REVIEW AND ANALYSIS 43 buyers, should also have access to a complete digital record of the building history, containing all defects and repairs over the building’s life. The overall approach we presented in this paper is inscribed within that vision because it enables any detected defect (whether identified through manual annotation or the orthophotos, or using advanced AI technologies) to be recorded in a structured and spatially accurate manner. Regarding AI–driven defect detection our results are, so far, much more mixed, highlighting the challenges faced. In particular the difficulty in collecting extensive and varied data for training AI models that are robust enough to successfully process orthophotos for any other roof. The lack of suitable training dataset is a common challenge with the application of AI throughout the construction sector. This challenge can be addressed in different ways. Firstly, the sector should work to create such datasets and open them to researchers. Given what is observed in other sectors, this would rapidly lead to the creation of new, powerful automated defect detection models that would enable safe and robust visual roof monitoring at scale. Meanwhile, alternative methods exist to mitigate the lack of extensive datasets, including developing human-in-theloop solutions, where the AI system continuously adjusts its learning from feedback created by users. In conclusion, while much work is needed to achieve more automated roof visual monitoring and repair decision making, works such as this show that progress is rapid. Practitioners should remain curious about those developments as solutions usable in practice are likely to emerge in the near future. Frédéric Bosché PhD is Reader in Construction Informatics at the University of Edinburgh, and Past President of the International Association for Automation and Robotics in Construction. Jiajin Li is a PhD student at the University of Edinburgh, working with Historic Environment Scotland on the monitoring of traditional building fabric using UAV and computer vision. Figure 3: Stage B– Orthophoto processing for one of the Duff House roof panels. (a) Orthophoto of a roof panel obtained after Stage A. (b) Identification of the slated area of the roof panel orthophoto. Legend: grey-background, red-slates, blue-leadwork, brown-other. (c) Defective slate detection: organic growth (top) and structural defects (bottom). documentation and communication, just like facade orthophotos are known to be for walls. Subsequently, for the roof panel orthophoto analysis (Figure 3), each orthophoto is automatically divided into its main roof materials—slates, leadwork, and other components (Figures 3a and 3b). This process, known as ‘semantic segmentation’, classifies each pixel in the orthophoto into one of those three categories. For this, we employ a state-of-theart AI model (deep learning) that we have trained with manually-labelled example orthophotos. Our experimental evaluations indicate that this method is also quite robust, yielding accuracies of 89% for slated area detection. Once the slated area is isolated, slate defects are automatically detected (Figure 3c). This step can be approached in different ways: one way is to directly detect defects in the orthophoto; another way is to first detect slates and then classify them as defective or in good condition. We have explored both approaches and the results in Figure 3 show results obtained using the second method. The results visually appear encouraging, but the reality is that neither method has yet yielded good enough performance: while the accuracy in the detection of organic growth is better (60%), the accuracy in the detection of slates with structural defects (broken, slipped, missing) is much lower (45%). TAKE AWAY FOR PRACTITIONERS So, what can roof contractors and building owners take away from this? Answering this question requires us to consider two distinct aspects of our work: information management, and AI–driven defect detection and repair decision making. From an information management perspective, there are strong arguments for digital building passports. Just like car repair shops can now access the entire history of a car through digital records (and car owners should in fact be able to do so), a building owner, the companies they contract for monitoring and maintenance, and even potential
RkJQdWJsaXNoZXIy MzI0Mzk=