Simulating and predicting soil water dynamics using three models for the Taihu Lake region of China

Abstract

Drought stress under a changing climate can significantly affect agricultural production. Simulation of soil water dynamics in field conditions becomes necessary to understand changes of soil water conditions to develop irrigation guidelines. In this study, three models including Auto-Regressive Integrated Moving Average (ARIMA), Back-Propagation Artificial Neural Network (BP-ANN), and Least Squares Support Vector Machine (LS-SVM) were used to simulate the soil water content in the 0–14 cm and 14–33 cm soil layers across the Taihu Lake region of China. Rainfall, evaporation, temperature, humidity and wind speed that affect soil water content changes were considered in the BP-ANN and LS-SVM, but not in ARIMA. The results showed that the variability of soil water content in the 0–14 cm soil layer was greater than that in 14–33 cm. Correlation coefficients (r) of soil water content between simulations and observations were highest (0.9827) using LS-SVM in the 14–33 cm soil layer, while they were the lowest (0.7019) using ARIMA in the 0–14 cm soil layer; but no significant difference in r values was observed between the two soil layers with the BP-ANN model. Compared with the other two models, the LS-SVM model seems to be more accurate for forecasting the dynamics of soil moisture. The results suggested that agro-climatic data can be used to predict the severity of soil drought stress and provide guidance for irrigation to increase crop production in the Taihu Lake region of China.