Calibration-Enhanced Multi-Awareness Network for Joint Classification of Hyperspectral and LiDAR Data

Hyperspectral image (HSI) and light detection and ranging (LiDAR) data joint classification has been applied in the field of ground category recognition. However, existing methods still perform poorly in extracting high-dimensional features and elevation information, resulting in insufficient data classification accuracy. To address this challenge, we propose a novel and efficient Calibration-Enhanced Multi-Awareness Network (CEMA-Net), which exploits the joint spectral–spatial–elevation features in depth to realize the accurate identification of land cover categories. Specifically, we propose a novel multi-way feature retention (MFR) module that explores deep spectral–spatial–elevation semantic information in the data through multiple paths. In addition, we propose spectral–spatial-aware enhancement (SAE) and elevation-aware enhancement (EAE) modules, which effectively enhance the awareness of ground objects that are sensitive to spectral and elevation information. Furthermore, to address the significant representation disparities and spatial misalignments between multi-source features, we propose a spectral–spatial–elevation feature calibration fusion (SFCF) module to efficiently integrate complementary characteristics from heterogeneous features. It incorporates two key advantages: (1) efficient learning of discriminative features from multi-source data, and (2) adaptive calibration of spatial differences. Comparative experimental results on the MUUFL, Trento, and Augsburg datasets demonstrate that CEMA-Net outperforms existing state-of-the-art methods, achieving superior classification accuracy with better feature map precision and minimal noise.