Predicting building age from urban form at large scale.

• Assessing generalizability of age prediction across cities and countries based on urban form. • Evaluating amount of local data needed to fill data gaps. • Detailed analysis of predictive performance across regions, construction periods, settlement and buildings types. • Assessing the usability of predictions to improve prioritization of large-scale retrofits. • Highlighting climate relevance of scalable, spatially explicit building attribute prediction. • Filling data gaps within countries is possible. • 10% local data allow inference of remaining unknown building ages. • Massive training data improves generalization across regions. • Generalizing across countries is not (yet) possible. • Age predictions can inform retrofit policies and may significantly increase energy savings. To stay within 1.5 °C of global warming, reducing energy-related emissions in the building sector is essential. Rather than generic climate recommendations, this requires tailored, low-carbon urban planning solutions and spatially explicit methods that can inform policy measures at urban, street and building scale. Here, we propose a scalable method that is able to predict building age information in different European countries using only open urban morphology data. We find that spatially cross-validated regression models are sufficiently robust to generalize and predict building age in unseen cities with a mean absolute error (MAE) between 15.3 years (Netherlands) and 19.9 years (Spain). Our experiments show that large-scale models improve generalization for predicting across cities, but are not needed to infer missing data within known cities. Filling data gaps within known cities is possible with a MAE between 9.6 years (Netherlands) and 16.7 years (Spain). Overall, our results demonstrate the feasibility of generating missing age data in different contexts across Europe and informing climate mitigation policies such as large-scale energy retrofits. For the French residential building stock, we find that using age predictions to target retrofit efforts can increase energy savings by more than 50% compared to missing age data. Finally, we highlight challenges posed by data inconsistencies and urban form differences between countries that need to be addressed for an actual roll-out of such methods. [ABSTRACT FROM AUTHOR]