Automatic DTM and Building Footprint Extraction from Imageries and Point Clouds in Indonesia’s Land Registration Drone Survey: A Roadmap Towards Reconstruction of LOD1 3D building model
Ruli Andaru, Trias Aditya, Bambang Kun Cahyono, Purnama Budi Santosa, Yulaikhah, Septein Paramia Swantika
Accurate and automatic Digital Terrain Model (DTM) and building footprint extraction from
drone survey has become essential and challenging work for cadastre verification,
modernization and updating. In the context of multipurpose cadastre, the integration of land
parcels with other spatial information such as building footprint, terrain elevation, and 3D
model, allows for detailed representation of land information. This facilitates spatial analysis
and adjacency information within the real-world objects above the ground. In this paper, we
introduce an approach for automatic DTM and building footprint extraction by implementing
deep-learning methods (i.e., YOLO v8 and CNN) using true-orthoimage UAV and point
clouds. We first apply photogrammetric processing through SfM pipelines to produce 3D
point clouds and true-orthophoto. To extract DTM, CNN deep learning is implemented to
classify point clouds into ground and non-ground objects. The detection of building footprint,
as an important spatial information in the cadastral intelligence, is performed by
implementing YOLO v8 deep-learning using custom trained data. To ensure that users,
irrespective of their technical skill levels, can easily navigate and utilize those two algorithms,
we build a GUI for a desktop application using Python, namely Geo-Carta (Geospatial-Cadastre
with Artificial Intelligence for Generating LOD 3D City Model). It consists of four
features dealing with the detection of building footprint from orthophotos, ground extraction
from point clouds, land parcel editing feature, and generation of LOD-1 3D building models.
We tested the Geo-CARTA app for detecting building footprints across various building types
(with different shapes and patterns) in several provinces in Indonesia, i.e., Papua, West
Sulawesi, East Borneo, Riau, West Java, and Yogyakarta. The results show that the detection
of building footprint reached the accuracies of 88.47%. For the accuracy assessment of
ground extraction, we tested with the UAV dense clouds in West Java and Yogyakarta,
achieved an accuracy of 0.969. The resulting building footprints, DTM, and DSM were then
used for reconstructing 3D building models in LOD1 which were implemented automatically
using Geo-CARTA app and exported the 3D model into cityjson format.
Event: 12th International FIG Workshop on LADM & 3D LA
Only personal, non-commercial use of this document is allowed.