HSI and LiDAR Joint Classification Network With Interactive Extraction and Heterogeneous Balance

To improve the representation of spectral features and enhance the complementarity of heterogeneous spatial features in the joint classification of hyperspectral image (HSI) and light detection and ranging image (LiDAR), this paper proposes a joint classification network with interactive extraction and heterogeneous balance (JCNIH). On the one hand, this paper designs an interactive method for information exchange between large-span spectral features and small-span spectral features, which improves the adaptability and representation of spectral features. On the other hand, this paper designs a balance loss to constrain the extraction of heterogenous spatial features, which can keep the feature extraction balance and enhance the spatial complementarity. Then, the extracted spectral features and heterogeneous spatial features are aggregated for classification. Finally, comparative and ablation experiments demonstrate the effectiveness of JCNIH, and the experimental results show that the JCNIH outperforms the state-of-the-art methods in overall accuracy by 10.33% on three public datasets.