Using neural networks for predicting soil water storage based in situ soil moisture observations

Previous studies have shown that Long-Short Term Memory networks (LSTMs) offer a large potential for data-based learning and hydrological predictions. This study focuses on exploring the largely untapped potential of such modeling approach using a large sample of in-situ soil moisture data based on Time-Domain Reflectometry (TDR) measurements, collected across the Attert experimental basins in Luxembourg. Soil moisture plays a critical role in various hydrological processes, influencing groundwater recharge, governing infiltration dynamics, and contributing significantly to the generation of overland flow. Additionally, it stands as a key determinant for the water supply essential for sustaining vegetation and agricultural crops. Here we introduce an LSTM model that has been trained on extensive long-term in-situ soil moisture observations with the objective of extrapolating the dynamics of soil moisture across spatial dimensions, temporal scales, and depths. A key challenge in this context is how to deal with multiscale variability of TDR observations, which arising from small scale variations in soil texture scales as well as larger scale spatial variability of physiographic and meteorological characteristics. Acknowledging, that this multiscale variability of soil moisture is difficult to disentangle by standard available predictors and their gradients, we place particular emphasis on data processing and understanding of such variability. This emphasis is crucial to mitigate potential confusion within the model, ensuring a more accurate representation of soil moisture dynamics. For this purpose, we evaluate the efficacy of LSTMs in capturing soil moisture dynamics, while concurrently aiming to clarify variability and address uncertainty. Furthermore, employing clustering techniques and network theory approaches, our aim is to discern systematic variability and patterns, considering model performance and the relationships within soil moisture measurement time series. As a result, we demonstrate the advantages of employing LSTMs to assess soil moisture dynamics at the catchment scale, while emphasizing the exploration of drawbacks and limitations inherent in purely data-based learning. This analysis provides a valuable guide for future modeling attempts, offering an opportunity to depict spatial and temporal variations in soil water storage. Such representations prove beneficial in the development of early-warning systems for potential dry events.