Reconstruction of missing groundwater level data by using Long Short-Term Memory (LSTM) deep neural network

Monitoring groundwater level (GWL) over long time periods is critical in understanding the variability of resources present context global changes. However, Normandy (France) for example, GWLs have only been systematically monitored ~20 to 50 years. This study evaluates Long Short-Term Memory (LSTM) neural network modeling reconstruct GWLs, fill gaps and extend existing time-series. The approach illustrated by using available monitoring fluctuations piezometers implanted chalk aquifer region, Northern France. Here GWL data recorded years at 31 northwestern employed perform prediction. To optimize performance, most influential factors that impact accuracy prediction are first determined, such as architecture, quantity quality. resulting adopted measurements step with an increment missing observation time. requires no calibration time-lag processing implementation relies on retrieve targeted piezometers.