Dynamic Common and Unique Feature Fusion Network for Hyperspectral and LiDAR Data Classification

With the advancement of multimodal technology, the combination of hyperspectral image (HSI) and light detection and ranging (LiDAR) data has become a prominent research topic in land use and land cover (LULC) classification. There are complex correlations between HSI and LiDAR. However, many studies focus on the spectral, spatial, and other attribute features of multimodal data while ignoring these important relationships, thereby limiting the classification performance. To address these, we propose a dynamic common and unique feature fusion network (DCU-Net) that establishes the dependence relationship between HSI and LiDAR and mines their shared and complementary information. A multiscale attribute feature extraction block is employed to capture spectral–spatial information of HSI and spatial–elevation information of LiDAR data, which effectively reduces the effect of scale differences among objects. In addition, we introduce a novel common-unique transformer (CUT) block with cross dynamic-agent-attention to extract common features of HSI and LiDAR data and depthwise convolution modules focusing on their unique features. By associating common and unique features of HSI and LiDAR with their dependencies, the robustness and classification accuracy of the model are significantly improved. In the fusion stage, common and unique features are adaptively reconstructed into discriminative features containing high-level semantic information. Extensive experiments on four popular HSI and LiDAR datasets illustrate the superiority and effectiveness of our proposed model, which showcases the great potential for multimodal remote sensing (RS) data analysis. The source code of the proposed method is available publicly at https://github.com/wanglei1588/Comon_Unique