1 (edited by jscheingross 2010-08-05 13:42:21)

Topic: distorted images within GCPS optimization window

I've been running GCPS Optimization with NAPP aerial photos (scanned at 14 microns) as the slave image and a shaded relief from LiDAR (gridded to 0.5 meters) as the master image.

During optimization, I've noticed several of the tiles for my GCP seem to be distorted from the original image (this occurs in both the statistical and frequency options).  This distortion has varying degrees.  Here are some examples:

1. A rather bad case:

http://www.joelscheingross.com/cosi_corr/gcp11.jpg

and a view from the original images:

http://www.joelscheingross.com/cosi_corr/gcp11_compare.jpg

2. A mildly distorted image

http://www.joelscheingross.com/cosi_corr/gcp25.jpg

and the original:
http://www.joelscheingross.com/cosi_corr/gcp25_compare.jpg

3. A non-distorted case:

http://www.joelscheingross.com/cosi_corr/gcp07.jpg

and the originals:

http://www.joelscheingross.com/cosi_corr/gcp07_compare.jpg


In case it makes any differences, here's the hillshade I'm using as my master image:

http://www.joelscheingross.com/cosi_corr/s12_shd.jpg

It's LiDAR coverage over my main area of interest, with a 10 m DEM (from USGS seamless, resampled to match the resolution of the LiDAR) covering the areas where there's no LiDAR coverage.

For the time being, I've been removing the badly distorted tie points from my GCPS file, and re-optimizing with only the mildly distorted and un-distorted points.  This gives marginally better results when looking at the "convergence quality report" in the GCPS optimization report. 

I was wondering if anyone else has experienced similar problems and know why they arise?  Is simply removing the badly distorted GCPs and moving on the best strategy?  Or is there something more here I should be concerned about?

I have used my pruned set of GCPS to go on and make a resampled image, and these images appear to match well to the LiDAR hillshade, so I'm assuming this is OK, but thought I should check here before I start processing tons of images.

Also, a bit unrelated, I've noticed when using the frequency correlator my convergence often increases (sometimes by a factor of 10 or more) during my iterations.  Is this normal?

Thanks,
Joel

2

Re: distorted images within GCPS optimization window

Hi Joel,

This pb is the same one you had with your resampling. You've simply defined GCP in areas where you DEM has large unphysical variations. In the GCP optimization window, in the resampling options, you can fix the resampling distances to be around 1.2 and it should fix that pb. However, you'll still see weird distortions in your image patches because this fix won't correct your DEM. Since your DEM cannot really be corrected, removing these points is probably the best thing to do indeed.

What do you mean when you say that convergence increases with the frequency correlator? You mean it's better or it's worse?

Cheers,
Sebastien

3

Re: distorted images within GCPS optimization window

Thanks Sebastien, that explanation makes sense.

For GCPS Optimization with the frequency correlator, I'm seeing the value of the Avg X and Avg Y jump dramatically, here's an example:

;Convergence quality report:
;Loop number, Avg x, Avg y, Norm Avg xy, Stdev x, Stdev y, Norm Stdev xy
;     1    -0.978851    -0.134636     0.988067     5.081732     7.402703     8.979088
;     2    -0.418247     1.224246     1.293719    12.079823    10.010157    15.688383
;     3     0.975763    -1.516300     1.803131     8.763812    12.059406    14.907504
;     4    -0.317347     2.918323     2.935527    10.568963    10.876655    15.165903
;     5     5.031858     1.609554     5.283016    12.847443     9.700179    16.098145

I assume this is due to a large variance in the fit?

My statistic correlator optimizations behave much better:
;Convergence quality report:
;Loop number, Avg x, Avg y, Norm Avg xy, Stdev x, Stdev y, Norm Stdev xy
;     1     1.271418     1.848821     2.243801     0.832150     0.595195     1.023099
;     2    -0.050483     0.045001     0.067629     0.710341     0.622665     0.944614
;     3     0.095880    -0.075135     0.121813     0.758313     0.545650     0.934223
;     4    -0.025789     0.058485     0.063918     0.706566     0.603806     0.929418
;     5     0.042342     0.013991     0.044593     0.692383     0.605555     0.919832

4

Re: distorted images within GCPS optimization window

Hi Joel,

It looks like the optimization diverges when using the frequency correlation on the shaded DEM. Although it sometimes gives good results, the statistical correlation tend to be more robust when GCP are optimized with respect to the shaded DEM. It's however weird that the optimization completely diverges. Are the GCPs well spread in your image? If they only cover a small part of the image or if they are aligned, that could explain why.

Sebastien

5

Re: distorted images within GCPS optimization window

Hi Joel,

When I look at the grouping of your GCP from the previous posts, your GCP are incredibly close from each other. Ideally, you'd like them to cover the whole image or you can't constrain the global geometry of your image.

Sebastien

6

Re: distorted images within GCPS optimization window

I picked my GCP close together (within the aerial image) because I have a rather small spatial extent for the my DEM.  II thought it would be best to have the most accuracy around my area of interest, because I don't care much if the other parts of the photo are not correctly ortho-rectified. 

I have an ~2 km LiDAR swath running along the San Andreas Fault.  I could use the full extent of the swath to pick GCPs over a larger area of the image, but if I want to have GCPs covering the entire image, I would need to combine the LiDAR with a 10 m dem (or perhaps attempt to make a higher resolution DEM using my aerial photos in sequence as stereo pairs).  Do you have a recommendation?

Thanks,
Joel

7

Re: distorted images within GCPS optimization window

Given the circumstances, I would suggest picking GCP as spread out as possible, within the tile of your LIDAR DEM.

Sebastien