Wavelet denoising of GNSS time series through Machine Learning. Application to the Campi Flegrei caldera

The great potential of the Global Navigation Satellite System (GNSS) in monitoring ground deformation in different geodynamic contexts (e.g., plate kinematics, active volcanic and seismic areas) is nowadays widely recognized. GNSS time series, indeed, can provide unique information on the ongoing dynamics of a geological system and can help assess the associated hazard. However, as for other geophysical data, time series retrieved from permanent GNSS stations can be severely affected by noise, which can sometimes reach such large amplitudes as to hide elusive ground deformation signals. For this reason, over the years several denoising techniques aimed at improving the signal-to-noise ratio have been proposed for detecting low amplitude signals. One of the most effective denoising techniques has proven to be the Wavelet decomposition. This is a multiresolution decomposition of the signal through the discrete wavelet transform (DWT). Through the DWT, it is possible to decompose a signal into a series of orthogonal wavelets. The input signal is spectrally decomposed into distinct bands and the information in each band can be analyzed independently according to the characteristics of the source. However, Wavelet analysis requires long series of data in order to be effective and its calculation can be time-consuming. These characteristics hampers its use as a real-time monitoring tool. In the present study, we aim to overcome these limitations by training a neural network to perform the equivalent of wavelet analysis on GNSS data. The proposed approach can be summarized as follows: i) the wavelet analysis is performed on GNSS data coming from different sites in a permanent network; ii) a neural network is trained using original time-series as input and “Wavelet processed” series as target; iii) the trained model is used on newly recorded GNSS data to perform real-time denoising. Our analysis focuses on GNSS time series collected over the last twenty years at Campi Flegrei (Naples, Italy), a volcanic caldera world-renowned for its slow ground deformation, called bradyseism. The preliminary results are promising as the trained model shows a high accuracy with test data (newly collected data). We expect the quality of the prediction could even increase over time, as new data will be used to train the model.