When performing supervised imaged classification using machine learning classifiers such as random forest or SVMs, there is a possibility of using point or polygon training data. In the context of pixel-based classification, I find that point data make more sense since a polygon will cover many pixels and confuse the classifier even if the researcher makes an effort to select training sample polygons that are as homogeneous as possible.
In my opinion, I think polygon training samples should be reserved for object-based classification. What are your opinions on this?
Fundamentally, it depends on the representation of your feature vector that represents the attributes that need to be classified. As such, there is no difference between points and polygons. The only difference may come at the density of the attributes that will be extracted from points or polygons. Indeed, the density will make a big difference since it riches the training samples for the positive and negative examples and indirectly it helps to optimize the classification surface or boundary.
Dear Faustin,
You may consider excellent explanation by Dr Gamal Seedahmed., ie., is enough.
I use polygons when I am interested in incorporating the values of neighboring cells.
Do you mean resolved (object radiates on multiple pixels) vs unresolved (object radiates on single pixel) image classification?
Dear Faustin Gashakamba,
First, I partially understand your question. Please be careful when asking the question.
When we do supervised image classification, in the selection of training data, we are offered the following options:
1. Draw Polygon
2. Draw Rectangle
3. Draw Circle
I partially agree with Professor Gamal Seedahmed's answer. I can not understand how in the case when we do supervised classification there is no difference between whether we choose the circle or the polygon in the training selection of data! The professor himself is claiming that density plays an extraordinary role. So, how can the selection of the circle (point as you call it) or the polygon not play a role – referring density?!
I recommend that depending on the pixels you want to cover, use the circle or polygon – as it suits you best.
All the best,
Albert Berila
Thank you Albert Berila for your contribution to this discussion.
Actually, when I said "point" I literally meant it; not a circle as you say.
I use R's caret package and even when you use polygons (circles, squares or irregular shape), you can still extract ground-truthed values at pixel level and use that as your model training data.
Single point can be more accurate when the polygon usually cover more unwanted points.
- Badges
- Science topic