Effect of training strategy on PUL-SVM classification for cropland mapping by Landsat imagery

Positive and unlabeled learning (PUL) algorithm, an one-class classifier which is trained by positive samples and unlabeled samples, has been used in remote sensing classification. However, the effect of training strategy of PUL has not been investigated. This study tested the performances of PUL-SVM on cropland mapping by Landsat TM data using the training samples with different sizes and different purity levels. It is found that the highest accuracy is achieved when the sizes of positive sample and unlabeled sample are comparable if using the random strategy. In contrast, if using the purer positive samples, it is more difficult to find the optimal unlabeled sample size. Therefore, it is recommended the random strategy for the positive samples, and the balanced sizes for positive and unlabeled samples when using PUL-SVM.