GPR Signal processing in frequency domain using Artificial Neural Network for water content prediction in unsaturated subgrade

Basing on the recent outcomes of an investigation on the water content evaluation from processing the GPR signal in the frequency domain, an accurate model for the prediction of the frequency spectrum of the reflected GPR signal as the moisture content changes is proposed. This method uses an Artificial Neural Network approach. After the training, the ANN is reasonably able to predict the frequency spectrum for a water content in a specific soil (the error in the spectrum generation computed on a validation set, after 105 training epochs, is about or less than 10% for the same soil, if the ANN is used for all kinds of soil the error increases to about 15-20%). Of course this method can be inverted generating with the trained ANN a catalogue of different spectra for different values of the soil water moisture. Using this inverse approach it is possible to predict the water content of a specific soil just comparing the real current spectrum of the reflected GPR signal with the spectra of the catalogue. This method has been successfully tested on experimental data.