2D3D-Netherlands dataset and 2D3D-France dataset

We have compiled two distinct datasets comprising both HSR images and LiDAR point clouds. One is the 2D3DScene-Netherlands, which was collected in Amsterdam and Hague, Netherlands. The other is the 2D3DScene-France, which was collected in Montpellier, Lyon and Marseille, France. The HSR images utilized in both datasets were sourced from Google Earth, and captured at a resolution of 0.5 meters. All the images were cropped and resampled to 256×256 pixels. The LiDAR point cloud data in the 2D3DScene-Netherlands dataset originated from the 4-th version of Actueel Hoogtebestand Nederland (AHN-4) project (https://www.ahn.nl/) conducted by the Dutch Water Commission. The AHN-4 project utilized the most advanced airborne LiDAR sensors, such as RIEGL LMS-Q1560, for point cloud data collection in 2020, resulting in an approximate point density ranging between 6 to 10 points per square meter. Meanwhile, the LiDAR point cloud data in the 2D3DScene-France dataset were derived from the LIDAR HD project (https://geoservices.ign.fr/lidarhd), initiated and led by the French National Geographic Forestry Information Institute. The LIDAR HD project utilized the most advanced sensors, such as RIEGL VQ-1560II and Leica ALS80, for point cloud collection in 2021, resulting in an average point density ranging from 10 to 20 points per square meter. It is important to note that this study exclusively employes the x-y-z coordinates of the point cloud data, while omitting other attributes such as intensity. The 2D3DScene-Netherlands and the 2D3DScene-France datasets were collected over an area of 318.4 km² and 344.75 km², respectively. Moreover, co-registration between LiDAR data and HSR images was conducted by firstly automatically detecting correspondence points from the them, and secondly calculating 3D transformation parameters. Based on the computed 3D transformation parameters, the point clouds were reprojected and a registration accuracy of less than 1 cm was achieved.Each of the two datasets encompasses nine distinctive categories that exemplify variances in both 2D and 3D representations, including transport (TR), industrial area (IA), office building (OB), sport field (SF), farmland (FA), forest (FO), residential area (RA), church (CH), and meadow (ME). Categories are manually labeled through expert visual interpretation, utilizing both image and point cloud data. In addition, the POI and area of interest (AOI) information of Google Map are referenced to verify labeling accuracy. In both datasets, we have labeled 250 samples per category, and these samples are distributed over the whole study area. Such a design fosters equilibrium among categories and ensures uniformity in sample sizes for comprehensive analysis and evaluations. All these samples will be uploaded after this paper is accepted.