Machine learning-based modeling and analysis of surface cover impacts on hydrological parameters influencing urban runoff

• Various covers, such as turfgrass and gravel reinforced with geocells, were tested under rainfall simulation. Vegetative and geocell covers effectively reduced runoff and increased time to runoff initiation. • Hydrological parameters were modeled using machine learning algorithms. XGBoost performed the best, reducing peak flow (PF) by up to 70% and runoff coefficient (C) by up to 65%. • Gravel with three layers of geocell (GGE3) was the most effective cover in reducing runoff, decreasing PF by up to 70%. • Sensitivity analysis showed that rainfall intensity and cover coefficient had the most significant impact on predicting runoff parameters. Urban development and the expansion of impervious surfaces have increased the volume and intensity of surface runoff, leading to flash floods. This study aims to evaluate the impact of various surface covers and develop an accurate model for predicting runoff parameters. A rainfall simulator was used to experimentally investigate four hydrological parameters: peak flow (PF), runoff coefficient (C), time to runoff (TR), and base time (TB). Ten cover types, including soils, turfgrass, rosemary plants, and gravel reinforced with geocell, were examined at slopes of 0% and 5% under six rainfall intensities. Hydrological parameters were modeled using seven machine learning algorithms: multilayer perceptron neural network (MLP), random forest (RF), gradient boosting (GB), bagging, support vector regression (SVR), XGBoost, and LightGBM. Experimental results showed that covers with three geocell layers (GGE3) and plant-geocell combinations effectively reduced runoff, decreasing PF by up to 70% and C by up to 65%, while increasing TR by over 80 minutes. Under heavy rainfall, GGE3 reduced TB by approximately 30% compared to impervious covers, improving surface drainage. In modeling, XGBoost demonstrated superior performance in predicting PF, C, and TB, achieving R² = 0.993, MAE ≈ 0.025, and NRMSE < 9% for C. For TB, XGBoost and GB yielded R² > 0.97 and NRMSE < 8%, showing high accuracy. For TR prediction, SVR in the testing phase achieved R² = 0.973 and MAE < 4 minutes, outperforming other models.