Better continental-scale streamflow predictions for Australia: LSTM as a land surface model post-processor and standalone hydrological model

Abstract. Accurate large-scale hydrological predictions are essential for water resource planning. However, many land surface models encounter difficulties in capturing streamflow timing and magnitudes, particularly in large catchments and when calibrated across broad regions and multiple hydrological variables. In this study, two Long Short-Term Memory (LSTM)-based approaches are assessed to enhance streamflow predictions across Australia: (i) LSTM-C, a standalone rainfall–runoff LSTM that relies solely on precipitation and potential evapotranspiration as inputs, and (ii) LSTM-QC, a rainfall–runoff LSTM that incorporates runoff outputs from the Australian Water Resources Assessment–Landscape model (AWRA-L), which can also be interpreted as a post-processor for AWRA-L. These approaches are tested in 218 minimally impacted catchments from the CAMELS-AUS dataset under three cross-validation strategies – temporally out-of-sample, spatially out-of-sample, and spatiotemporal out-of-sample – to evaluate their robustness for historical reconstructions, predictions in ungauged basins, and a proxy for climate-projection scenarios. The results indicate that both LSTM-QC and LSTM-C consistently outperform AWRA-L runoff across nearly all catchments and exceed the predictive skill of a widely used conceptual model (GR4J) in most basins. Under a temporally out-of-sample framework, LSTM-QC demonstrates a performance advantage over LSTM-C by leveraging information embedded in AWRA-L, particularly when fine-tuned to local catchment observed data. This advantage is primarily attributed to the LSTM's ability to correct systematic biases in AWRA-L and enhance channel-routing signals. However, under spatial and spatiotemporal cross-validation LSTM-C performs comparably well, suggesting that a purely data-driven approach can generalize effectively to ungauged or future conditions without reliance on AWRA-L.