In airborne hyperspectral (1m), can signatures of target objects be extracted from training pixel instead of field campaign? This is to reduce cost; but does this affect information extraction performance? The application is tree species identification in mixed urban and rural environment.
Hi Abdelmounaime,
This is an excellent and interesting question raised. This is a broad topic as the exact application is not mentioned. There are many research articles based on either of the approaches.
First of all, the success will depend on how 'pure' your 1m pixels are. You should be able to find sufficient number of 'pure' pixels corresponding to all the LULC classes. This will also require a very good knowledge of the study area and good quality ground truth information (for selecting training pixels).
I prefer image derived spectra for forests as these spectra represent the top-of-canopy conditions whereas in most cases field campaigns collect spectra of leaf. So you will not encounter the uncertainties related scaling behavior of spectra.
One can also use the standard spectral libraries (for example from USGS) instead of field campaigns. But many of the land surface features (for example vegetation) may have different spectral properties in different geographical areas.
I expect a large number of responses to this question and hope to see more experiences shared!
-Ani
The unsupervised classifications are tacitly built this way.
In case you intend to develop your own algorithm, try to select also windows of more pixels than one, for example 3x3 (it depends on the crown spread) and analyze the results as respect to the "one-pixel" training.
[ Reason: the spatial resolution of 1 m is too close to the target size, so the identification results may be affected if, for example, a major part of some training pixels is not covered by the supposed trees; the idea is to size a window for training so that it is mostly covered by the crown of the tree to be identified].
I agree with Aniruddha and Ruxandra,
YES abdo, Since you are looking to IDENTIFY your target species, in this case the field work is "highly required". on the other hand, you can go on by using JUST your airborne data and to direct your research to identify the limitations airborne data for species identification without field campaign, which would be great as well.
i suggest for you to see what CSIRO are doing on this issue as well
regards
I aggree with Aniruddha, the benefit of using field spectra is qustionable in case of vegetation cover, because all pixels are mixed ones in this case. Most importat is: the spatial and frequency sampling should be the same for image and field, if field spectra will be used, and accurate atmospheric correction should be applied. If image based reference spectra will be derived, very accuraate geocoding is necessary in turn. If target size is large, the reference pixel size should be increased, so you shold lower image spatial resolution for superwised classification. I applied fire ladder to collect field spectra for Ambrosia Artemisia identificaton. Leaf spectra are useless. Confusion matrix evaluation tells you which of the choises are best. With DGPS you are able to choos several test reference sites/spectra. Using the you will be suprised, how big differences in classification results arose. So be patient, and do experiments with image based superwised classification on a small test area? choose the best one, and appply it to a larger region. Choose the best fenological state for the target tree, when the biggest difference can be eperienced.
Unfortunately, I can only refer to my article in Hungarian, but the images are self-informative.
A parlagfű detektálás távérzékelési módszereinek vizsgálata. I. rész, Magyar Gyomkutatás és Technológia (Hungarian Weed Research and
Technology) Vol.VI. No 2. pp. 54-67.
I'll upload it on ResearchGate.
- Badges
- Science topic
- Similar topics
- Geoscience
- Atmospheric Sciences
- Atmosphere