Private:MEG-AM W08

  •  My favorite aspect of the session was the way you explain the classifiers. It was very helpful for me to understand it.

W08:

The best classification I got with training areas and ndvi, rabidEye band 4, landsat band 5 and landsat band 8 and the classifier maximum likelyhood. The only problem was with some water in the city where there is no water. I could not get my classification without this water.

classification

 

I use r.kappa from processing tools in qGIS.

I use this site for help:

http://landmap.mimas.ac.uk/index.php/Learning-Materials/Image-Processing-for-GRASS/6.4-Assessing-Training-Data-Quality

I made test area like training area and converted to raster for r.kappa. Then I used this raster and classification to get the kappa index.

The result:

matrix

Cohen’s Kappa index of agreement tests if two people/algorithms would make similar classification.
The value is very bad, but maybe it is because there is class with id 0, and this is not in my
classification. It is from the test area raster. I did not find how to tell qGIS to ignore 0.

 

One thought on “Private:MEG-AM W08

  1. Thomas Nauss

    Where is the Kappa value in the figure? I only see the percentage of correctly observed pixels. Why don’t you compute it by hand after having the contingency matrix ready? This would overcome the “0″ problem and it is quite simple and straight forward once you have the row and column sums.

    The classification looks quite ok from a visual perspective and you have obviously selected quite a variety of training and validation sites which is excellent.

    Cheers,
    TN

Leave a Reply