Building façades are among the most exposed components of structures, subject to degradation mechanisms that compromise safety, performance, and aesthetics. In recent years, Unmanned Aerial Systems (UAS) and Convolutional Neural Networks (CNNs) have increasingly been used to automate façade inspections. Although previous studies have demonstrated high accuracy in crack detection, most rely on close-range, high-contrast images captured under controlled conditions, limiting applicability in real-world operational settings. This study addresses this gap by developing and comparing four CNN-based models (MC, CC, RA, and RC) for crack detection in plastered and ceramic-tiled façades using RGB, thermal, and contrast-enhanced images acquired by UAS under operational conditions. Results show that the MC model achieved the highest recall (90.3%), mAP (73.2%), and F1-score (73.0%) for plastered façades, whereas the RC model achieved the highest precision (71.9%), mAP (81.1%), and F1-score (81.0%) for ceramic façades. Models based on AlexNet and ResNet achieved higher probability thresholds (MC: 98.5%; CC: 99.3%) than YOLOv11 models. The main contribution of this research lies in the comparative evaluation of multiple CNN architectures across different façade typologies using realistic UAS-acquired imagery. By demonstrating the feasibility of automated crack detection under non-ideal conditions, the study advances technical support for façade inspections and maintenance planning, bridging the gap between academic performance benchmarks and practical implementation in building pathology and rehabilitation. These findings highlight the potential of combining UASs and CNNs to improve efficiency, accuracy, and safety in large-scale building inspections.
Santana et al. (Fri,) studied this question.