Supervised classification

[sbht]

Hi everyone,

I need to that how many pixel must collection for each of classes in supervised classification?

Edited June 1, 2013 by sbht

[HTC]

the more the better!!!! :D

[rahmansunbeam]

the more the better!!!! :D

theoretically it works, but in practice the result can be distracting.

Usually we do not take pixels as training sites, but group of pixels together. When the software converts training sites to signatures, it computes the statistics (e.g., mean DN) of the spectral responses within each training class and within each band. This statistics happens within sites and neighboring pixels. If you take too much of training sites for class A than B, the overall statistics will be biased towards A. How is that?

Consider class A has a mean DN ranges from 6 to 8, class B has 10-12. Too much variation among A can push the range to 4-10, making class B sit between 11-12... means confusing class boundaries and misplacement of class definitions (mixed classification).

So for optimum or near-optimum result, use equal number of training sites for all class, with same/similar number of pixels in each.

[Arhanghelul]

theoretically it works, but in practice the result can be distracting.

Usually we do not take pixels as training sites, but group of pixels together. When the software converts training sites to signatures, it computes the statistics (e.g., mean DN) of the spectral responses within each training class and within each band. This statistics happens within sites and neighboring pixels. If you take too much of training sites for class A than B, the overall statistics will be biased towards A. How is that?

Consider class A has a mean DN ranges from 6 to 8, class B has 10-12. Too much variation among A can push the range to 4-10, making class B sit between 11-12... means confusing class boundaries and misplacement of class definitions (mixed classification).

So for optimum or near-optimum result, use equal number of training sites for all class, with same/similar number of pixels in each.

And still you will have some confusion, due to similar spectral responses, especially when you whant to extract more classes. Classes that have a very similar spectral response will create the most error confusiion classification.

A solution would be to use an OBIA software.

OBIA moderates the confusions, by creating and classifing image objects (group of pixels with similar spectral properties) and not individual pixels ( like in ERDAS or ENVI).

But sometimes, I find more better to use the single-pixel classification than the OBIA, for some satellite images. For example, LANDSAT images. They have a low-medium resolution (30 m). I think that using OBIA for classification is too much.

It would be ok to use a per-pixel classification (ENVI-ERDAS) for this tipe of low-medium resolution images.

This is my opinion.

OBIA for high-resolution images with many classes and per-pixel (ENVI-ERDAS) classification for low-medium resolution images with only a few classes ( too many classes will create confusion > depends on the classification methods: Maximum Likelihood or something more advance like a hybrid method would be apropiate)

My opinion....

Edited June 4, 2013 by Arhanghelul

[jj_kan]

I can to offer the example of of practice experience results. RS Chief specialists have recommend in classification Landsat data use the principle: 30 - 50 training plots (AOI or ROI) for 1 class.  Minimum size one plot - 3x3 pixels (1 ha).