Does grouping watersheds by hydrographic regions offer any advantages in fine-tuning transfer learning model for temporal and spatial streamflow predictions?

Khoshkalam, Yegane ORCID: https://orcid.org/0000-0001-8885-936X; Rousseau, Alain N. ORCID: https://orcid.org/0000-0002-3439-2124; Rahmani, Farshid; Shen, Chaopeng ORCID: https://orcid.org/0000-0002-0685-1901 et Abbasnezhadi, Kian (2025). Does grouping watersheds by hydrographic regions offer any advantages in fine-tuning transfer learning model for temporal and spatial streamflow predictions? Journal of Hydrology , vol. 650 . p. 132540. DOI: 10.1016/j.jhydrol.2024.132540.

Ce document n'est pas hébergé sur EspaceINRS.

Résumé

Predicting accurate streamflow for data-limited regions and poorly gauged watersheds remains a global challenge. The complex calibration of physically based models (PBMs) and the substantial data requirements of machine learning models (MLs), such as Long Short-term Memory (LSTM) networks complicate this task. However, it is uncertain whether grouping watersheds by similar hydrological and ecological characteristics provides advantages in temporal and spatial prediction during fine-tuning. To answer this question, Transfer Learning (TL) based on LSTM modeling was used to transfer knowledge from 671 U.S. watersheds to 31 watersheds in Southern Québec while using outputs from a widely used PBM, HYDROTEL, and a data integration (DI) method using lagged streamflow. We ran experiments to determine: (i) whether temporal prediction performance of the TL model, when coupled with DI, and fine-tuning on the target region benefits from watershed grouping, (ii) the effectiveness of the TL model, improved with physics data from HYDROTEL (physics-guided TL) in predicting ungauged regions (PUR) with fine-tuning based on watershed grouping, and (iii) if the performance of physics-guided TL model in predicting ungauged basins (PUB) improved when more watersheds were randomly used for fine-tuning, irrespective of watershed grouping. Fine-tuning based on watershed grouping led to advanced performance (median Kling-Gupta-efficiency (KGE) value of 0.93) in temporal prediction experiments, surpassing or approaching the performance of the DI-TL model fine-tuned on the entire dataset. This advantage was more pronounced in TL than in local training, likely due to a limited target region dataset. In PUR experiments, performance of the physics-guided TL model was moderate (median KGE of 0.56), reflecting the challenge of spatial prediction and the little impact of watershed grouping. However, physics-guided TL model’s performance in PUB experiments matched (median KGE of 0.85) that of the calibrated HYDROTEL on the same watersheds, underscoring the need to collect more data in spatial prediction.

Type de document:	Article
Mots-clés libres:	LSTM; streamflow prediction; physically based model; HYDROTEL; deep learning; spatial prediction
Centre:	Centre Eau Terre Environnement
Date de dépôt:	01 avr. 2025 19:33
Dernière modification:	01 avr. 2025 19:33
URI:	https://espace.inrs.ca/id/eprint/16250

Gestion Actions (Identification requise)

Modifier la notice