The justification in using Positive and Unlabeled (PU) Learning for Satellite Imageries is the high cost of labeling the pixels and that when we want to extract only one object, say Road, we just collect labeled data for that class. However, if we could have small number of labeled data for other classes, say the negative class, besides the target class, then we could use less complicated methods such as semi-supervised learning instead of PU leaning. But, I've seen papers working on PU leaning for classification of Satellite Imageries. Since, Remote Sensing is not my main focus, I was wondering if someone could help me and let me know if there is a strong justification in using PU learning for the classification of Satellite Imageries? Thanks.
Traditional supervised learning is inefficient for one-class classification because it requires all classes that occur in the image to be exhaustively assigned labels. Positive and unlabeled learning (PUL) algorithm is used for one-class classifications of two scenes of a high-spatial-resolution aerial images. The PUL algorithm trains a classifier on positive and unlabeled data, estimates the probability that a positive training sample has been labeled, and generates binary predictions for test samples using an adjusted threshold.
Refer the following:
Article An Ensemble of Classifiers Based on Positive and Unlabeled D...
Article A Positive and Unlabeled Learning Algorithm for One-Class Cl...
Two quick points:
1. @Mohammad Eshghi, PU learning IS a semi-supervised method.
2. @Kavitha Pk, PU learning is different from the "one-class classifications" approaches (see one-class SVM for example). PU learning is a binary (binomial) classification task.
The main reasons (justification) of using PU algorithms, as you mentioned Mohammad, is to reduce the high costs of manually labeling training examples and also to leverage the abundance of unlabeled data to provide additional information to the algorithm (which improves final classification performance).
Thank you for your answers Kavitha Pk and Tom Arjannikov
My main concern is that PU learning is so hard in terms of developing a new method and accuracy metric. And, there are debates that even with very small labeled data from other classes (note that it doesn't need to exhaust all the other possible classes), non-PU methods could be developed with high accuracy. Therefore, I was wondering considering that argument, how one could still justify using PU learning for the classification of satellite images?
Thanks
@Mohammad Eshghi, if you have labels from other classes, then you should definitely use them as they will provide better information than what you would get in the PU setting. The PU approaches are designed for those situations when you only have some positively labeled examples and no negative ones.
I don't think you could justify to use PU methods over fully-supervised ones when you have both positive and negative labels. However, if you have few labeled examples of both classes and a large quantity of unlabeled examples, then semi-supervised approaches could help you improve the final classification model. I recommend to use them in addition to (not instead of) the fully supervised methods.
- Badges
- Science topic
- Similar topics
- Papers