Climate controls on runoff generation in an experimental blue–green roof revealed by high-resolution monitoring and machine learning

Blue–green roofs (BGRs) are increasingly adopted for urban stormwater management, yet their hydrological response remains difficult to predict due to climatic and antecedent conditions. This study investigates runoff generation in an experimental Mediterranean BGR using four years (2021–2025) of high-resolution (10-min) monitoring data and machine learning (ML). Results reveal a nonlinear, state-dependent rainfall–runoff relationship. An empirical rainfall breakpoint of approximately 17 mm was identified, above which runoff occurrence and variability generally increased; however, runoff occurred below this value and remained limited for some events above it, confirming that rainfall depth alone is insufficient to explain system behaviour. Antecedent dry period and air temperature emerged as important predictors, suggesting associations with pre-event storage conditions. ML models achieved moderate predictive performance (AUC ≈ 0.68), indicating that atmospheric variables provide useful information beyond rainfall magnitude. These findings support climate-adaptive, performance-based approaches for modelling and designing BGR systems.