Cross-Modal Consistency-Guided Information Fusion for Hyperspectral and LiDAR Classification

The joint classification of hyperspectral image (HSI) and light detection and ranging (LiDAR) data plays an important role in remote sensing image interpretation and landcover mapping. However, the scarcity of labeled samples, coupled with the inherent heterogeneity of multisource data, poses significant challenges for achieving satisfactory classification performance. In this article, we propose a cross-modal consistency-guided information fusion framework that progressively exploits the inherent associations of same-class landcovers in terms of feature representation, local semantic context, and boundary structure, thereby guiding multisource information fusion for landcover classification under limited labeled samples. First, a multisource landcover feature construction scheme integrates spectral features, LiDAR-derived geometric features, and multiattribute spatial features from both modalities to establish complementary and diverse feature representations. Second, a prior-knowledge-guided pseudo-sample screening mechanism identifies high-reliability pseudo samples by evaluating joint spectral-elevation similarity within local windows centered on existing labeled samples for sample augmentation. Finally, a dual-gradient Markov random field (MRF), built upon Sobel-derived spectral and geometric gradient fields as complementary edge barriers, refines the initial classification map to enforce spatial coherence while preserving class boundaries. Experiments on three benchmark HSI–LiDAR datasets demonstrate the effectiveness of the proposed framework.