Effect of training strategy for positive and unlabelled learning classification: test on Landsat imagery

摘要

The positive and unlabelled learning (PUL) algorithm, which transforms the traditional binary classifier into a one-class classifier based on Bayes' rule, has received attention in the field of remote sensing classification. PUL requires positive and unlabelled sample sets for the training procedure. As a one-class classifier, only the positive samples are labelled by the users, whereas the unlabelled samples are generated randomly from all the pixels. However, the effect of the training strategy on PUL has not yet been investigated. This study tested the performance of PUL-SVM (support vector machine) using training samples with different sizes, purity levels and quality levels. It is found that the balanced sizes of the positive sample set and the unlabelled sample set is an optimal strategy if the sets were sampled randomly. However, pure positive samples, which are preferred to be selected for guaranteeing correctness, make determining the optimal sample size difficult. The quality of positive samples is also an important factor that affects classification accuracy. A positive sample set that contains random-error samples drastically reduces classification accuracy, whereas confusion-error samples with spectra similar to those of the target class slightly affect classification accuracy. On the basis of these findings, we recommend the random sampling strategy for positive samples and balanced sizes for positive and unlabelled sample sets. Meanwhile, random-error samples must be avoided, whereas confusion-error samples can be highly tolerated.