Topic: Blurred Resampled Images

Following is a message from Duncan:

I am having difficulty orthorectifying a couple of SPOT IV images at the moment. My GCPs are nicely optimized, and I have a good quality DEM (20m from SPOT), but the resampled image is quite blurred when compared to the original – see attached. This has obvious implications for my next step of correlation.

Does anything spring to mind that might be the cause of this?

By the way I should also say that I have rectified two SPOT V images in exactly the same manner as I do the SPOT IV, and they are co-registered to better than a pixel and the detail is retained perfectly.

Any pointers greatly appreciated!

Best wishes,


My answer:

Dear Duncan, you are experiencing a classical issue in COSI-Corr, but don't worry, there is a way around it. First, the most likely reason for this pb:

The orthorectification mapping generates an association of coordinates between raw image coordinates and ground coordinates. The local deformation and irregularities of this mapping depend primarily of the local perturbation and slope variations of the topography. However, for simplicity, the resampling kernel used does not vary over a given image. This means that we are using a kernel that takes into account the worst deformation possible in the image to resample it (we have an adaptive version of the resampling overcoming this approximation, but we haven't released yet, it will be included in the next release in a few months). Thus, if the topography of your scene has areas of very steep gradients, and areas of flat topography, high frequencies of the image will be lost on areas of flat topography.

However, in most practical cases, the frequency content variations in a medium resolution image are negligible. In your case, because your image (10m) and your DEM (20m) do not have a very high resolution, it is very likely that steep topography gradients are introduced by outliers in your DEM.

The maximum distortion in the ortho-rectification mapping, hence the image frequency content, is determined by the variables dx and dy that are recorded in the header file of your ortho-image. If you edit the header file of your ortho-image, in the description section, you will find a line that looks like this:
"Kernel: Sinc Sinc Size : 25 Distance Max dx=1.2937 dy=1.1723"
Please post this line.

In your case, you must have fairly large dx and dy values. What are these values for your SPOT 5 image?

The possible fixes, by order of likeliness:

1 - Check that you don't have any missing values or outliers in your DEM. In ENVI, you can open your DEM file in a display, then right click on it and "Quick Stat". The min-max values will tell you if all elevations are within a physical range, and they should be. If they're not, they should be replaced and interpolated to be in a physical range. When this is fixed, the ortho + resampling should be recomputed.

2- If no outliers or missing values are detected in the DEM, it is possible to compute de resampling distances dx and dy only on a selected subset of the mapping matrices. You can also estimate them using the tool "matrices interdistances" and force them manually during resampling. Please refer to the COSI-Corr manual for these functions, or we can discuss it here later if needed.

3- You will need an adaptive resampling kernel. In your case, it shouldn't be necessary unless your image has a large incidence angle. We could potentially send you an early beta version of the next COSI-Corr release if this is what you need.

Let us know,



Re: Blurred Resampled Images

Dear Sebastien,

Thanks for your detailed reply - it's much appreciated.

First things first - the dx and dy values for the blurred SPOT IV image I sent you were as follows:

Kernel: Sinc Sinc Size : 25 Distance Max dx=3.2705dy=1.1566

Indicating, as you suggest, that the distortion is high.

(Incidentally, the values for the SPOT V image (5m) that rectified without problems are dx=1.4931dy=1.2680 - much better!)

For your possible fixes:

1. My DEM values range between 2100 and 7700m - I have already trimmed the border off the DEM to get rid of some null and negative values and reset the xstart and ystart to 1 and 1 so I think it should be ok...

2. I have subset the area over which the resampling distances are computed to a very small area of the DEM - that resulted in much lower values (dx=1.7503 dy=1.1485) and produced a much finer image. But I guess the question then is over which area should the distances be computed (low topography/high topography/varied), and if I fix the distances during resampling, how is it best to arrive at dx and dy - is it simply the lower the better or should it be guided by the automatically computed values?

3. The incidence angle of the image is 26.9 degs. Hopefully I shouldn't need your beta version (based on 2. above), but I appreciate your willingness to offer this.

Thanks again!



Re: Blurred Resampled Images

Dear Duncan,

The resampling distances should be computed in areas were you need the correlation to be the most accurate, I would say over an area that encompasses the glaciers of interest. Regarding dx, dy, we cannot say the smaller the better. If they need to be large, they should be large, but by experience, large dx and dy are often produced by DEM inconsistencies.

In your case, you're working with an image that has a high incidence angle, with distortion mostly across-track, and the pb seem to be on the dx distance, which is consistent with a large incidence angle. Maybe you're simply encountering one of the limitation due to a fixed kernel size, the reason we implemented a locally adaptive version. If the resampling distances are still too large when defined only over your area of interest, we'll send you a beta version of the adaptive resampling (it's implementation is not optimized yet and is a little slow).

The SPOT 5 image has more reasonable dx and dy. Is it a close to nadir acquisition?

If you're correlating images with large incidence angle differences, the correlation will be sensitive to the glacier elevation change between acquisitions, unless you possess a DEM for each acquisition. Otherwise, the horizontal vector measured will be a linear combination of the horizontal and vertical displacements. It's something to keep in mind when interpreting the results. Any vertical change will introduce some bias along the epipolar direction, which is mostly oriented in the EW direction for SPOT images.



Re: Blurred Resampled Images

Dear Sebastien,

Thanks for another very useful response.

I'm understanding what is going on here now. Obviously, using lower dx and dy (either by selecting a small "smooth" area for computation, or simply by forcing the values prior to resampling) produces good results in flat areas (e.g. the glacier surfaces) where distortion is low, but does not adequately compensate for the high distortion in steep areas - hence the need for an adaptive kernel. This is exactly the case with my image - using low dx/dy the 10m data are resampled almost perfectly over the glacier but the hill-tops/mountain slopes are very messy. I can live with this if necessary, as I mainly want to derive glacier velocity (it doesn't matter too much about the steep slopes in this instance), but if you would be willing to release your adaptive sampling version to me that would be great - a fully corrected image would clearly be preferable...

Yes, the SPOT V data are almost nadir - 1 or 2 degs off. I have another SPOT V image (2.5m) that is 13 degs from nadir and another SPOT IV image (10m) that is 20 degs from nadir, in both cases I'm getting quite high dx/dy and therefore an element of blurring over the glacier surfaces - clearly the incidence angle is critical in areas of extreme topography.

Thanks for the tip re. glacier elevation changes - not something I had really thought about but very valid (although the surface elevation changes should be very low on these glaciers for any given 12 month period).



Re: Blurred Resampled Images

Hi Duncan,

We'll try to include a preliminary version of the adaptive resampling over the next month, we'll let you know. As for now, you can certainly obtain good results defining your resampling distances over the areas of interest.