Dear Colleagues,
i am a bit indecisive about choosing a convenient, though accurate implementation to fit two rasters, consisting of Canopy Height Models from different 3D data sources, across a closed canopy forest landscape. The hitherto tested experiences returned pretty mixed results.
I appreciate any hints on this.
Cheers, Hooman
Sood, A. K., Ohdar, R. K., & Mahapatra, S. S. (2010). Parametric appraisal of mechanical property of fused deposition modelling processed parts. Materials & Design, 31(1), 287-295.
Kovalerchuk, B., Doucette, P., Seedahmed, G., Brigantic, R., Kovalerchuk, M., & Graff, B. (2008, May). Automated vector-to-raster image registration. In Algorithms and technologies for multispectral, hyperspectral, and ultraspectral imagery XIV(Vol. 6966, p. 69660W). International Society for Optics and Photonics.
Sood, A. K., Ohdar, R. K., & Mahapatra, S. S. (2010). Parametric appraisal of mechanical property of fused deposition modelling processed parts. Materials & Design, 31(1), 287-295.
Kovalerchuk, B., Doucette, P., Seedahmed, G., Brigantic, R., Kovalerchuk, M., & Graff, B. (2008, May). Automated vector-to-raster image registration. In Algorithms and technologies for multispectral, hyperspectral, and ultraspectral imagery XIV(Vol. 6966, p. 69660W). International Society for Optics and Photonics.
Thanks Sobhan Sarkar for the hints. However, I am rather looking for ready-to-use implementations rather than technical papers that introduce algorithmic workflows. These could embrace a range from r or python scripts to plugins in open source or even commercial domains for raster-to-raster (actually DEM/DSM to DEM/DSM fitting). Could you (or other colleagues) help on this? Thnx
Hi Hooman,
not sure if I got the question right or if this has been solved already, however, if your problem is about fine registration of two (roughly) co-located rasters, the function coregisterImages() in R package RStoolbox comes to my mind. It will perform a simple linear translation in x-y direction. Another solution might be otbcli_FineRegistration in Orfeo Toolbox (Open Source), which allows for a more complex transform, as it will "locally warp" one raster to match another. Hope that helps, if you already found another way I'd be curious to learn more about it!
Best regards,
Sebastian
https://www.rdocumentation.org/packages/RStoolbox/versions/0.2.3/topics/coregisterImages
https://www.orfeo-toolbox.org/CookBook/Applications/app_FineRegistration.html?highlight=fine%20registration
P.S.:
Last time I used otbcli_FineRegistration it had a minor bug. You had to run it twice to actually warp an image. First calculate the disparity image and only after that run again specifying the image to warp (using -w flag) and an output name (-wo flag)...
E.g. the workaround would look something like this:
otbcli_FineRegistration -ref StereoFixed.png -sec StereoMoving.png -out FineRegistration.tif
otbcli_FineRegistration -ref StereoFixed.png -sec StereoMoving.png -out FineRegistration.tif -w StereoMoving.png -wo MatchedOut.png
Great, Sebastian. Thanks a lot for the comprehensive reply. At the first glance it seems to me like the OTB solution might fit better to our problem, which is registering two CHMs to eachother. I will give it a try :)
Thanks again!
Hi Hooman,
maybe a simple systematic shifting (in all directions and by different number of pixels) between the two images and calculating a simple metric like the RMSE between the pixel values of both images (for each examined shift and then search for the minimal error) might be another option.
We implemented something similar to match UAV images to Sentinel-2 data and it worked farely well.
What I don't fully understand is the purpose for matching two DEMs from different sources. Are they supposed to show the same information?
Hi Fabian Ewald Fassnacht , thanks for the useful hint. The main idea is pretty simple: The are a set of Phantom UAV imagery recorded solely with an on-board GPS-IMU, and the data lacks decent accurate DGPS measurements for the external orientation. It needs to be fitted into a LiDAR-based DSM, in order to increase its comparability with other data sources within the study (including a fixed wing UAV and a LiDAR ).
Hi Hooman,
The most robust way would probably to fit the point clouds and not the rasters. This also allows you to better filter or adjust what you actually want to fit, e.g. only ground points/returns.
There are plenty of free and automized options to do so, e.g. Lastools or Cloud Compare.
The most robust approach would be to extract GCPs from the LiDAR data (corners, etc...) and use these GCPs during the SFM processing (e.g. Agisoft). Therby you minimize the mismatch caused by disstortions of the photogrammetric bundle adjustment.
Cheers, Teja
Teja Kattenborn : Taking GCPs from ALS data in forest conditions is a real challenge, since finding keypoints like corners etc is mostly not feasible :(.
To my experience its actually quite trivial. Take for instance the tree tops of some very distinct trees and feed their coordinates (x,y,z) in the SFM project. You can even use pre-aligned SFM data and use the "reference trees" derrived from the LiDAR data for a subsequent optimization of the bundle adjustment (this way you can directly identify the trees in the SFM project that were derived from the LiDAR data).
If this is too much work (takes maybe 10 Minutes per UAV flight and ~6 GCP, excluding processing time), then you could still try the second option mentioned above: i.e. use automatic point cloud matching.
In my opinion these options are much more robust, than the raster-based approach, since the latter does not really allow to account for a shift in orientation/rotation (which is very likely to occur for standard GPS-based UAV data).
Cheers, Teja
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Computer Graphics and Animation
- 3D