I have 40 cm resolution aerial image with visible and near-infrared bands. Purpose is to extract dried trees in forest. I used multiresolution segmentation algorithm to extract individual trees and used NDVI to classify dried and healthy trees, but there are clear cut areas in the study area, and they give the same NDVI value as dried trees. I even checked RGB and NIR values of segments and they also give the same value as dried trees. Any suggestions?
-
1Not too sure, if this is the right way to go, but you could search on alternative vegetation indices like SAVI: "The soil-adjusted vegetation index was developed as a modification of the Normalized Difference Vegetation Index to correct for the influence of soil brightness when vegetative cover is low." (src)– ulrichCommented Apr 12, 2016 at 11:23
-
@tareq, thanks for your advice, I tried SAVI, but met the same problem unfortunately.– SherCommented Apr 12, 2016 at 11:39
-
1Do you have any associated elevation data like LIDAR? I would have imagined clear-cut areas would return a different height in first return? You could then distinguish between areas that are returning similar NDVI values for dried trees by their elevation.– HornbyddCommented Apr 12, 2016 at 11:59
-
@Hornbydd, thanks, it is good advice, but I haven't got any elevation data for this image.– SherCommented Apr 12, 2016 at 13:02
-
1How did you "check" the spectra? Did you perform a separability analysis? I have performed this type of modeling on numerous occasions and there has always been separability in standing dead and down woody debris. I would follow @Tom Dilts advice and utilize the texture (focal variance) of the red and NIR bands. You could also explore other segmentation algorithms that use texture as a variance reduction criteria.– Jeffrey EvansCommented Apr 12, 2016 at 18:37
|
Show 2 more comments
2 Answers
If you can't differentiate them spectrally then it has to be based on some other criteria. You state that you are using a "multiresolution segmentation algorithm" which suggests that you might have access to software that can do object-oriented classification. You might want to consider textural, shape, or spatial characteristics that differentiate individual dead tree crowns from clear cuts. Size and crown shape are two differences that come to my mind.
-
thank you for your advices. Yes have access to eGognition software, and created multiresolution segmentation with this software which takes into account, shape and texture, but clear cuts are not so big areas, their size is quite similar to trees, it is difficult to differentiate them with dead trees.– SherCommented Apr 13, 2016 at 9:38
-
I tried different texture information of objects but couldn't see the variation– SherCommented Apr 13, 2016 at 10:19
In addition to the spectral data you would need another way to differentiate the two. I know point cloud data from LiDAR surveys would allow you to do this, but you might not have it available or in the same time frame as your photos. It's a vague hint, but looking for elevation datasets that show a difference between the bare earth DSM and some of the point returns might be your best bet.
