The Application of Artificial Intelligence in Predicting the Effect of Gravel and Vegetation Cover on the Urban Runoff Volume Using Experimental Data

Low-Impact Development Best Management Practices (LID-BMPs) methods can be effective in solving problems related to surface runoff. This study investigates the effectiveness of LID-BMPs in urban runoff reduction through experimental and artificial intelligence approaches. Laboratory tests evaluated 10 surface treatments - including impervious control, sandy loam soil, gravel, gravel with one geocell layer, gravel with two geocell layers, gravel with three geocell layers (GGE3), rosemary vegetation (R), rosemary with geocell, turf, and turf with geocell under 6 rainfall intensities (45–200 mm/h) and 2 slopes (0%, 5%). Using a rainfall simulator, 7200 runoff volume measurements were collected and analyzed through four computational models: Artificial Neural Network, Multiple Linear Regression, Recurrent Neural Network, and Long Short-Term Memory (LSTM). Model inputs included bed slope, rainfall intensity, Runoff harvest times in each experiment (t), and coverage coefficient, with performance evaluated using correlation coefficient (R²), root mean square error (RMSE), Mean Absolute Percentage Error (MAPE), Normalized Root Mean Square Error (NRMSE) and mean absolute error (MAE). Key findings demonstrate that GGE3 achieved optimal runoff reduction (up to 99.6% versus control). The LSTM model outperformed others in predictive accuracy (R²=0.9965, MAE = 0.7093, RMSE = 1.1437, NRMSE = 1.5133, MAPE = 13.9253), with sensitivity analysis identifying t as the most influential parameter. These results provide actionable insights for urban stormwater management, combining empirical validation with advanced machine learning techniques to optimize LID-BMP implementation strategies. Graphical Abstract