Perturbed peer model with joint confidence for semi-supervised object-based classification in urban area

Abstract

ABSTRACT The application of deep learning (DL) improves the accuracy of object-based classification in urban area, but the huge numbers of labelled samples required for training DL models are difficult to obtain. To address that, we propose a novel semi-supervised object-based classification method that combines pseudo-labelling method and consistency regularization method, named as perturbed peer model (PPM). To evaluate the quality of unlabelled object samples, a new object-level joint confidence is designed to assess the confidence of the samples’ prediction, which provides guidance for selecting unlabelled object samples in semi-supervised object-based classification. Instead of only using the high-confidence samples as usual, both the high-confidence samples with discriminative power and the low-confidence samples with a large range of structural features are exploited using pseudo-labelling method and consistency regularization method, respectively, facilitating the fusion of the two methods in the proposed PPM. In addition, two types of perturbations, dropout and difference augmentation, are integrated into the model to drive the consistency regularization method as well as to enhance the difference to facilitate the pseudo-labelling method. Experimental results show that the classification accuracy of PPM is better than the widely-used self-training, co-training, noisy-student, unsupervised data augmentation for consistency training, and FreeMatch methods in urban area.