Fanfeng Shi and Fang Zheng
Polarimetric SAR (PolSAR), image classification, semi-supervised learning, hybrid convolutional neural network (HCNN)
In computer vision applications, convolutional neural networks (CNNs) have demonstrated significant effectiveness and achieved remarkable success, largely because of the abundance of well-labelled datasets. However, acquiring high-quality labels for polarimetric synthetic aperture radar (PolSAR) images is both time-consuming and expensive. Over the past few years, semi-supervised learning has gained attention as an effective approach to reduce dependency on labelled samples by utilising a combination of both labelled and unlabelled data. This paper proposes a novel semi-supervised method that utilises a hybrid CNN architecture with cross-pseudo supervision (CPS). This method transforms the PolSAR coherency matrix into two forms and trains two distinct CNN models to handle these different inputs,which addresses two critical challenges in semi- supervised PolSAR classification: (1) preserving phase coherence in complex-valued data to resolve ambiguities in terrains with overlapping magnitude responses and (2) mitigating pseudo-label noise propagation in label-scarce scenarios. This dual innovation enables robust classification in label-scarce PolSAR applications. Experimental results obtained from two different datasets validate the effectiveness of the suggested approach. The overall accuracies of two datasets are 99.19% and 96.01% with using 0.8% and 0.08% training samples per class.
Important Links:
Go Back
