GNSS-IR soil moisture estimation using deep learning with Bayesian optimization for hyperparameter tuning

One of the ways for measuring water content is using global navigation satellite system (GNSS) multipath signals. By analyzing those signals, we will get useful information about the reflection surface. This technique is called GNSS interferometric reflectometry. Some receivers can record signal-to-noise ratio (SNR) data, which includes the SNR component of reflected signals and is strongly related to soil moisture. Here, we will use the data for station P038 in Mexico for 4 years, from 2017 until 2020. The calculation steps include extracting SNR data from RINEX files, estimating the prior reflector height and phase, calculating SNR metrics, and removing the vegetation effect to obtain volumetric water content (VWC). The results show that the VWC level has increased from 8.88 to 12.49% from 2017 to 2020. We have used long short-term memory neural networks with tuned hyperparameters by Bayesian optimization for predicting the time series of soil moisture. Our model is trained using 80% of the station observations. The accuracy of the network is checked using different metrics on the train, test, and all data. The mean absolute error, root mean square error, and a20-index of the test data are obtained as 0.002, 0.041, and 0.727, respectively. The modeling results will help farmers arrange their irrigation schedules more professionally.