A Reference-based Urban Evaporation Partitioning framework for urban interception estimation using multi-source observations

Evaporation from rainfall intercepted by the urban canopy layer (E_i) is a key but highly uncertain component of urban water and energy balances, with important implications for runoff generation and stormwater management. Quantifying E_i remains challenging due to the strong spatial heterogeneity of urban vegetation and impervious surfaces, as well as the difficulty in separating interception evaporation from other evaporation components during wet periods. In this study, we develop and validate a Reference-based Urban Evaporation Partitioning (RUEP) framework to quantify urban canopy interception evaporation by integrating multi-source observations with data-driven modeling. The proposed framework combines a hybrid machine learning model (HM_Ets) and a deep learning model (ML_Ei). HM_Ets is trained using dry-period observations to estimate total non-interception evaporation, including transpiration and soil evaporation (E_ts), and is subsequently applied to wet periods. Interception evaporation (E_i) is derived as the residual between observed wet-period evaporation and modeled E_ts, and is further simulated using ML_Ei. The framework is evaluated using multi-source datasets from an urban flux site in Vancouver, Canada, including eddy covariance flux measurements, meteorological observations, remote sensing products, and GIS-derived urban morphology data. Results demonstrate that the RUEP framework effectively reproduces both E_ts and E_i dynamics, with R² values of 0.80 and 0.90 and Nash–Sutcliffe efficiencies of 0.55 and 0.81 for dry and wet periods, respectively. Event-based interception ratios (E_i/P) exhibit pronounced seasonal variability, peaking in autumn (0.47) and reaching minimum values in spring (0.09), while E_i/E ratios peak in spring (0.14) and are lowest in autumn (0.08). At the street-block scale, E_i shows strong spatial heterogeneity and a non-monotonic “high–low–high” pattern along the combined normalized difference vegetation index (NDVI) and impervious surface fraction (ISF) gradient. Areas with either high vegetation cover or large impervious fractions exhibit elevated E_i, with vegetation height further modulating E_i under high-NDVI conditions. Random forest analysis identifies wind speed, vegetation structure (NDVI and vegetation height), and precipitation characteristics as the dominant controls on urban interception evaporation. Overall, the proposed RUEP framework provides a practical approach for quantifying interception evaporation in heterogeneous urban environments, offering new insights for improving urban hydrological modeling and supporting vegetation-informed stormwater management and urban design.