- Moisture stains indicate ongoing degradation processes and may reveal areas of the roof slab where water infiltration occurs, compromising the performance and durability of the building system. During inspections of roofing systems, an inspector's field of vision differs from that of drones during overflights. As a result, traditional inspections might not always detect the presence and severity of stains, making maintenance on flat roofs a complex task. In this context, this experimental study aims to analyze deep learning-based semantic segmentation with images obtained from drones to map and monitor damp patches during automated building inspections of flat roof systems. The research tested two convolutional neural networks for semantic segmentation: the Fully Convolutional Network (FCN) with a ResNet50 backbone and DeepLabV3 with a ResNet101 backbone, as well as a transformer-based deep artificial neural network called SegFormer with a MiT-B1 backbone. We evaluated three optimizersMoisture stains indicate ongoing degradation processes and may reveal areas of the roof slab where water infiltration occurs, compromising the performance and durability of the building system. During inspections of roofing systems, an inspector's field of vision differs from that of drones during overflights. As a result, traditional inspections might not always detect the presence and severity of stains, making maintenance on flat roofs a complex task. In this context, this experimental study aims to analyze deep learning-based semantic segmentation with images obtained from drones to map and monitor damp patches during automated building inspections of flat roof systems. The research tested two convolutional neural networks for semantic segmentation: the Fully Convolutional Network (FCN) with a ResNet50 backbone and DeepLabV3 with a ResNet101 backbone, as well as a transformer-based deep artificial neural network called SegFormer with a MiT-B1 backbone. We evaluated three optimizers for each model—Adam, Adagrad, and SGD—along with learning rates of 1e-2, 1e-3, and 1e-4. The models were compared using four performance metrics. The FCN, optimized with Adagrad at a learning rate of 1e-2, showed the best results. The average metrics obtained in this case were as follows: precision: 79.69 %, recall: 67.81 %, F-score: 73.09 %, and Intersection over Union (IoU): 57.70 %.…
