Urban Land-use Features Mapping from LiDAR and Remote Sensing Images using Visual Transformer Network Model

Abstract. With the rapid development of science and technology in the acceleration of urbanization, it is important to achieve efficient and accurate monitoring and mapping of urban features. Traditional urban feature mapping methods often rely on a single data source, such as optical remote sensing images or LiDAR, which often encounter many challenges in complex urban environments, such as shading, occlusion, and land cover changes. LiDAR has relatively accurate three-dimensional spatial information, while remote sensing image has rich spectral information. Thus, the fusion of spatial-spectral features can improve the accuracy and robustness of automatic classification efficiency for urban feature mapping. Recently deep learning technology has achieved a profound impact on remote sensing data processing. However, some existing deep models have not effectively fused spatial-spectral information In addition, the lack of semantic information optimization could confuse classification, especially for some high spectral heterogeneity areas. Hence, this study proposed a visual Transformer model to achieve automatic mapping combined with LiDAR and remote sensing images. In addition, this study improved the global attention mechanism for adaptive enhancing spectral-spatial fusion. Finally, it is found that the proposed algorithm is generally better than other representative methods, and the classification accuracy using remote sensing data and LiDAR is improved. The proposed modules can improve the Kappa coefficient by 5%.