42 YEARBOOK 2025 of Roofing Contractors Scotland), the CyberBuild Lab have thus undertaken to explore how Industry 4.0 technologies could be developed and deployed to address the above limitations of current practice in visual roof condition monitoring. Specifically, the team has been developing a digital twin-based approach as illustrated in Figure 1. For each building roof, a twin virtual roof is continuously updated with information on the physical roof condition (for example, defects) automatically extracted from UAV imagery (which could be complemented with other reality capture data). This information is Figure 2: Stage A - Orthophoto generation process for one of the Duff House roof panels. recorded in a structured, spatially accurate manner which enables monitoring over time (that is, from imagery acquired at multiple successive epochs) and thereby data-driven maintenance and repair decision making. Figure 1 shows that our proposed method for roof condition visual evaluation has two stages: • Stage A: A high-quality orthophoto is generated for each roof panel from UAV-acquired imagery. • Stage B: Each orthophoto is processed in two steps: the orthophoto is first segmented into its main material components— slates, leadwork and other—using a specifically-developed AI model. Then, visible deteriorations and defects are detected within each area of the roof panel; we focus here on the slated area. Because the spatial relationship between each roof panel 3D geometry and its corresponding orthophoto is known, the detections are precisely mapped in space. This way, defect information is systematically documented, enabling condition monitoring over time (for example, evolution of a specific defect over time or patterns over time). Note that our approach assumes that an existing semantically rich 3D model (BIM model) of the building roof exists. This assumption can be relaxed, but we argue that it is also reasonable because: (1) our expectations for the 3D model are low with only the need to have a coarse 3D representation of the roof; and (2) it is very likely that in the medium to long term buildings will have a twin virtual 3D representation as part of their digital building passport. Figures 2 and 3 illustrate our results with an example building: Duff House, a historic Georgian mansion in Banff, Scotland, with complex roof geometry and architectural details. For the orthophoto generation (Figure 2), structure-from-motion photogrammetry is applied to UAVacquired pictures, which automatically outputs a 3D reconstruction of the roof in its current state. If the UAV data comes with precise geo-referencing, then the 3D reconstruction aligns with the existing 3D model representation of the roof in the digital twin. This alignment then enables the automated generation of orthophotos for each roof panel. This process employs various computational methods that aim to robustly merge the visual information coming from the many UAV pictures, reduce blur and other visual artefacts, and remove occluders (chimneys or other nearby roof panels for example). As Figure 2 illustrates, this process is quite robust. The reader will observe that the generated orthophotos are useful even to support current manual image annotation-based approaches to roof condition evaluation, easing visual evaluation, quantification, Figure 1: Three-stage DT-based roof condition visual evaluation model
RkJQdWJsaXNoZXIy MzI0Mzk=