RECLNet: Riemannian Manifold Enhanced Contrastive Learning Framework for PolSAR Image Few-Shot Classification

In recent years, contrastive learning (CL) methods have achieved remarkable success in few-shot classification of polarimetric synthetic aperture radar (PolSAR) images. However, existing CL models based on Euclidean metric typically vectorize PolSAR data into real-valued or complex-valued vectors with independent channels, disrupting the inherent correlations between polarimetric channels. To address this issue, we propose a novel CL framework, the Riemannian-Euclidean CL network (RECLNet). The proposed RECLNet allows direct input of polarimetric covariance matrices, overcoming the limitations of conventional Euclidean-based CL models that require vectorizing PolSAR data. First, RECLNet constructs hard positive samples by leveraging polarimetric information within the Riemannian space. Second, a series of carefully designed Riemannian manifold operation (RMO) layers are used to extract the Riemannian geometric features of PolSAR data while preserving its matrix structure. Finally, the vision transformer (ViT) is adopted as the backbone of the Euclidean metric encoding to capture contextual information. Experiments conducted on two widely used PolSAR datasets demonstrate that the proposed approach achieves superior performance compared with existing state-of-the-art methods.