Associations between deep learning runoff predictions and hydrogeological conditions in Australia

• Catchment attributes provide context for interpreting long short-term memory (LSTM) prediction performance. • This multiple-catchment model integrates hydrogeological domain knowledge. • Catchment characteristics include climate, topography, geology and surface conditions. • Strengths: frequent or variable rainfall; hilly terrain; low permeability. • Weaknesses: flat land; sporadic, slow flows; loss to groundwater; summer wet season. To capture the complexity of hydrological systems across regions, multidimensional domain knowledge (e.g. climate, soils, geology and topography) can be incorporated into deep learning models of streamflow behaviour. Such integration has demonstrated notable improvements in streamflow predictions, thereby enhancing accuracy and offering valuable insights for sustainable water resource management. However, this catchment-specific domain information also holds potential for assessing the suitability of deep learning models for runoff predictions under varied conditions. This study explores the wide-ranging performance of deep learning streamflow predictions across the diverse landscape of Australian catchments through the leveraging of newly-available, comprehensive hydrological and hydrogeological datasets. Data from CAMELS-AUS (the Australian adaptation of CAMELS [Catchment Attributes and MEteorology for Large-sample Studies]) and a nationwide set of hydrogeological catchment attributes are integrated at a continental scale to probe associations between deep learning prediction performance and catchment attributes. The study encompasses three steps: 1) unsupervised learning to identify common patterns of catchment attributes; 2) a continent-wide, deep learning time series model (long short-term memory [LSTM]) incorporating catchment attributes into concurrent predictions across hundreds of basins; and 3) visualising and investigating associations between high (or low) runoff prediction performance and various catchment attributes. The resulting visual analytical tool provides insights into continent-wide differences in performance and also facilitates analysis at the individual catchment level. Key findings reveal a) enhanced LSTM performance in catchments characterised by frequent or variable rainfall, hilly terrain, and low permeability; and b) challenges encountered by the LSTM in flat catchments with slow, infrequent flows, high permeability, and in predicting runoff peaks in regions of substantial summer rainfall. Understanding these performance patterns can help inform the application of global LSTMs in water resource management and hydrological forecasting. Future work may involve assessing how such domain knowledge could improve the extrapolation of predictions to ungauged catchments within each attribute cluster. This multi-catchment study highlights the scalability advantages of machine learning techniques for gaining hydrological insights at a continental scale.