In recent years, the widespread use of mobile devices equipped with LiDAR sensors has made it possible to easily capture 3D point clouds. However, these devices suffer from cumulative self-localization errors, limiting accurate 3D measurements to relatively small areas. Existing approaches to wide-area georeferencing of mobile LiDAR point clouds typically rely on GNSS receivers, direct 3D point-cloud matching, or alignment to pre-existing high-precision urban 3D models. These approaches can be costly, computationally demanding, or impractical where such reference data are unavailable.

Therefore, this study proposes a low-cost method for constructing a high-precision wide-area 3D map using only LiDAR-equipped mobile devices and open data. It is intended to help municipalities and citizen communities without expensive equipment or professional surveying skills develop local digital twin infrastructure. The resulting 3D maps can support 3D visualization of potential disaster impacts and community disaster preparedness.

The proposed processing pipeline accepts 3D point clouds from mobile devices and Japanese governmental open geospatial data as input and outputs open point-cloud formats such as LAS/LAZ for seamless use in FOSS4G software such as QGIS. The workflow consists of two stages: registration of adjacent point clouds and georeferencing of the integrated point cloud within a global geodetic coordinate system. Rather than directly estimating rigid transformations in full 3D space, the proposed method separates spatial information into horizontal and vertical components and optimizes them sequentially, improving registration stability for point clouds acquired by mobile devices. Geographic coordinates are assigned through image matching against open geospatial data instead of GNSS-based positioning. All processing, except point cloud acquisition, is implemented using open-source Python libraries widely used in the FOSS4G ecosystem. To support reproducibility, we will release the source code, processing workflow, example configuration files, and links to the open input datasets used in the experiment.

First, rough overlaps are extracted from the approximate positional information of each point cloud. CSF is then applied to separate each point cloud into ground and non-ground points. Height-based slices of the non-ground points are projected into bird’s-eye-view images representing wall surfaces. ORB feature-based image matching in OpenCV provides initial alignment, followed by horizontal refinement on the XY plane using the Point-to-Point ICP algorithm in Open3D. Finally, vertical and tilt errors are corrected using the ground points.

Next, absolute coordinates are assigned to the registered point clouds using Fundamental Geospatial Data provided by the Geospatial Information Authority of Japan. Although validation is performed with Japanese open data, the method is applicable in other regions where road edge vector data and digital elevation models (DEMs) are openly available. For horizontal georeferencing, road edge data is used. Using geopandas and shapely, a rough region around the point cloud is extracted and converted into a road edge image. The integrated point cloud is processed in the same manner to generate a wall surface image. OpenCV’s normalized cross-correlation (NCC) is then used to estimate the optimal translation and rotation parameters. For vertical georeferencing, a 5 m mesh DEM is referenced, and elevations are extracted with rasterio. A correction surface is generated by smoothing the elevation differences between the ground points and the DEM, and the resulting correction values are applied to the entire point cloud. This aligns absolute elevation with the DEM while preserving local terrain variations.

To verify the effectiveness of the proposed method, an experiment was conducted in the Jinaimachi district of Tondabayashi City, Osaka Prefecture. Measurements were conducted using an iPad Pro (11-inch, 4th generation), and the Scaniverse application. Five scan datasets were obtained from these measurements. In addition, ground truth data were prepared using measurements from a high-performance 3D scanner (Matterport Pro3) and a GNSS receiver (Drogger RZX.D). As evaluation metrics, both the RMSE between feature points of adjacent point clouds after registration (relative accuracy) and the RMSE between the final constructed 3D map and the ground truth data (absolute accuracy) were evaluated. The target accuracy was defined as an RMSE of within 0.1 m for relative accuracy. For absolute accuracy, the horizontal accuracy was required to be within 1.75 m, corresponding to “Map Information Level 2500” and suitable as a base map for hazard mapping, while the vertical accuracy was required to be within 0.30 m, suitable for flood simulation.

As a result of the evaluation, the average relative accuracy achieved an RMSE of 0.048 m, sufficiently satisfying the target value. Regarding absolute accuracy, the horizontal RMSE was 0.63 m and the vertical RMSE was 0.09 m, demonstrating favorable results. The results satisfy the requirements for Map Information Level 2500 horizontally and Level 500 vertically.

For the FOSS4G community, the proposed pipeline demonstrates how mobile LiDAR, governmental open data, and open-source geospatial libraries can be combined into a reproducible workflow for practical 3D hazard mapping. It also aligns with the conference’s emphasis on Asian geospatial initiatives by demonstrating a reproducible workflow based on Japanese governmental open data. Experimental results confirmed that the proposed method provides sufficient accuracy for 3D hazard maps that enable intuitive visualization of flood depths and landslide-affected areas. This method facilitates 3D data development in municipalities and citizen communities with limited budgets and can contribute to regional digital transformation. Future work includes improving robustness in wider and more diverse environments and developing a web system that integrates data from multiple devices to generate 3D maps automatically. Unlike workflows that depend on GNSS receivers or pre-existing high-precision 3D city models, the proposed method enables georeferenced wide-area 3D mapping using only mobile LiDAR, open geospatial data, and an open-source processing stack.