Tumushabe et al. (2026) Enhancing river flow prediction accuracy in data scarce catchments using a time varying ensemble deep learning approach

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-09-09
• Authors: Angel I. Tumushabe, Seith N. Mugume, Johanna Sörensen
• DOI: 10.1016/j.jhydrol.2026.136387

Research Groups

• Department of Civil and Environmental Engineering, Makerere University, Uganda
• Department of Water Resources Engineering, Lund University, Sweden
• United Nations University Hub on Water in a Changing Environment (WICE) at Lund University, Sweden

Short Summary

This study evaluates the performance of a Time-Varying Dynamic Model Averaging (TV-DMA) ensemble for river flow prediction in two data-scarce tropical catchments. The TV-DMA approach achieved superior performance compared to Bayesian Model Averaging (BMA) and process-based HEC-HMS model.

Objective

• Investigate the effectiveness of Time-Varying Dynamic Model Averaging (TV-DMA) ensemble for river flow prediction in data-scarce tropical catchments.
• Compare the performance of TV-DMA with other machine learning and deep learning models, including Bayesian Model Averaging (BMA).

Study Configuration

• Spatial Scale: Two hydrologically distinct, data-scarce tropical catchments in the Upper White Nile Basin.
• Temporal Scale: Daily river flow time series.

Methodology and Data

• Models used:
  • Time-Varying Dynamic Model Averaging (TV-DMA) ensemble
  • Bayesian Model Averaging (BMA)
  • Feedforward Neural Network (FFNN)
  • 1D-CNN
  • LSTM
  • CNN-LSTM
  • Diff-FFNN-LSTM
• Data sources: Satellite, observation, and reanalysis data.

Main Results

• The TV-DMA ensemble achieved the highest individual-model river flow predictive performance (NSE ≥ 0.980) in both catchments.
• The heterogeneous, first-order differential processing Diff-FFNN-LSTM model consistently outperformed other models due to its ability to stabilise non-stationary river flow time series.

Contributions

• This study highlights the principal value of TV-DMA’s dynamic update of individual model weights based on predictive error distributions, which enhances river flow predictive performance.
• The findings demonstrate that TV-DMA can improve the accuracy and efficiency of river flow modelling in data-scarce regions.

Funding

• This research was funded by [project/program name and reference code]

Citation

@article{Tumushabe2026Enhancing,
  author = {Tumushabe, Angel I. and Mugume, Seith N. and Sörensen, Johanna},
  title = {Enhancing river flow prediction accuracy in data scarce catchments using a time varying ensemble deep learning approach},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136387},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136387}
}