Improving soil heat and moisture forecasting for arid and semi-arid regions: A comparative study of four mathematical algorithms

Abstract

ABSTRACTMathematical modeling is extensively used for ecohydrological processes because it facilitates data acquisition. However, modeling of soil moisture and heat remains challenging in dry ecosystems. In this study, we examined the performance of four models in simulating hydrological processes in a semi-arid mountain grassland (SMG), and in shrubland forming a transitional zone between the desert and an oasis (desert–oasis ecotone; DOE) in northwestern China. We used precipitation, air temperature, humidity, atmospheric pressure, and other meteorological variables to estimate moisture and temperature at different soil depths. Four methods were used to test model performance, including partial least squares (PLS) regression, stepwise multiple linear regression (SMR), back-propagation artificial neural network (BPANN), and neural network time series. Our results showed that BPANN had the best prediction accuracy and supplied a robust modeling framework capable of capturing nonlinear environmental processes by improving the stability of the weight-learning process. Soil depth in SMG for which model performance was optimized was 20 cm for PLS and SMR. Additionally, artificial neural networks (ANNs) have a remarkable applicability compared to other algorithms for increased accuracy in time-series predictions; however, they could not depict soil moisture or temperature dynamics at 160 cm depth in SMG, and at 10 cm depth in DOE. Using conventional meteorological data as primary predictors, and avoiding the complexity of distributed hydrological models can be helpful in developing a regional capacity for soil moisture and heat forecasting.