A novel multi-view intelligent fusion network for automatic 3D urban green spaces mapping

Three-dimensional (3D) maps of urban green spaces (UGS) provide a significant dataset for estimating carbon sequestration and understanding urban ecosystem functions. However, existing 3D-UGS mapping methods are often designed for light laser detection and ranging (LiDAR) data with high cost and small coverage, limiting their scalability for city-level applications. In response to the above limitations, we developed a multi-view intelligent fusion network by using multi-view images to generate a 3D-UGS map. The proposed algorithm consists of three parts: (1) Height information estimation network; (2) UGS extraction network; and (3) 3D-UGS map generation module. A height information estimation network was used to retrieve detailed UGS height information. UGS extraction network-based multimodal feature fusion idea was proposed to extract 2D-UGS. 3D-UGS map generation module was developed to produce a fine-grained 3D-UGS map. The proposed algorithm yields a high-quality 3D-UGS map with a root mean square error (RMSE) ranging from 0.85 m to 1.06 m for the urban semantic 3D (US3D) dataset. The results showed that the proposed algorithm achieves remarkable 3D-UGS map performance with an average RMSE of 2.045 m at Beijing. Our research provides new insight into how multi-view images and artificial intelligence (AI) can be integrated to generate fine 3D-UGS maps at the city scale.