1 (edited by shekk 2012-01-23 19:52:21)

Topic: Associating Output Raster to Input Data

Hi all,

I am using COSI-Corr in the context of glaciology and feature tracking.

I would like to know how the resulting rasters relate to the original input data. I have noticed that the results are smaller in size compared to the input images. When overlaying both input and output data, the output leaves a margin compared to the input (see example image: http://img337.imageshack.us/img337/9022 … sicorr.jpg). I believe this is due to the search window and the size of this margin is probably defined by the window size.

I would now need to know, how the pixel positions of the original image are represented by the result. In my example I used the frequency engine with an initial window size of 128x128 and a final size of 32x32. The XY step was 16 pixels. When I compare input and output I have a left margin of 127 px and a top margin of 128 px. However, when I calculated the correlation with the same setup and the same input images but without map information (non-referenced images), the resulting rasters were 1 px (16 original px) narrower. In this case the left margin was 143 px.

- How can I determine, which original pixels belong to each pixel in the result? In other words, is the magnitude of displacement calculated from exactly the 16x16 pixels 'underneath' the resulting pixel?
- How do the margins of 127 and 128 px come about?
- Why changes the output size when using georeferenced images?

Thanks a lot.


Re: Associating Output Raster to Input Data


The sampling of the deformation field relates to the parameters you used. The window size gives you how many image pixel will be averaged to derive the deformation field, the step size, the distance in pixel between each sampling point. Make sure to carefully read the manual as this information is provided. If you have a 1000x1000 pixel image and you use a step of 10 pixels, the displacement field will be about 100x100 pixels, minus some pixels to account for edge effects.

You can increase the sampling density by reducing the correlation step.

Note that all results are georeferenced, and that is the information you need to go from the image to the displacement image. You cannot use the "link display" function in ENVI if the images don't have the same sampling. But in any case, you should be able to use the "geographic link".

When the map information is provided, the default output is said "gridded". It means that the origin of sampling grid is chosen so that it's a multiple of the image resolution. Using this trick, if you have N images, all correlated with the same step, they will all be aligned without further resampling needed.

Does that answer your question?