Application of physics informed neural networks to near-surface soil moisture dynamics

Abstract:

Soil moisture data is vital for many applications, including weather forecasts, hydrological modeling, and agricultural irrigation. Especially, near-surface soil moisture data is essential because of the interaction with the atmosphere and biosphere. Although a large amount of soil moisture data is becoming available through advanced remote sensing and direct sensing methods, the data analysis using models has not progressed as rapidly. The main factors that inhibit modeling progress include soil heterogeneity and difficulties in obtaining initial and boundary conditions for modeling purposes. We demonstrated the potential of physics informed neural networks (PINNs) for estimating soil hydraulic properties from volumetric water content without the need for prior knowledge of soil properties and initial and boundary conditions (Bandai and Ghezzehei, 2020 under review). However, the original PINNs method’s ability to approximate the solution of the Richardson-Richards equation (RRE) in describing soil moisture dynamics showed some weaknesses. Furthermore, the results of the PINNs were sensitive to the structure of neural networks. Here, we present an improved neural network structure and a novel learning algorithm proposed by Wang et al., 2020. We tested the revised model’s ability to obtain the forward solution of the RRE given initial and boundary conditions and compared the results with the previous implementation of PINNs. We then show a practical application of the PINNs to analyze soil moisture dynamics at the near-surface using synthetic soil moisture data.