Abstract: Low-lying plain polders, characterized by diverse underlying surfaces, present significant challenges for hydrological modeling and forecasting due to the complexity of runoff generation and concentration. To address these challenges, we developed a runoff model that accounts for multiple underlying surface types, including paddy fields, dry lands, forests, urban areas, and water surfaces. The MIKE 11 HD model was employed to simulate rive network flows, and a hydrological-hydrodynamic coupling model was proposed specifically for plain polders with complex underlying surface conditions. To further improve the model's accuracy, a BP neural network was integrated to correct forecasting errors. The model's performance was evaluated in the Nansha Jiaomen River drainage area in Guangzhou. In addition, using the “9·7 Shenzhen Rainstorm” of 2023 as the input scenario, a case study was conducted to simulate the impact of different drainage measures on river water levels. Results show that the model accurately simulates river water level dynamics during flood events, with average Nash efficiency coefficients of 0.86 and 0.91 during the calibration and validation periods, respectively. For 8 out of 10 flood events, the maximum water level simulation error was less than 0.05 m. After correction by the BP neural network, all flood events achieved a Nash efficiency coefficient greater than 0.9, meeting the required accuracy of flood forecasting. The study also highlights the risk of waterlogging under the simulated scenario, underscoring the need to enhance flood retention and drainage capacity within the polders.