Geolocating Images

Most satellite images are shipped without locational information, or is located by sensor position, which usually is not precise enough for connection with other images or map data. To "geolocate" the image, we "rectify" it using one of several processes.

1. Registration - colocating two images that may or may not be in proper coordinates.
2. Georeferencing - puts the image data into geographic coordinates associated with a projection and datum. "Ground control points" are used to locate the same feature on map and image.
3. Orthorectification - both georeferences the data and corrects for distortion due to terrain elevation (layover). It requires a DEM and information about the camera or sensor. The ground control points are 3D is this case.

The rectification should be done AFTER the classification or other image analysis, because the process usually degrades the quality of the original image by resampling or otherwise merge values in cells.

The process has three basic steps.

1. Finding control points
2. warping the image using a polynomial transformation

   the polynomial "order" determine how much distortion can be admitted into the warping model

   • a "linear" model allows only shifting in x and y, rotation, or skew in scale in x and y
   • a "polynomial" or "2nd order" transformation allows the scale to change with distance in x and /or y, plus the above
   • a "cubic" transformation allows the above plus localized distortion due to the lens (not common in satellites)

   The square root of the average of x and y errors sqaured for all points (geez dave, write that out) is called the Root Mean Square Error. In general is it best to use the LOWEST order transformation that you can get away with (RMSE <1 pixel).

3. resampling the image to create new pixel in the geolocated reference space
   • nearest neighbor selects the closest cell to the correct geographic location (computationally easy, but can result in shifts in linear features and edges)
   • bilinear interpolates the value for the DN in the cell from the adjacent cells in x and y (extreme data are lost, but image is "smoothed")
   • cubic convolution uses an array of 4 cells in both x and y fit to a polynomial (tougher for your CPU)
     
     see these pages for images on these different methods
     http://cct.rncan.gc.ca/resource/tutor/fundam/chapter4/04_e.php?p=1
     http://idlastro.gsfc.nasa.gov/idl_html_help/Interpolation_Methods.html

   if you are not going to classify or interpret the data first, you should only use nearest neighbor, because the others will average out the cell values and mute the original image.

much of this from www.coastal.uno.edu/coastal/high-edu/cs4096/lectures or the Univ AZ Remote sensing tutorials