Using physical method, machine learning and hybrid method to model soil water movement

Abstract

This study explores the performance of Phycically-based modelling (PBM), Machine learning (ML), and Hybrid modelling (HM) in soil water movement. Three types of models were tested on experiments under different soils and external pressure head conditions. In PBM, we proposed an adaptive step-length model named Time Cellular Automata (TCA), achieving an RMSE of 5.91, which outperforms HYDRUS (RMSE 7.92). In ML, Root Mean Square Error (RMSE) of all four tested models was below 1.5, with eXtreme Gradient Boosting (XGBoost) performing the best. The predictive accuracy of ML significantly outperformed PBM. SHapley Additive exPlanation was used to interpret the data and feature importance of machine learning. Middle-layer soil temperature, surface-layer soil salinity, water head and air temperature were identified as important parameters for ML. Heuristic algorithm can assist in searching for optimal parameters for TCA (Optimized TCA) and improve RMSE from 5.91 to 4.79. By integrating PBM and ML, developed a hybrid modeling strategy named HM. The HM was constructed using XGB and TCA, and achieved an error rate falling between Non-Optimized TCA (5.91) and Optimized TCA (5.51). This study presents a method for constructing HM from PBM and ML which is guided by data-driven approaches to make the analysis of soil water movement more efficient and economical.