Satellite imagery: Worldview-2 and Pleiades (MS = 2 m GSD, PAN = 0.5 m GSD)
Software: eCognition
Questions in detail:
1. Which GLCM is robust enough for a transfer in between satellite imagery, that are different in sensor type and varying in aquisition time, and indicates for forest structure? This means the method for forest structure analysis should not be based on a trial & error approach: e.g. training and ground truthing on each satellite imagery.
2. More simply: like 1., but not different in sensor type, e.g. satellite imagery just varying in aquisition time (e.g. WV-2)
3. Which pixel size is appropriate for forest structure analysis using texture measurements? using MS = 2 m, PAN = 0.5, or MS pansharped 0.5 m?
4. Which spectral band is appropriate for forest structure analysis using texture measurements? e.g. PAN, coastal, blue, green, yellow, red, NIR1, NIR2 (WV-2); PAN, blue, green, red, NIR (Pleaides); or better to use vegetation indices like NDVI, Simple ration (SR), EVI,...
5. Technical question: How to fasten GLCM calculation within eCognition?
From my experience, the GLCM features in eCognition are very consuming in hardware resources in general. Especially for larger extents, such as complete satellite scenes, this is hardly feasible.
Regarding question 5: There are also GLCM (quick 8/11) features. Those are much faster, than using the complete ones.
I am also curious on upcoming replies from the colleagues as GLCM features require quite some experience.
I also observed slow computations every time I tried them. I suggest to use Monteverdi application (http://www.orfeo-toolbox.org/otb/monteverdi.html). It's a free and open-source software. You can try different GLCM textures on a small area (fast computation) and then compute on the entire of your image (batch processing). Finally, you follow the suggestion of Philip Gartner to import back into eCognition.
I recommend running the GLCM on the NIR band or PAN band since you are doing forest analysis. NIR band will probably run quicker and I recommend it if you're looking at forest stands, but Pan may be better if you are interested in individual trees. GLCM in other software like ENVI or Monteverdi will be calculated in a moving window, not at the object level, so that's something you may want to think about.
Basically, I think the choice of which GLCM to use is not going to have a huge impact on your study. You may want to just pick a few and just check the descriptions/equations to make sure they are not too correlated.
To find out, which combination of GLCM is suitable - One could run a Feature Space Optimization on a small subset. Path to FSO: eCognition >> Classification >> Nearest Neighbor >> Feature Space Optimization.
I can't respond for ecognition software. We have used the orfeo toolbox implementation (http://www.orfeo-toolbox.org) for GLCM descriptors. The GLCM parametrization is complex and application dependant. In this paper, we have run an exhaustive research, with an automatic parametrisation for retrieving physical forest variables from a Quickbird image.
Results show that a multiscale approach combining pan et MS GLCM feature with different parametrizations gives the best results.
BEGUET, B. , CHEHATA, N. , BOUKIR, S. , GUYON, D., 2012, Retrieving forest structure variables from Very High Resolution satellite images using an automatic method. ISPRS'2012, XXII International Society for Photogrammetry & Remote Sensing Congress, ISPRS Annals of Photogrammetry, Remote Sensing and the Spatial Information Sciences, 25 août - 1er Septembre, Vol. I-7, pp. 1-6, Melbourne, Australie.
http://www.isprs-ann-photogramm-remote-sens-spatial-inf-sci.net/I-7/1/2012/isprsannals-I-7-1-2012.pdf
Regarding question 5: You need at least 16GB of RAM (see discussion at http://community.ecognition.com)
- Badges
- Science topic
- Similar topics
- Geoscience
- Remote Sensing