DCCL: Dual-View Cross-Modal Consistency Learning for Hyperspectral and LiDAR Data Classification Under Noisy Labels
Abstract
Joint hyperspectral and light detection and ranging (LiDAR) data classification shows great promise for object recognition and land-cover mapping. However, most existing approaches assume clean labels—an assumption that may not hold in practical remote sensing applications—leaving noisy-label learning for real-world hyperspectral image (HSI)–LiDAR classification largely unexplored. To address this gap, we formulate the problem of HSI–LiDAR data classification under noisy labels and establish a baseline. We further develop a dual-view cross-modal consistency learning (DCCL) framework to enhance robustness against both symmetric and asymmetric label noise. In particular, we introduce a dual-view augmentation strategy that applies strong and weak perturbations separately to HSI and LiDAR data, providing label-independent supervisory signals and encouraging perturbation-invariant and modality-coherent representations. Furthermore, we propose a cross-modal consistency learning strategy that enforces prediction consistency across views and representation consistency across modalities and views, enabling robust dual-modal feature learning under noisy labels. Experiments on the Houston 2013, Trento, and MUUFL datasets under both symmetric and asymmetric noisy-label settings demonstrate that the proposed method consistently outperforms existing approaches. The results highlight the vulnerability of current HSI–LiDAR classification methods to label noise and confirm the effectiveness of the proposed DCCL framework. The source code will be made publicly available at https://github.com/halfcoder1/DCCL