In some studies, different numbers of pixels per class were used for training. Therefore, the training data are unbalanced. However, many studies have shown that unbalanced training sets do bias the results. Should we use equal number of pixel per class for training everytime on image classification? Any answer will be appreciated!
Thank you for your answer Michel.Would you please give us more specific/precise example of it within remote sensing image classification ? I already read the articles on it.
Mustafa, try transforming pixels to another domain where number of pixels do not matter. For instance, you can try taking a histogram.
Dear Mustafa,
Simple (and dummy) approach, that works depending on characteristics of the data, is to normalize the samples (here - rescale image). That usually means loose of the information but sometimes it is acceptable (mind that resizing of the images may use different algorithms that provide different results, i.e. Lanchos is useful when considering texts while bilinear/bicubic in case of the landscapes).
Other method is to extract invariant property vectors (i.e. histogram as proposed Mr. Osman Topcu). There are many possibilities here and it is really hard to provide direct and appropriate solution without more details and deeper knowledge on the data you're using.
Hope it helps,
Piotr
Hi Mustafa, have you ever tried to assign a weight factor to each class? Given the entire amount of pixels of all training samples, you could say that weight of class A is inversely proportional to the amount of pixels provided by class 1 in training set and so on.
Thank you all for your answer.
@Dear Tiago....Do you mean weights for each class on classification or training data?
@Dear Piotr I am working on optical imagery for image classification. Sure it would be depending on the data and classification method.
@Dear Osman Thank you for your answer. I am looking at it.
Hi,
if I've understood your problem correctly.
Divide the images to equal parts, and then extract equal number of features (e.g. statistical features) from each parts, and finally concatenate all features, can be useful in some cases.
Hi there!
You may try to augment training data (through simple copying or copying with transformation, e.g. adding noise) so to level (balance) data training data for each class. Another way is to modify your loss function and add larger weights to classes with fewer training samples. These techniques usually works for neural networks.
- Badges
- Science topic
- Similar topics
- Geoscience
- Remote Sensing