A benchmark comparison of AI-based modeling of soil infiltration rates

ABSTRACT Infiltration is crucial in the hydrological cycle, serving as the primary process that increases soil moisture. This study investigates soil infiltration rate (IR) prediction using various techniques, including GMDH, Gaussian Process, SVM, ANN, and MARS. 190 field observations were collected from Alashtar sub-watersheds in Lorestan, Iran. 70% of the observations were used for model preparation, while 30% were used for validation. The input variables for the study are Time, Sand, Clay, Silt, pH, Electrical Conductivity, Moisture Content, Soil Bulk Density, Porosity, Calcium Carbonate, Phosphorus, Organic Carbon, Organic Matter, Nitrogen, and Temperature, while IR is the output variable. Obtained results indicate that the ANN has a higher accuracy with coefficient of correlation values as 0.9366, 0.8624, mean absolute error values as 0.0607, 0.1000, Nash Sutcliffe model efficiency values as 0.8732, 0.7350, scattering index values as 0.3108, 0.5003, and Legates and McCabe's Index values as 0.6585, 0.5654 by using training and testing data sets, respectively. A sensitivity analysis highlighted that time is the parameter that most influences estimating the IR. The study underscores the precision of ANN in predicting soil infiltration rates and the need for AI-based models in hydrological models to improve accuracy and reliability in IR prediction.