PDCNet: A Polarimetric Data-Enhanced Contrastive Learning Network for PolSAR Land Cover Classification
Polarimetric synthetic aperture radar (PolSAR) has rich polarization information, offering an efficient and reliable means of collecting information. However, how to effectively leverage these complex data to extract polarization features remains a key challenge. Recently, contrastive learning has b...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10948157/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Polarimetric synthetic aperture radar (PolSAR) has rich polarization information, offering an efficient and reliable means of collecting information. However, how to effectively leverage these complex data to extract polarization features remains a key challenge. Recently, contrastive learning has been successful in computer vision, with fewer labeled and a large amount of unlabeled data. Inspired by it, a polarimetric data-enhanced contrastive learning network (PDCNet) for PolSAR land cover classification is proposed. The design process for polarimetric contrastive learning involves the construction of positive samples, the establishment of a PolSAR-based network architecture for contrastive learning, and the formulation of the loss function. The proposed method first constructs positive sample pairs for contrastive learning by leveraging the linear relationship between the coherency matrix (T-matrix) and the covariance matrix (C-matrix). In addition, since both of them are <inline-formula><tex-math notation="LaTeX">$\text{3}\times \text{3}$</tex-math></inline-formula> complex conjugate symmetric matrices, we design an adaptive contrastive learning network to extract real and complex information. Specifically, the encoder of PDCNet is designed as an extraction module for a real-convolutional composite complex convolutional network. Thirdly, a loss function is designed to effectively compute the real and complex high-level semantics of the T- and C-matrix's features separately. We conducted experiments with labels of different quantities in four data sets. The results indicate that our proposed model outperforms state-of-the-art contrastive learning models. |
|---|---|
| ISSN: | 1939-1404 2151-1535 |