A deep-level feature fusion architecture via self-attention integrated U-Net combining ultra-hyperspectral image and LiDAR point cloud data towards detailed land cover classification

Land cover classification is essential for applications such as ecological monitoring, resource management and sustainable development. With the rapid development of deep learning, multi-modal data that integrate spectral and spatial information have become increasingly important for fine-grained classification. Hyperspectral imagery provides rich spectral information, while LiDAR offers precise three-dimensional structural data. However, the increasing spectral and spatial resolution also introduces challenges in effectively exploiting and fusing ultra-fine features. To address these challenges, we propose a deep learning-based spectral-spatial fusion framework for ultra-hyperspectral imagery and LiDAR data. A dual-branch U-Net architecture is adopted to extract modality-specific features, and a Channel-wise Competitive Attention Alignment (CCA) mechanism is introduced to perform adaptive cross-modal fusion under a sum-to-one constraint. This design enables effective interaction between hyperspectral and LiDAR features while reducing redundancy. Experiments are conducted using ultra-hyperspectral imagery acquired by the AISAIBIS sensor and LiDAR data collected by the RIEGL VQ-780II sensor over Danzhou city, Hainan Province, China. The proposed method achieves an overall accuracy of 87.87% in pixel-level land cover classification, outperforming single-modality approaches. The results demonstrate that the proposed fusion strategy effectively exploits the complementary characteristics of spectral and structural information, providing a robust solution for fine-grained land cover classification.