Multivariate modeling of soil suction response to various rainfall by multi-gene genetic programing

Abstract

In this study, field monitoring testing and machine learning are used to analyze vegetated soil’s response to various rainfall events under natural environmental conditions. Parameters that reflect the soil, vegetation, and atmosphere of three monitoring points at different distances from a tree (0.5 m, 1.5 m, and 3.0 m) at a constant depth (0.2 m) are quantified and used for multivariate model development. A machine learning method, multi-gene genetic programming (MGGP), is used to formulate the relationships between two indices representing vegetated soil response and six selected influential parameters. Analysis indicates that for a complicated system, the MGGP method is suitable for establishing an efficient computational model under conditions of limited data. Global sensitivity analysis and parametric study are conducted, based on the obtained multivariate models, to reveal the effect of each influential parameter, indicating that rainfall pattern has much the same impact on variations in soil suction as rainfall amount and intensity and tree canopy do. An advanced rainfall pattern can trigger a more rapid response of vegetated soil than intermediate and delayed rainfall patterns can. Rainfall pattern’s effect on the descent rate of soil suction is nonlinear.