I need to classify Radarsat 2 and Alospalsar data set using Support vector machine. I have the 5 multi component scattering parameters layer stacked individually for Radarsat 2 and AlosPalsar. I have total of eight classes namely Urban, Dense Vegetation, Sparse Vegetation, Barren land, Harvested land, Dry water bodies, Water bodies, Coal mining area. To get the ROI's of barren land, coal mining area, water bodies and dry water bodies are difficult as it gives almost the same backscatter values in the surface scattering component. Are there any ways of classifying the above classes in the mentioned data sets?
Radarsat-2 is C-Band SAR and ALOS is L-Band SAR. Both the bands have their unique advantages and uncertainties. Pre-processing i.e) speckle removal is mandatory.
http://www.indiana.edu/~act/files/publications/2012/12.03.pdf look at this pdf.
After Speckle removal u observe backscattering values in the image (take care so that your texture is not lost). Waterbodies have the least values in terms of -50 to -20 db approx, dried waterbodies are having moisture content with smooth surface that can be observed in L-Band because of little bit of penetration capacity of larger wavelengths. for Urban lands have high Backscattering.
You let me know for what you are using you final product, so that one can suggest accordingly. If you are trying to build automated or semi automated technique you need sample sets.
Simply use ERDAS-Imagine or ENVI software and use Level-1 LULC classification built in; but make sure you deactivate the Anderson classification classes that are not relevant to you terrain.
You did not mention the type of data your are using - amplitude/SLC, polarization etc. Depending upon the data type, you may use Wishart/ Min Distance classifier/rule based classification approach. You could have explored the optical data for reducing the confusion from some of the classes.
I agree with Prof.Gloaguen. If the data is not matched with your objectives, kindly change it. Or try data fusion with optical data.
If you solely use the RADARSAT data, you may add textural features or decomposition methods to improve the classification. There are some methods that can help you for better spectral seperability. You may check the article below:
Effect of Different Target Decomposition Techniques on Classification Accuracy for Polarimetric SAR Data
I also suggest you to analyse the data spectral response and training data characteristic. You may define the methods more proper up to the result of it.
See this article as well.
Haack BN, Khatiwada G; Comparison and integration of optical and quadpolarization radar imagery for land cover/use delineation. J. Appl. Remote Sens. 0001;4(1):043507-043507-16.
http://remotesensing.spiedigitallibrary.org/article.aspx?articleid=707916
- Badges
- Science topic
- Similar topics
- Geoscience
- Remote Sensing