Unravelling groundwater time series patterns: Visual analytics-aided deep learning in the Namoi region of Australia

Understanding the sustainability of current groundwater extractions is critical in the face of changing climate and anthropogenic conditions, but this proves challenging in areas with complex, and not well understood, hydro geology. A combination of unsupervised (self-organizing map, SOM) and supervised (long short-term memory, LSTM) models is demonstrated here to effectively abstract prevalent patterns from a diverse set of groundwater monitoring time series in the dry and hydrogeologically complicated Namoi region, enabling predictions of water levels based on climate and anthropogenic conditions to be made using a set of regional deep-learning based neural networks. By drawing on shared pattern information from across the Namoi system, the SOM reduces the complexity of the multiple time series, shares information between sparse time series which could not be modelled with the LSTM individually, adds a spatial aspect to the LSTM analysis, and provides a valuable visual analysis that enhances communication and decision-making.