A Semi-Automatic Roof Type-Driven Workflow for LoD2 Building Reconstruction from Orthophotos and Photogrammetric Point Clouds
This study presents a semi-automatic, roof type-driven workflow that integrates digitized roof outlines from orthophotos with elevation data from photogrammetric point clouds to reconstruct accurate, semantically structured Level of Detail 2 (LoD2) building models in CityJSON and CityGML formats.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Cities are more than just collections of streets and buildings; they are complex, three-dimensional organisms that planners, engineers, and historians need to understand in full depth. To map this complexity, geographers and computer scientists have developed a standard way of describing buildings in digital space, known as the "Level of Detail 2" model. Unlike simple maps that only show the outline of a building on the ground, these detailed models capture the actual shape of the roof, the height of the walls, and the specific geometry that makes a house a house. This level of detail is crucial for tasks ranging from estimating how much sunlight a neighborhood receives to planning how a city might respond to extreme weather. However, creating these accurate 3D models from aerial photographs and laser scans has traditionally been a difficult, time-consuming process, often requiring expensive software or highly specialized skills that are not always available.
In a recent study, researchers from Universitas Gadjah Mada in Indonesia proposed a new way to build these digital models that balances human insight with computer automation. Instead of trying to force a computer to guess the shape of every roof from scratch, or relying entirely on a human to draw every single line, they developed a semi-automatic workflow that treats the type of roof as the central guide. The team focused on five common roof shapes found in the city of Singkawang: flat roofs, gable roofs (which look like a classic house with a triangular peak), hip roofs (where all sides slope down), shed roofs (a single slanted plane), and pyramid roofs. By first identifying which of these five types a building has, the researchers could apply a specific set of logical rules to reconstruct the building's 3D shape accurately.
The process began with a collection of high-resolution aerial photographs and a dense cloud of points generated by laser scanning, which acts like a digital skeleton of the city's surface. The researchers first used these data to create a digital map of the ground level, separate from the buildings and trees, and another map showing the very top of everything, including the highest points of the roofs. By comparing these two maps, they could determine exactly how tall each building was. Next, a human operator looked at the aerial photos and drew the outline of each roof, labeling it with its specific type. This step was crucial because it gave the computer a clear starting point and a clear instruction on what shape to build.
Once the outlines and labels were ready, a custom computer program took over. For a flat roof, the program simply lifted the outline to the correct height. For a gable roof, the program analyzed the slope and direction of the roof to find the exact line where the two sides meet at the peak, ensuring the angle matched the real-world data. For hip and pyramid roofs, the program calculated how the sides should slope inward to meet at a ridge or a single point. For shed roofs, it determined the direction of the slope and adjusted the height of each corner accordingly. The result was a complete, three-dimensional building model that included the ground, the walls, and the roof, all structured in a way that allows different computer systems to read and share the data easily.
The team tested this method on 371 buildings in the study area. The results were remarkably precise. When they compared the generated models against the original photographs, the buildings matched the real-world shapes almost perfectly, with the computer correctly identifying the roof planes in nearly every case. When they measured the dimensions of the models against the laser scan data, the average difference was less than one meter, a level of accuracy that meets the strict international standards for this type of digital mapping. One building in the test group had a minor error where an extra roof section was added, but the overall success rate remained extremely high.
This approach offers a practical middle ground between fully manual modeling, which is slow and expensive, and fully automatic systems, which can sometimes produce confusing or incorrect results when faced with complex cityscapes. By letting a human identify the roof type and then letting a computer apply the correct geometric rules, the researchers created a system that is both fast and reliable. The study demonstrates that with the right combination of human observation and rule-based automation, it is possible to generate high-quality, detailed 3D city models from standard aerial data without needing the most advanced or expensive technology. This method provides a clear, reproducible path for cities to build their own digital twins, helping planners and researchers visualize and understand the built environment with greater clarity than ever before.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.