Class-Anchor-Guided Shared Discriminative Learning for Open-Set Hyperspectral-LiDAR Classification
Abstract
Open-set joint classification of hyperspectral image (HSI) and light detection and ranging (LiDAR) data requires accurate known-class discrimination and reliable rejection of unseen classes. A key challenge is that the fused open-set decision space may be affected by inconsistent class structures learned from modality-specific HSI and LiDAR representations. To address this issue, this article proposes a class-anchor-guided shared discriminative learning framework for open-set HSI-LiDAR classification. The proposed method extracts modality-specific features using a hybrid HSI encoder and a structural LiDAR encoder, integrates them through low-rank cross-modal enhancement, and organizes HSI, LiDAR, and fused embeddings with learnable class anchors in a shared metric space. During inference, unknown samples are rejected only from the fused representation using an anchor-derived margin–energy score. Experiments on the Houston2013, Trento, and MUUFL datasets show that the proposed method improves known-class classification and unknown-class rejection over representative closed-set and open-set baselines.