Evaluation of Deep Learning and Physically Based Models for Urban Streamflow Estimation Under Historical and Future Climate Conditions

Abstract Accurate streamflow estimation in urban catchments remains challenging under accelerating hydroclimatic variability and urbanisation. Conventional models such as Runoff-Routing-Burroughs (RORB), widely adopted in Australian practice, often struggle to represent the non-linear rainfall-runoff dynamics, governing urban flood responses. This study develops and evaluates a Deep-Learning-Neural-Network (DLNN) framework for streamflow estimation in the highly urbanised Gardiners Creek catchment, situated in southeastern Melbourne, using a dual-simulation-approach, integrating event-based and continuous-modelling. Historical rainfall and streamflow data (1989–2021) were utilised for calibration and validation against the RORB model, while future rainfall inputs were derived from dynamically downscaled ACCESSS1-0-CCAM projections under RCP4.5 and RCP8.5 scenarios from the Victorian-Climate-Projections dataset. Results demonstrate the DLNN’s improved overall predictive accuracy and adaptability (R²: 79.2–95.3%, NSE/KGE: 0.83–0.95, VE < 10%) relative to RORB (R²: 72.3–91.8%, NSE/KGE: 0.68–0.89, VE: −17.13% to − 13.20%). The DLNN-model effectively captured short-term flood peaks and long-term runoff variability, maintaining stability across diverse hydroclimatic regimes. Future projections indicate increased short-duration peak flows under RCP8.5, highlighting heightened flash-flood risks and limitations of current design frameworks. These findings establish DLNN as a robust, parsimonious, and climate-adaptive alternative for supporting flood prediction and resilient water resource management.