Infiltration rate assessment of experimental rain garden using conventional models and machine learning algorithms

Rain gardens are recent best management practice adopted for mitigating urban floods. However, its design parameters lack a rational basis. The present study employs conventional and machine learning algorithms to predict the rain garden's infiltration characteristics. Four small rain gardens were constructed in the hydraulics laboratory of the Civil Engineering Department, NIT Kurukshetra, Haryana, India to perform experiments on perennial flower plants. The study includes three conventional models (Kostiakov model, Philips Model, and Multi-linear Regression) and two machine learning algorithms (Multivariate adaptive regression spline (MARS) and Group method of data Handling) for prophecies the infiltration rate of rain garden. Air temperature, soil water content, water inflow rate to a rain garden, number of plants, plant height, and time elapsed in various water level drops of accumulated water in rain garden are input parameters and infiltration rate as output parameter has been used for modeling. 75 % of experimental data was for training purposes and the rest was for validation. The performance of conventional and machine learning algorithm models was assessed using the statistical coefficients viz. correlation (CC), root mean square error (RMSE), and Nash-Sutcliffe efficiency (NSE). It is observed that the performance of the MARS is the best among all the models. For training data set values of CC, RMSE and NSE are obtained as 0.998, 0.129, and 0.996, respectively. And for validation data set values of CC, RMSE and NSE are obtained as 0.977, 0.141, and 0.948. The study's outcome will be useful in the accurate assessment of the infiltration rate of the rain garden.