Estimation of volumetric soil moisture from GNSS-IR and analysis of the resulting time series with LSTM artificial neural networks

One of the ways for measuring environmental parameters is using GNSS reflected signals from the Earth surface that are received by GNSS antennas. Environmental parameters include soil moisture, seasonal snow accumulation, ice thickness, vegetation cover and water level changes in dams, lakes and seas (tide). The focus of this research is on soil moisture.

Reflection of GNSS signals from a surface is called multipath. When the goal is positioning, multipath is one of the most significant sources of error in GNSS observations. But, by analyzing those reflected signals, we will get useful information about the reflection surface. This technique is called GNSS interferometric reflectometry (GNSS-IR). By this definition, GNSS-IR can be considered as a remote sensing technique for continuous and local monitoring of environmental parameters which can be performed in various weather conditions such as rainy and cloudy conditions, as well as different lighting conditions such as day and night.

Signal-to-Noise Ratio (SNR) is a measure of the strength of a signal relative to the background noise level. In GNSS, SNR is used to evaluate the quality of the received signal. It is calculated as the ratio of the power of the received signal to the power of the noise in the receiver's bandwidth. Some of the receivers can also record SNR data which includes SNR component of reflected signals.

In GNSS-IR, changes in soil moisture result in changes in the SNR component of the reflected signals. Specifically, as the soil moisture content increases, the dielectric constant of the soil increases, which causes the reflected signals to have higher amplitudes and higher SNR. Conversely, as the soil moisture content decreases, the reflected signals have lower amplitudes and lower SNR. Therefore, analyzing the SNR of the reflected signals can provide useful information about the soil moisture content. In addition to soil moisture, SNR can also be affected by other factors such as atmospheric conditions and receiver noise. Therefore, it is important to carefully analyze and process the SNR data to accurately estimate the soil moisture content.

The soil moisture algorithm to be used in this study is currently implemented at stations in the EarthScope PBO H2O network with the greatest variations in vegetation. Among the sites in the PBO H2O network, data from P038 in the New Mexico region is used. This site located in a flat area in a ecosystem characterized as grass land.

SNR data from the new L2C signals are used by this site because the quality of the data are higher than either the legacy L1 or L2P signals. The frequency of the L2C signal corresponds to a maximum penetration depth of 5 cm. The multipath signals are used to estimate soil moisture changes over a four-years period from 2017 to 2020.

The calculations were done in four main steps. In the first step, appropriate satellite tracks with elevation angle between 5 to 30 degrees were selected and SNR data were extracted from RINEX files. In the second step the initial reflector height is estimated for each track and then the phase is obtained for each satellite track on each day. In the third step, SNR metrics are calculated, and finally, vegetation cover effects are removed and the result is converted to volumetric water content. According to the estimations, the volumetric water content in 2017 was 8.88%, which increased to 11.74% in 2018, then slightly decreased to 10.88% in 2019 and finally increased to 12.49% in 2020. In the fourth step, the effectiveness of the LSTM neural network model in predicting the time series of volumetric soil moisture obtained from GNSS-IR signals is investigated. The LSTM neural network can maintain its content over a long period of time and essentially remember previous information. This prediction will help farmers to prepare their irrigation schedules more efficiently. For this purpose, it is suggested to use cheap GPS receivers in agricultural lands in rural areas. The model is trained using 80% of station observations. By updating the network status with observed values instead of predicted values, the root mean square error decreased from 0.09 to 0.04, and the predictions became more accurate.

Handling the location and type of receivers located in the Iranian Permanent GPS Network for Geodynamics (IPGN) and making the necessary settings in order to determine environmental parameters are suggested as by-products for IPGN.

Investigations have shown that performing GNSS observations produces more homogeneous reflective effects. Therefore, in order to increase the accuracy and quality of the results, it is suggested to use GNSS-IR instead of just GPS-IR.