Soil Matrix Infiltration Measurement and Modelling Across Diverse Land Uses With Machine Learning

ABSTRACT Evaluating soil matrix infiltration is essential for understanding water dynamics, runoff and groundwater recharge. Experiments using a double‐ring infiltrometer with and without fine sand were conducted in forestland, tea gardens and paddy fields. A methylene blue tracer helped analyse soil profiles to detect macropores. Fine sand effectively blocked macropores in forestland and tea gardens but was absent in paddy fields, likely due to annual puddling, high moisture, clay content and bulk density. The basic infiltration rate (BI) remained consistent across land types, whereas accumulated infiltration depth (AI) was influenced by fine sand, except in paddy fields. Data collection was labour intensive, leading to regression models predicting infiltration based on soil properties. Three machine learning techniques (partial least squares regression [PLSR], the group method of data handling [GMDH] and random forest [RF]) were evaluated via the coefficient of determination ( R 2 ) and root mean square error (RMSE) metrics. Model 2 of PLSR performed best for forestland ( R 2 : 0.933, RMSE: 0.908 cm) and tea gardens ( R 2 : 0.838, RMSE: 1.156 cm), whereas Model 1 of PLSR was optimal for paddy fields ( R 2 : 0.780, RMSE: 0.370 cm). GMDH yielded the best overall results, with RF ranking third but remaining useful.