Adombi (2026) Ensembling or not? On the value of boosting and bagging for physics-aware machine learning in hydrology

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-09-10
• Authors: Adoubi Vincent De Paul Adombi
• DOI: 10.1016/j.jhydrol.2026.136397

Research Groups

• R2Eau, Centre d’´etudes sur les ressources min´erales, Universit´e du Qu´ebec `a Chicoutimi

Short Summary

This study investigates the value of ensemble machine learning strategies when applied to physics-aware machine learning models for hydrological prediction. The results show that ensemble strategies do not yield substantial improvements over a properly configured standalone PaML model.

Objective

• Investigate whether ensemble machine learning strategies provide tangible added value when applied to physics-aware machine learning models for hydrological prediction.

Study Configuration

• Spatial Scale: Catchment scale, focusing on five representative catchments from the CAMELS-US dataset.
• Temporal Scale: Daily time step, with a focus on streamflow simulation and prediction.

Methodology and Data

• Models used:
  • Physics-aware machine learning (PaML) models as base learners for ensemble construction.
  • Bagging and boosting ensemble strategies developed using PaML as the base learner.
  • Classical sequential boosting approach and growth-based boosting strategy inspired by GrowNet.
• Data sources: CAMELS-US dataset, providing hydrological, meteorological, and catchment information for 671 river basins across the contiguous United States.

Main Results

• Ensemble strategies do not yield substantial improvements over a properly configured standalone PaML model.
• Bagging provides limited additional benefits.
• Growth-based boosting formulation exhibits non-monotonic and basin-dependent behavior.
• Classical boosting can compensate for intentionally reduced model capacity by aggregating low-capacity PaML learners, but the resulting gains remain marginal relative to a moderately expressive single model.

Contributions

• This study contributes to clarifying the practical role of ensemble learning in physics-aware hydrological modeling.
• The findings suggest that refining the architecture and capacity of a single PaML model is generally more effective than adopting ensemble strategies.

Funding

• Not specified in the provided text.

Citation

@article{Adombi2026Ensembling,
  author = {Adombi, Adoubi Vincent De Paul},
  title = {Ensembling or not? On the value of boosting and bagging for physics-aware machine learning in hydrology},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136397},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136397}
}