Shokri et al. (2026) Better continental-scale streamflow predictions for Australia: LSTM as a land surface model post-processor and standalone hydrological model

Identification

• Journal: Hydrology and earth system sciences
• Year: 2026
• Date: 2026-02-09
• Authors: Ashkan Shokri, James C. Bennett, David Robertson, Durga Lal Shrestha, Andrew J. Frost, Eric Lehmann
• DOI: 10.5194/hess-30-757-2026

Research Groups

• Commonwealth Scientific and Industrial Research Organisation (CSIRO), Australia
• Australian Bureau of Meteorology (BoM), Australia

Short Summary

This study evaluates two Long Short-Term Memory (LSTM)-based models (standalone and hybrid with AWRA-L) for continental-scale streamflow prediction in Australia, demonstrating their superior performance over traditional land surface and conceptual hydrological models across various validation scenarios. The findings highlight the potential of deep learning to enhance water resource management and climate adaptation strategies.

Objective

• To assess the performance of an AWRA-LSTM hybrid for streamflow predictions in Australia.
• To establish the performance of LSTMs as both a post-processor for AWRA-L and a standalone hydrological model in Australia.
• To investigate optimal implementation of the AWRA-LSTM hybrid, including the choice of dynamic and static predictors.
• To diagnose the relative contributions of bias-correction and routing improvements to LSTM performance.

Study Configuration

• Spatial Scale: Continental-scale Australia, utilizing 218 minimally impacted catchments from the CAMELS-AUS dataset. Gridded data (AWRA-L, AGCD) at 0.05° resolution.
• Temporal Scale: Daily time step, with data spanning 1975–2014 (CAMELS-AUS v1) and 1975–2022 (CAMELS-AUS v2). AWRA-L historical outputs from 1911 onward. LSTM models used a sequence length of 365 days.

Methodology and Data

• Models used:
  • Long Short-Term Memory (LSTM) networks:
    • LSTM-C: Standalone rainfall–runoff model.
    • LSTM-QC: Hybrid model incorporating AWRA-L runoff, acting as a post-processor for AWRA-L.
  • Australian Water Resources Assessment–Landscape model (AWRA-L v7): Land surface model.
  • GR4J: Conceptual rainfall–runoff model (used for benchmarking).
• Data sources:
  • CAMELS-AUS dataset (v1 and v2): Observed daily streamflow (predictand), catchment-averaged rainfall (from Australian Gridded Climate Data - AGCD), potential evapotranspiration (from SILO database), and 12 static/quasi-static catchment attributes (e.g., catchment area, mean slope, mean annual precipitation, aridity).
  • AWRA-L v7 outputs: Gridded total runoff (Qtot) as a dynamic predictor for LSTM-QC.

Main Results

• Both LSTM-QC and LSTM-C consistently outperformed AWRA-L and the conceptual GR4J model in streamflow predictions across most Australian catchments.
• Under temporally out-of-sample (TooS) cross-validation, fine-tuned LSTM-QC significantly outperformed GR4J in 77.5% of catchments, leveraging information from AWRA-L.
• Under spatially out-of-sample (SooS) and spatiotemporal out-of-sample (TSooS) cross-validation, LSTM-QC generally outperformed GR4J (69% and 67% of catchments, respectively), with LSTM-C performing comparably well, suggesting effective generalization of purely data-driven approaches.
• Fine-tuning LSTM models to local catchment data significantly improved performance, benefiting 95.4% of catchments in TooS experiments.
• Analysis of sequence length revealed that LSTM-QC's improvements over AWRA-L stem from both systematic bias correction (even at short sequence lengths) and enhanced channel routing signals (with longer sequence lengths).
• Recalculating quasi-static climatic predictors for each cross-validation fold was found to be crucial to prevent information leakage and ensure accurate performance assessment.

Contributions

• First comprehensive assessment of an AWRA-LSTM hybrid model for continental-scale streamflow prediction in Australia.
• Demonstrates the superior predictive skill of LSTM-based models (both standalone and hybrid) compared to established land surface (AWRA-L) and conceptual (GR4J) models in an Australian context.
• Provides insights into optimal LSTM implementation, including the value of fine-tuning and the careful selection and cross-validation of dynamic and static predictors.
• Decomposes the sources of improvement from LSTM post-processing, distinguishing between bias-correction and routing enhancement.
• Validates the robustness and generalizability of LSTM models for critical hydrological applications, including historical reconstructions, predictions in ungauged basins, and as a proxy for climate projection scenarios.

Funding

• Commonwealth Scientific and Industrial Research Organisation (CSIRO) AquaWatch Australia Mission
• CSIRO AI4Missions program
• National Water Commission (Murray-Darling Basin Sustainable Yields project)

Citation

@article{Shokri2026Better,
  author = {Shokri, Ashkan and Bennett, James C. and Robertson, David and Shrestha, Durga Lal and Frost, Andrew J. and Lehmann, Eric},
  title = {Better continental-scale streamflow predictions for Australia: LSTM as a land surface model post-processor and standalone hydrological model},
  journal = {Hydrology and earth system sciences},
  year = {2026},
  doi = {10.5194/hess-30-757-2026},
  url = {https://doi.org/10.5194/hess-30-757-2026}
}