Multichannel semi-supervised active learning for PolSAR image classification

Abstract

Deep neural networks have recently been extensively utilized for Polarimetric synthetic aperture radar (PolSAR) image classification. However, this heavily relies on extensive labeled data which is both costly and labor-intensive. To lower the collection of labeling data and enhance the classification performance, a novel multichannel semi-supervised active learning (MSSAL) method is proposed for PolSAR image classification. First, a multichannel strategy-based committee model with cooperative representation classification is presented to explore more effective information in the limited training data. Second, a loss prediction (LP) module is designed to identify the most informative pixels, and an ensemble learning (EL) strategy is designed to select the pixels with the highest confidence. Then, the deep neural network is fine-tuned with the obtaining target pixels through LP and EL in each iteration. Finally, the trained deep model predicts labels for all unlabeled data, outputting the final classification results. The proposed method is evaluated on three real-world PolSAR datasets, demonstrating superior performance to other PolSAR image classification methods with limited labeled samples.