高光谱成像
遥感
激光雷达
计算机科学
人工智能
融合
传感器融合
特征提取
计算机视觉
人工神经网络
图像融合
模式识别(心理学)
雷达跟踪器
图像分辨率
图像分割
图像处理
作者
Wuli Wang,Jianchun Gao,Jing Sun,Jing Jia,Xujie Wang,J Zhang,Jianbu Wang,Peng Ren
标识
DOI:10.1109/lgrs.2026.3691834
摘要
Hyperspectral and LiDAR data fusion improves classification by combining spectral and structural information. However, accurate classification in real-world scenarios remains challenging. Key issues include large-scale variations within object classes, cross-modal noise interference, and limited labeled samples. To address these, this paper proposes a Multi-Granularity and High-Order Fusion Network (MG-HOF Net). The framework consists of three novel modules. First, the Multi-granularity Detail-preserving Large Kernel Attention (MD-LKA) module captures both long-range context and local details to handle scale variations. Second, the Gated Bi-directional Cross-Modal (GBCM) fusion module employs symmetric interaction and gated residuals, enabling selective information fusion while suppressing noise. Third, the Second-Order Rectified Distribution (SORD) head replaces first-order pooling with bilinear interaction and power normalization, which enhances feature discriminability and robustness in few-shot settings. Experiments on three datasets demonstrate that our MG-HOF Net achieves state-of-the-art performance under few-shot conditions.
科研通智能强力驱动
Strongly Powered by AbleSci AI