Remote Sensing Lab 6: Geometric Correction

Goal
The primary goal of this lab is to become familiar with the image preprocess technique known as geometric correction. We will be going over the two major types of geometric correction: Image-to-map rectification and Image-to-Image registration. Both are equally important and require the implementation of Ground Control Points (GCPs) to rectify distorted satellite images.

Methods
The first method we are going to perform is the Image-to-map rectification. For this we are using a United States Geological Survey (USGS) 1.5 minute digital raster graphic image of the Chicago Metropolitan Statistical Area to correct a Landsat TM image of the same location. To start correcting the Landsat image, we first have to establish Ground Control Points (GCPs). The GCPs are a Multispectral tool in Erdas that geometrically link the USGS image to the distorted Landsat image. We decided to use a 1st order Polynomial model and begin placing the four GCPs on each image. We choose four GCPs because with a 1st order polynomial, you need at least three to geometrically correct an image; so by using four, we are guaranteed  to be more accurate in the correction. Once we have placed four GCPs on the USGS image, we have to place the corresponding GCPs on the Landsat image. As you place the points down, you can see a RMS error appear next to each GCP pair below the image. In order to have the most accurate correction possible, you try and make the total RMS error as low as possible by modifying the GCP locations. The smaller the error, the more accurate your correction will be. On my model, in Figure 1 below, the total RMS error is .0009 which is very accurate and will allow us to run the model using the Nearest Neighbor method. After the model is run, and we lay the newly corrected image over the USGS image, we can see it is now geometrically accurate. There will always be error unless you can get your RMS error to be perfectly 0.  Now that the Landsat image is geometrically corrected, it is ready to undergo interpretative analysis.

The second method we use to geometrically rectify a distorted image is by doing Image-to-Image registration. For this method, we are using two images of Sierra Leone taken in 1991. One of the images is geometrically correct while the other is severely distorted. If you open both images in the same viewer and use the swipe tool, you can see first hand how distorted the image is compared to the accurate one. For this correction, we are going to use a 3rd order polynomial since the image is severely distorted. This will make us use more Ground Control Points which, in turn, will make the correction more accurate. To get started, create a GCP on the correct image, and place its pair in the same location on the distorted image. Repeat this process 12 more times and then check the RMS error. Remember, we want to get that error as low as possible by re-positioning the points on the images. In Figure 2 below, you can see my total RMS error of .0023. Once the total RMS error is less than .5 you can stop re-positioning the points and run the model using bilinear interpolation. The resulting image of Sierra Leone is a mirror image of the original correct image. Now that it is geometrically accurate, analysis can be conducted on it with accuracy.

Results

Figure 1 - The geometrically correct image on the right using 4 GCPs to correct the Landsat image on the left with a RMS error of .0009.

Figure 2 - The geometrically correct image on the right used to correct the image on the left with 12 GCPs and a RMS error of .0023.