Application of physics-informed neural network approach in soil moisture retrieval using GNSS reflectometry

This research aims to obtain soil moisture from reflected GNSS signals using physics-informed neural networks (PINN). GNSS reflectometry (GNSS-R) signals can be considered as a new remote sensing source to study soil moisture. Despite the high sensitivity between GNSS reflected signal power and soil moisture, the model between measurements and parameters is difficult to solve mathematically due to the complexity of the electromagnetic relationships. Although Neural Network (NN) algorithms have been applied successfully in GNSS-R soil moisture retrieval, neural networks are trained without respecting any laws of physics. In this work, a new framework referred to as “physics-informed neural networks (PINN)” was used which adds governing physical relationships between data parameters to neural networks to generate more robust models, with less data.

The proposed research advances GNSS-R soil moisture estimations, exploiting Cyclone Global Navigation Satellite Systems (CYGNSS) satellite signals using PINN methodology. In PINN’s structure, reflected GPS signals from CYGNSS and land surface geophysical parameters are used as input features. Since reflected signal power variations are not only sensitive to changes in soil moisture, but also to changes in vegetation, surface roughness, soil texture, and elevation angle, the effects of land surface geophysical parameters involved in physical relationships are considered in the model. For reference data, soil moisture measurements of the International Soil Moisture Network (ISMN) were used in both training and validation.

The proposed PINN model generates daily soil moisture values with a root mean squared error (RMSE) of 0.05 〖cm〗^3/ 〖cm〗^3, which is an improvement from 0.0774 〖cm〗^3/ 〖cm〗^3 for the underlying NN model due to adding physical models. Four different soil dielectric constant models (Dobson, Hallikainen, Mironov, and Wong models) have been used to investigate the impact of soil dielectric constant models as part of physical relations. The RMSE distinction and correlation coefficient difference of the best model (Hallikainen) and worst model (Mironov) is 0.02 and 0.13, respectively demonstrating PINN sensitivity to different soil dielectric constant models. Consequently, the soil dielectric constant model selection influences overall PINN results. Thus, calibration of soil dielectric models is necessary for GNSS-R soil moisture retrieval in the future.