Remote Sensing Lab 6

This lab was all about a image preprocessing technique called geometric correction. There are two major types of this technique and this lab was designed to develop and improve my skills when using this technique. The two kinds of geometric correction are image-to-map rectification and image-to-image registration. These are both applied to images before the images are read or interpreted to ensure accuracy of the image when it is interpreted.
 
The first kind of geometric correction I preformed in the lab was image-to-map rectification. In order to do this I brought in an image of Chicago that is skewed and not accurate, and also a map of Chicago that is spatially accurate. I then used GCPs or ground control points to correct the Chicago image. GCPs are points that I placed on the image and also the map in the same spot. In order t determine how many of these points I needed I had to look at what order polynomial this image needs. This is a measure of the distortion in the image, the more distortion there is in the image the higher of polynomial you need. This image had low distortion so it was a first order polynomial. For a first order polynomial 3 control points are required in order to accurately correct the image. In order to get these points as possible to the exact same spot on both images you use something called the Root Mean Square error or RMS. The closer the points are to exactly the same spot on the two images the lower the RMS. An ideal RMS value is below 2.0. If the RMS value is too high the correction you are trying to achieve wont happen and the output image will still be distorted. Once the RMS is at a good level the next thing to do is resample the image. Resampling was explained in my previous post but pretty much it is correcting pixels that are missing brightness values. For this first image I applied the nearest neighbor method of resampling. The resampled image is spatially accurate and can now be interpreted correctly.


GCPs being placed.

 

Corrected image on original image showing how distorted original was.

 
The second kind of correction is call image-to-image registration. I brought in an image of Sierra Leone that is very badly distorted and also brought in another image of Sierra Leone (instead of a map like the first method) that has already been corrected and is spatially accurate. In order to correct this image I used the same procedure as the first image. This image is much more distorted than the first image and because of that I needed to use more GCPs. This was a third order polynomial which requires at least 10 GCPs. I placed GCPs on the distorted image and accurate image in the same location and got a RMS value of less than 2.0. For this image the RMS value is under .5 to make sure that this output image is very accurate. Then just like the first image I resampled. This time I used a bilinear interpolation resample. This resample method is more spatially accurate than nearest neighbor and gives a smoother output image.

 


GCPs being placed.
 
 

Original distorted image on top of original corrected image showing the distortion.





Corrected image on top of original corrected image showing spatial accuracy of the correction.


 
Geometric correction is fairly easy to do if you know how to and improve your skills. The biggest challenge is getting that RMS value down to a good number. Most of the rest of the correction is done by a computer program but knowing what resampling method to use is also an important skill when doing this.

Sources:
ERDAS Image 2013
Earth Resources Observation and Science Center, United States Geological Survey