Multilevel Differential Aggregation With Gated Discrimination Network for Hyperspectral-LiDAR Joint Land Cover Classification

• الحاوية / القاعدة: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
• المؤلفون الرئيسيون: Haizhu Pan, Bopeng Ren, Xuehu Li, Haimiao Ge, Cuiping Shi, Moqi Liu, Chunxue Xia, Jingshu Lv
• التنسيق: مقال
• اللغة: الإنجليزية
• منشور في: IEEE 2025-01-01
• الموضوعات: Differential aggregation; gated discrimination; morphological strip convolution; multisource remote sensing; weight coupling
• الوصول للمادة أونلاين: https://ieeexplore.ieee.org/document/11144477/
• تدمد: 1939-1404; 2151-1535

With the rapid development of Earth observation technologies, the integration of multisource remote sensing data provides great potential for land cover classification. Hyperspectral images (HSI) contain abundant spectralspatial information, while light detection and ranging (LiDAR) offers accurate 3-D structural data. Their effective fusion can significantly enhance classification accuracy. However, existing methods often retain redundant information during feature extraction and fusion, and such redundancy weakens the discriminative power of networks. Moreover, the irregular spatial distributions in multisource data make rulebased meshing ineffective for modeling complex geometric structures, limiting the capture of subtle cross-domain spatial features. To address these challenges, we propose a multilevel differential aggregation with gated discriminative network (MDAGNet) based on dynamic feature selection for joint HSILiDAR classification. First, a multilevel differential aggregation module separately extracts HSI spectral features and LiDAR elevation features to generate discriminative multilevel representations. Second, an adaptive discrimination module with a gating mechanism distills redundant information through multiscale discriminative operations, achieving deep fusion of critical features. Third, a morphological strip convolution module captures irregular geometric distributions in HSI, enhancing fine-grained feature perception. Finally, adaptive multimodal fusion is achieved through weight coupling and dynamic fusion strategies. Experiments on three public datasets demonstrate that the proposed method achieves superior accuracy and generalization compared with state-of-the-art approaches.