left and geometrically corrected reference image right. Locational data for GCPs displayed in table at bottom of image.

I ended up adding 13 total GCPs and was able to reduce my total RMS error to ~0.17 through careful adjustments of the most poorly placed GCPs. Once finished I was ready to run an interpolation process on the image which would make it geometrically correct. Unlike the first image I used a bilinear interpolation method instead of nearest neighbor. I believe this to be a good choice because, due to the distortion of the original image, nearest neighbor would probably loose data from the image. One the process was finished I was quite surprised with how well the accuracy of the resampled image compared to the reference image. When running a swipe tool to compare them I was not able to discern any spatial difference until I magnified both images enough to see individual pixels. And even then the difference was minimal.

 All data used in this lab provided by:

Satellite images: Earth Resources Observation and Science Center, United States Geological Survey.

Digital raster graphic (DRG): Illinois Geospatial Data Clearing House.