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

Abstract. Many European countries mainly rely on groundwater for domestic water use. 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 to estimate wtd anomalies (wtd_a) using other hydrometeorological observations routinely available near real-time. In this work, we explore the potential of Long Short-Term Memory (LSTM) networks to produce monthly wtd_a, using monthly precipitation anomalies (pr_a) as input. LSTM networks are a special category of artificial neural networks, useful in detecting a long-term dependency within sequences, in our case time series, which is expected in the relationship between pr_a and wtd_a. To set up the methodology, spatio-temporally continuous data were obtained from daily terrestrial simulations (hereafter termed the TSMP-G2A data set) with a spatial resolution of 0.11°, ranging from the year 1996 to 2016. They 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 wtd_a maps from LSTM networks agreed well with TSMP-G2A wtd_a maps in 2003 and 2015 constituting drought years over Europe. Moreover, we categorized test performances of the networks based on yearly averaged wtd, evapotranspiration (ET), soil moisture (θ), snow water equivalent (S_w), and soil type (S_t) 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 S_w < 10 mm, revealing a significant impact of the local factors on the ability of the networks to process information. Furthermore, results of cross-wavelet transform (XWT) showed a change in the temporal pattern between TSMP-G2A pr_a and wtd_a at some selected pixels, which can be a reason for undesired network behavior. Our results demonstrate that LSTM networks are useful to produce high-quality wtd_a based on other hydrometeorological data measured and predicted at large scales, such as pr. This contribution may facilitate the establishment of an effective groundwater monitoring system over Europe relevant to water management.