Bani et al. (2026) Application of artificial intelligence-based modelling to investigate spring streamflow predictability under ENSO and IOD forcing

Identification

• Journal: Modeling Earth Systems and Environment
• Year: 2026
• Date: 2026-02-28
• Authors: Sabrina Bani, Monzur Alam Imteaz, Md. Iqbal Hossain, Patrick Morrison
• DOI: 10.1007/s40808-026-02743-6

Research Groups

• Department of Civil & Construction Engineering, Swinburne University of Technology, Melbourne, Australia
• Monash University, Melbourne, Australia

Short Summary

This study developed Artificial Neural Network (ANN) models, driven by lagged El Niño Southern Oscillation (ENSO) and Indian Ocean Dipole (IOD) indices, to forecast spring streamflow in Victoria, Australia. The ANN models consistently and substantially outperformed traditional Multiple Linear Regression (MLR) across diverse catchments, demonstrating enhanced predictive accuracy and better representation of nonlinear climate-streamflow interactions.

Objective

• To investigate spring streamflow predictability in Victoria, Australia, by developing Artificial Neural Network (ANN) models driven by lagged El Niño Southern Oscillation (ENSO) and Indian Ocean Dipole (IOD) climate predictors.
• To compare the performance of the ANN models against a Multiple Linear Regression (MLR) benchmark, particularly focusing on capturing nonlinear and time-lagged relationships under strict chronological validation.

Study Configuration

• Spatial Scale: Victoria, Australia (approximately 227,100 km²), focusing on nine streamflow gauging stations across eastern, central, and western regions. Catchment areas ranged from 163.0 km² to over 1953.0 km².
• Temporal Scale: Seasonal mean streamflow for spring (September–November, SON) over 58 water years (1966–2023). Lagged climate predictors were considered at 1-month, 3-month seasonal average, and 6-month half-year average antecedent time frames.

Methodology and Data

• Models used:
  • Artificial Neural Network (ANN) with a feed-forward architecture, trained using the Levenberg–Marquardt (LM) algorithm.
  • Multiple Linear Regression (MLR) as a benchmark model.
• Data sources:
  • Streamflow: Monthly mean discharge (Q) in cubic meters per second (m³ s⁻¹) from the Victorian Water Measurement Information System (WMIS), managed by the Department of Energy, Environment, and Climate Action (DEECA).
  • Climate predictors: Niño 3.4 index (for ENSO) and Dipole Mode Index (DMI) (for IOD), sourced as monthly values from http://climexp.knmi.nl/.
  • Data preprocessing: Missing streamflow records (less than 3%) imputed, all variables normalized using min–max scaling to a [0, 1] range.
  • Data splitting: Chronological partitioning (70% training, 15% validation, 15% testing) and (75% training, 15% validation, 10% testing) to ensure temporal independence. An ensemble-based ANN framework with 150 independent runs was used.

Main Results

• Superior ANN Performance: ANN models consistently outperformed MLR across all stations and lag configurations.
  • Error Reduction: ANN achieved significantly lower Mean Squared Error (MSE) (typically 0.008–0.035) compared to MLR (0.02–0.06), alongside reduced Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE).
  • Correlation Improvement: ANN models showed markedly stronger agreement with observed streamflow, with Pearson Correlation Coefficients (R) generally exceeding 0.70 and reaching up to 0.91, whereas MLR correlations remained weak to moderate (R = 0.17–0.47).
• Regional Performance:
  • Eastern Catchments: ANN validation and testing R values of 0.68–0.86, with MSE reduced from 0.04–0.06 (MLR) to 0.015–0.03 (ANN).
  • Central Regions: ANN predictive accuracy increased from R values of 0.18–0.38 (MLR) to 0.60–0.88, with corresponding RMSE decreases from approximately 0.17–0.22 to 0.10–0.15.
  • Western Catchments: ANN reduced MAE from 0.11–0.18 (MLR) to 0.08–0.12.
• Lagged Predictor Effectiveness: Predictive performance was strongest and most stable at short-to-intermediate lead times (1–3 months), with a physically consistent attenuation at the 6-month lag due to diminishing teleconnection strength and increasing climatic uncertainty.
• Bias Control: ANN models maintained physically plausible bias levels, with Percent Bias (PBIAS) values typically within ±3.5% during validation and testing, unlike MLR which exhibited zero PBIAS due to its mean-preserving nature.

Contributions

• Demonstrates that integrating lagged ENSO and IOD predictors within an ANN framework significantly enhances the predictability of spring streamflow in Victoria, Australia, compared to linear models.
• Advances nonlinear hydroclimatic modeling approaches by effectively representing delayed teleconnection effects on seasonal streamflow.
• Introduces a robust ANN framework designed for chronological data split analysis, explicitly mitigating overfitting and providing a pragmatic balance between computational efficiency, interpretability, and predictive accuracy.
• Offers a framework that provides deterministic estimates of seasonal streamflow magnitude, enabling direct attribution of forecast skill to climate drivers, antecedent lag structures, and station characteristics.
• Provides insights into climate-streamflow interactions in southeastern Australia, with potential transferability to other teleconnection-sensitive regions globally.

Funding

Open Access funding enabled and organized by CAUL and its Member Institutions. Specific projects, programs, and reference codes for funding are not explicitly stated in the provided text.

Citation

@article{Bani2026Application,
  author = {Bani, Sabrina and Imteaz, Monzur Alam and Hossain, Md. Iqbal and Morrison, Patrick},
  title = {Application of artificial intelligence-based modelling to investigate spring streamflow predictability under ENSO and IOD forcing},
  journal = {Modeling Earth Systems and Environment},
  year = {2026},
  doi = {10.1007/s40808-026-02743-6},
  url = {https://doi.org/10.1007/s40808-026-02743-6}
}