Liu et al. (2026) Predicting Lake Surface Water Temperature With Transfer‐Based Physics‐Informed Deep Learning

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Water Resources Research
• Year: 2026
• Date: 2026-04-01
• Authors: Muyuan Liu, R. Iestyn Woolway, Chunqiang Xu, Yan Tong, Weijia Wang, Haoran Shi, N. Alsulaiman, James Tlhomole, Robert Ladwig, Matthew D. Piggott
• DOI: 10.1029/2025wr041062

Research Groups

Not explicitly mentioned in the abstract.

Short Summary

This study introduces Transfer-PIDL, a transfer learning framework, to enhance the generalizability of physics-informed deep learning (PIDL) for lake surface temperature prediction, demonstrating superior accuracy and reduced data requirements across diverse lakes.

Objective

• To develop and evaluate a transfer learning framework (Transfer-PIDL) that improves the generalizability and broad applicability of physics-informed deep learning (PIDL) for large-scale lake surface temperature prediction.

Study Configuration

• Spatial Scale: Multi-lake (40 source lakes, 43 additional lakes), regional to global scale (869 lakes).
• Temporal Scale: Continuous prediction of lake surface temperature, relevant for assessing ongoing climate change impacts.

Methodology and Data

• Models used: Physics-informed deep learning (PIDL), Transfer-PIDL (a three-stage training strategy: pre-training, re-training, fine-tuning), purely data-driven deep learning (DL), process-based (PB) models.
• Data sources: Large-scale satellite observations, process-based (PB) model simulations, local measurements, in situ temperature observations.

Main Results

• Transfer-PIDL outperformed local PIDL by 20%–39% in validation root-mean-square-error (RMSE) when sufficiently pre-trained (e.g., on 40 source lakes).
• Transfer-PIDL required less fine-tuning data compared to local PIDL and purely data-driven DL for comparable performance across initial cases.
• Across 43 additional lakes, Transfer-PIDL achieved a mean validation RMSE of 1.2 °C, surpassing local PIDL (1.6 °C), DL (1.8 °C), and PB (1.9 °C) models.
• In a global experiment involving 869 lakes, Transfer-PIDL demonstrated cross-thermal-system transferability, with the poorest-performing scenario yielding an RMSE of 1.5 ± 0.36 °C, a mean absolute error of 1.1 ± 0.25 °C, and an R² of 0.85 ± 0.12 (mean ± standard deviation).

Contributions

• Introduces Transfer-PIDL, a novel framework that integrates transfer learning with physics-informed deep learning to overcome the site-specificity challenge of PIDL.
• Significantly enhances the generalizability and accuracy of lake surface temperature modeling at large scales.
• Demonstrates reduced data requirements for fine-tuning compared to existing deep learning approaches.
• Offers a promising and robust approach for large-scale lake temperature modeling, crucial for sustainable water management under climate change.

Funding

Not explicitly mentioned in the abstract.

Citation

@article{Liu2026Predicting,
  author = {Liu, Muyuan and Woolway, R. Iestyn and Xu, Chunqiang and Tong, Yan and Wang, Weijia and Shi, Haoran and Alsulaiman, N. and Tlhomole, James and Ladwig, Robert and Piggott, Matthew D.},
  title = {Predicting Lake Surface Water Temperature With Transfer‐Based Physics‐Informed Deep Learning},
  journal = {Water Resources Research},
  year = {2026},
  doi = {10.1029/2025wr041062},
  url = {https://doi.org/10.1029/2025wr041062}
}