Hierarchical feature interactive fusion network for hyperspectral image and LiDAR data classification

With the advancement of remote sensing (RS) technology, the integrated application of multimodal images in surface monitoring and analysis has become a significant research direction. Hyperspectral images (HSI) can accurately capture spatial-spectral features, while detailed elevation data can be provided by light detection and ranging (LiDAR). However, most existing studies employ single-level feature fusion, failing to effectively capture cross-level feature interactions, which limits deep information fusion between multimodal data. Therefore, a hierarchical feature interactive fusion network (HFIFN) is proposed, aiming to realize full interactive fusion among multi-level features. First, a dual-branch convolutional neural network architecture is employed to perform initial feature extraction. Then, an interactive fusion transformer (IFT) is leveraged to learn the semantic relationship between modality features, mining the spatial correlations of modalities to achieve feature interaction. By exploring the inter-modal spatial correlation through the query mechanism, and combining the attention guidance block (AGB) to constrain the correlation, IFT optimizes the fusion effect while preserving the feature dissimilarity. In addition, the hierarchical interactive fusion module (HIFM) is employed to capture the correlation between multi-level features and achieve comprehensive integration of features across different hierarchies, thus enhancing the fusion effect. The final experimental results obtained on three common datasets validate the superiority of the investigated method.