Cross-modal three-dimensional intrinsic image decomposition for hyperspectral and LiDAR image joint classification

Existing remote sensing technologies capture discriminant information of on-ground objects and materials from a distance for accurate land-cover identification. Specifically, hyperspectral and light detection and ranging (LiDAR) images record spectral and spatial information and elevation information, respectively. Aligning and fusing both RS modalities comprehensively is beneficial to improving classification performance. However, this alignment and fusion are always challenged by the enormous RS modality gap. In this paper, a cross-modal three-dimensional intrinsic image decomposition (CM3DIID) is presented for hyperspectral and LiDAR image joint classification. Intrinsic image decomposition contributes to enhancing semantic information and making RS modalities clear. Our motivation is to construct a unique cross-modal intrinsic image decomposition for three-dimensional data cubes specifically, facilitating robust discriminant information extraction. Specifically, the hyperspectral image contributes to restricting the decomposed reflectance component, making the samples with similar spectral curves very close. The LiDAR image is considered to search samples with similar elevations, which should be closed in the shading component. The whole optimization process is clearly analyzed and achieved by constructing a deep learning network. The proposed CM3DIID constructs a connection bridge between hyperspectral and LiDAR images for information switching and modality fusing. Multiple experiments are conducted on two widely used hyperspectral and LiDAR datasets to verify the effectiveness and feasibility of the proposed CM3DIID. The final results clearly prove that the proposed CM3DIID is capable of generating better classification performance than other state-of-the-art comparison techniques.