Estimation of infiltration rate using data-driven models

Abstract

The infiltration rate is one of the primary processes of the hydrological cycle. It is the property of water by which it moves through the soil particles. Good knowledge of the infiltration rate is useful in calculating the natural and artificial groundwater recharge, soil erosion, and surface runoff. In this study, actual field measurements such as bulk density (B), water content (Wc), percentage of sand (Sa), silt (Si) and clay (C), and time (T) were used by five data-driven models to predict the infiltration rate of the soil. These data-driven models are multi-linear regression (MLR), neural network (NN), M5P model tree (M5P), random forest regression (RF), and Gaussian process (GP). The study area is situated in the Islamic Republic of Iran. The dataset contains 155 experimental measurements of the infiltration rate, which was collected using a double-ring infiltrometer. Out of 155 experimental observations arbitrarily chosen, 105 measurements were selected for training, whereas residual 50 were selected for testing the models. Three statistical parameters, root mean square error (RMSE), Nash-Sutcliffe model efficiency (NSE), and coefficient of correlation (C.C), were selected to compare the efficiency of all models. All data-driven models are capable of predicting the infiltration rate precisely. The comparative analysis of result suggests that the NN has very high performance with C.C = 0.9300, followed by GP, RF, M5P, and MLR (C.C = 0.9022, 0.8844, 0.6873, and 0.6016, respectively). Also, a comparison is made with the past studies, which indicated the NN model is best to predict the infiltration rate. Finally, the results revealed that time is the most important parameter for estimating the infiltration rate.