How can I achieve high quality supervised classification?

Question

For the supervised classification I use: QGIS, Semi-Automatic-Classification Plugin, a Landsat-8 Satallite Image and the CORINE Land Cover classes (http://uls.eionet.europa.eu/CLC2000/classes/index_html)

My goal is to perfom a high quality classification.

So far I know it is important to have many reference-areas with high correlation between actual surface and created class. Sounds easy! But it isn't! Because the entire landscape doesn't fit into just some classes.

For example: What is the "better" way to go to classify urban areas? Creating one large polygon and ignoring some green areas or creating many small/tiny polygons but therefor having less different pixels inside the polygon.

;

Also how do I figure out how many classes are needed? So far I decided to perform a level-1 CORINE classification, which means 5 classes. What else is important to know in terms of a high quality classification?

Answer 1

Most (if not all) of the classification methodologies available in the plugin are pixel-based classifications. As such, it takes each pixel in the image and compares that to those in your training areas. With this in mind, the big polygon gives it a much more nebulous set of information to compare with, while the small polygons give condensed information.
Consider the case with the big polygon - several of the pixels that you have told the methodology to classify as "Urban" are thoroughly "Vegetation" pixels. While most classification methods are able to deal with some noise in their training data, there is no reason to willfully mislead the methods. You don't want the classification to actually classify those trees (and all the other trees in your image) as urban areas. As such, you should in general give your classification method clear and pure training areas, while potentially considering going for a soft classification, where the result is percentage chance that a given pixel is in a given class, rather than a hard classification, which simply gives you the output class.

It should be noted that everything above relates specifically to pixel-based classifications and not object-oriented classifications. When working on the object level, the rules change significantly and it becomes a lot harder to make general rules of thumb.