Machine learning for monitoring groundwater resources over Europe

Groundwater (GW) is an important natural resource for Europe and the world, and has been affected by extreme weather and climate, e.g., summer heat waves and droughts, and human overexploitation. As climate change and human interventions increase, extreme events and GW depletion are expected to become more frequent and severe in many parts of Europe in the future, aggravating the vulnerability of GW systems. This emphasizes the necessity of GW monitoring in GW management. Up to date, however, it is still challenging to monitor GW at the large, continental scale, mainly due to the lack of water table depth (wtd) observations. In order to address the challenge, the PhD work proposes an indirect, generic methodology based on advanced machine learning (ML) techniques, that are Long Short-Term Memory (LSTM) networks andtransfer learning (TL), to produce reliable monthly wtd anomaly (wtda) estimates at the continental scale. The methodology is named LSTM-TL. While in this work, LSTM-TL has been implemented over Europe, it is transferable to other regions in the world. The methodology relies on the close connection between GW and other atmospheric and terrestrial compartments in the water cycle, using precipitation and soil moisture anomalies (pra and θa) as input, which have data available at large scales from, e.g., remotely sensed observations. Several steps were involved in the development of LSTM-TL for GW monitoring. In the first step, LSTM networks were applied in combination with spatiotemporally continuous pra and wtda data from uncalibrated integrated hydrologic simulation results (named the TSMP-G2A data set) over Europe to capture the time-varying and time-lagged relationship between pra and wtda in order to obtain reliable networks to estimate wtda at the individual pixel level assuming that pra is a useful proxy for wtda. In most European regions, LSTM networks showed good skill with respect to the TSMP-G2A data set inpredicting wtda with pra as input. The results indicated that the local factors, that are yearly averaged wtd, evapotranspiration (ET), soil moisture (θ), and snow water equivalent (SWE), had a significant impact on the performance of the LSTM networks. Moreover, the decrease in the network test performance at some pixels was attributed to a change in the temporal TSMP-G2A pra-wtda pattern during the study period. In the second step, a number of input hydrometeorological variables, in addition to pra, were included in the construction of LSTM networks to arrive at improved wtda estimates at individual pixels over Europe in various experiments. All input and target data were derived from the TSMP-G2A data set. Improved LSTM networks were found with pra and θa as input. Considering θa strongly increased the network testperformance particularly in the areas with wtd ≤ 3 m (i.e., the major wtd category of Europe), suggesting the substantial contribution of θa to the estimation of wtda over Europe. The results highlight the importance to combine θ information with precipitation information in quantifying and predicting wtda