Urban areas, now hosting over half of the global population and consuming 60–80% of total energy, make urban energy management a critical global priority. Historic cities, with their centuries-old and listed buildings, pose unique challenges but also offer significant opportunities for improving energy efficiency. Effective energy management requires comprehensive knowledge, supported by accurate diagnostics of building quality. Non-destructive testing (NDT) techniques, valued for their precision and non-invasive nature, offer effective solutions for complex urban environments. Among these, infrared thermography (IRT) plays a central role by enabling rapid and non-invasive acquisition of surface temperature distributions to support both qualitative and quantitative analyses. Drone-assisted IRT (UAV-IRT) further enhances these capabilities by facilitating energy audits, diagnosing heat losses, assessing HVAC efficiency, and inspecting renewable energy systems at the urban scale. Unlike ground-based imaging, UAV-IRT can access hard-to-reach areas and high-altitude structures (e.g., rooftops), without accessibility constraints, thereby improving the scope, accuracy, and efficiency of diagnostics.

Interpreting thermograms acquired via UAV-IRT for urban roof diagnostics remains challenging due to heterogeneous construction materials, complex thermal patterns, and the large volume of data generated. Manual inspection is often inefficient, subjective, and error-prone, particularly when distinguishing among insulation defects, moisture ingress, thermal bridges, and potential structural damage. This study presents a dataset of thermograms collected over two years from multiple historic buildings in Italy and applies machine learning for automated interpretation. The AI-based approach allows for the automatic detection and classification of thermal anomalies by reducing subjectivity and improving accuracy and consistency. It also enables surface segmentation by distinguishing between different materials and functional zones (e.g., roof coverings, mechanical systems, structural elements), revealing correlations between anomalies and underlying structural or energy-related conditions. This method demonstrates the feasibility of faster, more reliable, and predictive roof diagnostics, thereby supporting evidence-based strategies for energy management and conservation in historic cities.