Automated urban energy assessment: From thermal flyover to AI-driven retrofit prioritization for sustainable cities

Cities urgently need scalable methods to assess building energy performance for retrofit prioritization, yet current approaches are costly and slow for comprehensive urban energy planning. We present an automated pipeline that transforms thermal flyovers into actionable energy assessments using AI-driven analysis. Our approach segments mid-wave infrared orthomosaics into building patches using an unsupervised multi-modal pipeline (Thermal-SAM), extracts physically interpretable ring and context features, and benchmarks energy-use intensity (EUI) derived from aggregated EPC records. We exclude implausible EUI records above 1000 kWh/m2/yr from primary model-performance evaluation. On this screened dataset (n=1190), leakage-free non-thermal covariates achieve RMSE = 26.84 and MAE = 16.13 kWh/m2/yr, within the range reported by recent EPC-based prediction studies. An EPC-full upper-bound model that retains energy-consumption proxies achieves lower error (RMSE = 9.36, MAE = 1.31 kWh/m2/yr), but this reflects feature–target proximity rather than independent prediction. Thermal features are therefore not presented as routine direct EUI predictors. Instead, the excluded high-EUI tail is treated as a diagnostic anomaly set: in the unfiltered diagnostic analysis, thermal features help rank the largest thermal–EPC discrepancies (AUC = 0.779, 95% CI 0.774–0.784, 100 seeds), while this signal collapses after the implausible records are removed. These findings reposition thermal flyovers as a scalable tool for triaging buildings with unusually high or inconsistent EPC-derived energy use, while highlighting the limitations of single-pass thermal data for accurate EUI regression. Reproducible code (available upon request) supports rapid, citywide screening and targeted follow-up audits.