Deep learning for earthquake hydrology? Insights from the karst Gran Sasso aquifer in central Italy

Abstract

By leveraging monitoring data for the Gran Sasso carbonate aquifer during two significant seismic sequences that hit central Italy in recent years, this study investigates the possibility of using memory-enabled deep learning algorithms as meaningful tools for an enhanced modelling of the hydrological response of karst aquifers subject to earthquake phenomena. Meteorological, hydrological and seismic data are used to train and validate long short-term memory networks (LSTM) in one- and multiple-day ahead flow forecasting exercises, aimed at assessing model sensitivities to input variables and modelling choices (training data and parameters of the models). Results indicate that the models fairly reproduce the flow patterns for the considered spring in the Gran Sasso aquifer, thus supporting the potential use of these models for hydrological applications in similar areas, provided that sufficient data are available for the training of the network.