I need efficient methods which increase the accuracy of the classification of medium resolution hyperspectral images (between 17 and 30 meter) based on the collected spectral signatures from ASD FieldSpec 4 Hi-Res Portable Spectroradiometer.
I used the methods available in ENVI such as spectral angle mapper, spectral information divergence, SVM...etc. But they are not efficient enough to separate different features in the image. Sometime improving spatial resolution of the images to 10 m is not showing significant improvement
Could you elaborate on what are you mapping, such as crops, grasses, semi-vegetation etc? If you can describe your study area then may be some suggestions are possible.
Well, it really depends on which features you want to discriminate. If you properly captured them with the ASD. If the sampling at the field was the adequate for the pixel size of the images. And more importantly, how good are the radiometric calibrations of both ground and airborne instruments (or, the atmospheric correction in case you work in reflectance).
In any case, Abdulwasey Mohammed is right, you should provide more details in order to provide a proper answer.
hi M Awad,
in case this is useful,
we published a method that used hyperspectral reflectance of forest (and shrub and mixed ground covers) canopies collected with a hand held radiometer from a helicopter platform to both provide a high accuracy atmospheric correction of satellite Hyperion data as well as create and validate (with nadir photography) characteristic spectral (or endmembers) calculated from the site-specific canopy reflectance datasets (both helicopter radiometer and atmospherically corrected Hyperion data). To identify the underlying charateristic spectra we used a software that did not require known spectra as inputs. the PVA characteristic spectral were used to classify the entire Hyperion image. if interested, you can get the set of 3 papers published in IJRS in 2005 by authors Ramsey et al....
Dear Dr. Abdulwasey, Dr. Alonso and Dr. Gloaguen
Hi, sure I can provide more details. The satellite hyperspectral images are of medium resolution. One of them covers the wavelength from Visible to SWIR , the other from Visible to NIR 400nm 1050 nm. The spatial resolution of both can vary between 17 m and 30 m. However, the resolution of one them can be enhanced to 10 m.
What I need is to be able to separate different forest species in a mixed sometime few pixels forest area. This applies to the same crop type I want to map different type of crops such as potato, wheat....etc. The problem is that sometime the area is less than a 1 ha and sometime it is 0.1 hectare what ASD FieldSpec 4 Hi-Res Portable Spectroradiomete can do to help me map forests species and crops correctly?
How to take spectral signature samples? what can cause errors in collecting these samples?
I tried to compare the end members created by SMACC in ENVI with the collected ASD FieldSpec 4 Hi-Res Portable Spectroradiomete spectral signatures but there is still some uncertainty in the results.
Dear Dr, Elijah Ramsey III,
Thank you sure I can read these papers please send them.
Mr. Awad, may be you can try spectral unmixing as you have strong mixed pixels scenario. I am suggesting alternate method that is based on Object based fuzzy logic classification using membership functions wherein you can have shape index, textural information and your spectral library data collected from field, in a hierarchical tree based classification as sub objects and super objects. This would give you a better segmentation than any spectral classifier. There's a ton of literature on Object based classification techniques.
In your case, should you have temporal resolution of your study area, then phenology, along with other ancillary information would have been a very useful metric to discern wheat from potatoes and grasses. However, I guess you lack any temporal resolution, so give a try with object based techniques.
Thank You Dr. Elijah for sending the papers. I will investigate them for a solution for the problem.
Dear Dr. AbdelWassy,
thank you for your concern. Sure there are plenty of segmentation algorithm which works on object or pixel based. Some of them use fuzzy logic, ANN or other type of parametric or non-parametric methods. These are good for both unsupervised multispectral and hyperspectral images in this case Spectral signatures or any other field based collection of data can be used for verification. But, the quality of the results again depends on the algorithm, spatial resolution, the quality of image and the accuracy of the field collected data.
How I can avoid the above problems by simply using supervised classification ( user controlled) based on improved spatial resolution and corrected image with spectral signatures based on ASD spectroradiometer ?
Please read Chapter 7 in Signal Theory Methods in MS RS by David A. Landgrebe, may be you will find your answer.
- Badges
- Science topic
- Similar topics
- Geoscience
- Atmospheric Sciences
- Atmosphere