Using Long Short-Term Memory networks to connect water table depth anomalies to precipitation anomalies over Europe

Abstract

Abstract. Many European countries rely on groundwater for public and industrial water supply. Due to a scarcity of near-real-time water table depth (wtd) observations, establishing a spatially consistent groundwater monitoring system at the continental scale is a challenge. Hence, it is necessary to develop alternative methods for estimating wtd anomalies (wtda) using other hydrometeorological observations routinely available near real time. In this work, we explore the potential of Long Short-Term Memory (LSTM) networks for producing monthly wtda using monthly precipitation anomalies (pra) as input. LSTM networks are a special category of artificial neural networks that are useful for detecting a long-term dependency within sequences, in our case time series, which is expected in the relationship between pra and wtda. In the proposed methodology, spatiotemporally continuous data were obtained from daily terrestrial simulations of the Terrestrial Systems Modeling Platform (TSMP) over Europe (hereafter termed the TSMP-G2A data set), with a spatial resolution of 0.11∘, ranging from the years 1996 to 2016. The data were separated into a training set (1996–2012), a validation set (2013–2014), and a test set (2015–2016) to establish local networks at selected pixels across Europe. The modeled wtda maps from LSTM networks agreed well with TSMP-G2A wtda maps on spatially distributed dry and wet events, with 2003 and 2015 constituting drought years over Europe. Moreover, we categorized the test performances of the networks based on intervals of yearly averaged wtd, evapotranspiration (ET), soil moisture (θ), snow water equivalent (Sw), soil type (St), and dominant plant functional type (PFT). Superior test performance was found at the pixels with wtd < 3 m, ET > 200 mm, θ>0.15 m3 m−3, and Sw<10 mm, revealing a significant impact of the local factors on the ability of the networks to process information. Furthermore, results of the cross-wavelet transform (XWT) showed a change in the temporal pattern between TSMP-G2A pra and wtda at some selected pixels, which can be a reason for undesired network behavior. Our results demonstrate that LSTM networks are useful for producing high-quality wtda based on other hydrometeorological data measured and predicted at large scales, such as pra. This contribution may facilitate the establishment of an effective groundwater monitoring system over Europe that is relevant to water management.