Comparison of the use of a physical-based model with data assimilation and machine learning methods for simulating soil water dynamics

摘要

Soil moisture plays a critical role as an essential component of the global water resources by regulating mass and energy exchange between land surface and atmosphere. Quantification of these exchange processes requires accurate characterization and simulation of soil water movement. Physically-based models (PBMs) and machine learning methods (MLMs) can both be used in soil moisture simulation. However, their performances in soil water simulation have only been compared in a limited number of cases. Moreover, almost all of them are conducted in field studies each with fixed soil, initial condition, and boundary condition. Here, we developed three artificial neural network (ANN) frameworks, and made clearer and more systematic comparisons between them and a PBM-Ross numerical model solving Richards equation and parameter estimation using a data assimilation approach (iterative ensemble smoother, Ross-IES) in synthetic and real-world conditions. Compared with the ANNs, Ross-IES is more significantly affected by physical model uncertainties such as soil heterogeneity, initial and boundary conditions, while both methods are affected by observation noise. For Ross-IES, the errors from boundary conditions and hydraulic parameter conceptualization are found to be more prominent than that of observation noise and therefore are suggested to be identified first. Meanwhile, the ANNs have difficulty in simulating the peaks and troughs of the soil water time series as well as in situations where the soil moisture is constantly saturated. ANNs yield a superior simulation when the nonlinear relationship between the response variables and driving data is weak, while the performance of Ross-IES is governed by the prior soil hydraulic information. In addition, Ross-IES approach requires much higher computational cost than the ANNs. ANN-MS performs best among the three ANN-based machine learning models and demonstrates great data mining ability and robustness against overfitting.