Hi,
While using the HySpex camera, I am facing the problem of Keystone distortion across the scan line. Although I am using the camera for close range imaging, this is a well-known problem in remote sensing community and therefore, I will appreciate if someone can help me to correct it. I have the sensor model file as well, but I am not sure how to use it for correction...
Here the key question is if the keystone is or not the same for all your bands.
If it is, then it is just a geometrical problem, where you reproject your pixels to the proper position.
If its not, then you will have a problem of difficult solution, since the first bands of your spectrum will belong to a different point in space than the last ones. This will affect most near the edges of your objects.
What is your case?
@Luis : The distortion is same for all bands and the problem is now somehow solved by reprojecting the pixels. However, since it is kind of rescaling, I am not sure if it will create some change at the pixel level (for example mixing of spectra). What is your experience with such processing effects?
you might get changes in pixel's footprint, from one portion of the swath to another, but that is also happening because of the view geometry. So, it shouldn't bother you (even without keystone, the larger the FOV the larger the change in footprint).
If the effect is the same for all bands, it means that every band from a given pixel belongs to the same portion of the imaged object. So, self consistent, and the only remaining problem would be to find the proper spatial location, that, as you already mention is achieved by reprojection.
There might be spectral distortions if you do resampling, e.g. during rectification (i.e. resampling to an orthogonal grid, such as a map projection). But this problem is also present even without keystone.
My recommendation for any spectral imaging is to work always at sensor geometry (processing and analysis) as long as your algorithm does not require world geometry. In this way you always preserve spectral information. Then, once you have obtained your product, you can safely resample to whatever geometric grid you need.
Hope this helps, if still some doubts, let me know.
@Luis : Thank you so much for providing the explanation. I will work on that and will contact you in case of some doubt. :)
After the investigation on images, we found that the distortion is not keystone, and it was the initial misunderstanding about the type of distortion that we saw. In fact, there is variation in pixel size while going from one direction of line scan to the other (x-axis). This distortion is corrected by using the sensor model being provided by HySpex. There is another distortion present in the y-axis as well, where the squares in a grid are slightly altered an they appear like parallelograms. With affine transformation, this issue is also addressed.
Meanwhile, HySpex has created software for post-processing of these images and it will be available soon after tests. So the issue is solved.
Thanks :)
I noticed this post and would like to make some clarifying remarks from HySpex. As Haris correctly explained in his last post, the initial issue he raised had nothing to do with keystone (which is very small in our cameras….typically less than +\- 10 % of a pixel), but rather a variation in pixel size across the FOV which is normal to a large or small degree in any camera. The described variation in pixel size is characterized in great detail during production and these data (the sensor model) are supplied with the camera (along with all the other calibration data, detailed test report, etc), allowing the user to manipulate the data and sampling grid as preferred.
It should be straightforward to resample the images to get equal pixel size, e.g. in MatLab (by using the sensor model data).
So far, no-one else in our customer base has requested a dedicated SW tool to do this kind of resampling, so writing a MatLab script was what we originally proposed by email before the question was posted. Since we heard nothing after suggesting that, we assumed the problem had been solved and were a bit surprised that it was brought up in this forum, and even more surprised that it referred to a keystone effect.
For our airborne users, the sensor model has always been used directly in the georeferencing SW (e.g. PARGE) to georeference the data and resample to the chosen map grid, but the PARGE is not directly suitable for lab applications.
Based on Haris’ need, we have now developed a simple resampling SW tool for this purpose to use for lab measurements.
But we fully support Luis Alonso's approach as explained above....better to resample the processed data to the preferred grid after the data analysis, in order to preserve the best quality and full integrity of the data you are analyzing.
Regarding the “distortion” mentioned in the last post, this is could be due to a combination of sub-optimal acquisition settings (e.g. speed) and that the square samples were not aligned with the scan axis (which makes the object look skewed) during the scan. The camera could also be mounted with the linear FOV not being perpendicular to the scan axis, and/or the sample could have some height different from one end to the other. It seems like it has been elegantly solved in SW by Haris, but it is of course better to take care of this during the acquisition!
- Badges
- Science topic
- Similar topics
- Geoscience
- Remote Sensing