Lin et al. (2026) Non-contact discharge estimation using physics-informed neural networks and CNN-enhanced image velocimetry

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-07-13
• Authors: Yen-Cheng Lin, C L Wu, Hao‐Che Ho
• DOI: 10.1016/j.jhydrol.2026.136015

Research Groups

• Department of Civil Engineering, National Taiwan University, Taipei, Taiwan

Short Summary

This study proposes a physics-informed framework for non-contact river discharge estimation by integrating CNN-enhanced Large-Scale Particle Image Velocimetry (CNN-LSPIV) with physics-informed neural networks (PINNs). The framework is validated using laboratory flume experiments under three bed configurations and further evaluated through unmanned aerial vehicle (UAV)-based field observations.

Objective

• Develop a non-contact river discharge estimation method that integrates CNN-enhanced LSPIV with PINN-based physical constraints to improve the physical consistency and relative accuracy of discharge estimation.

Study Configuration

• Spatial Scale: Laboratory flume experiments under three bed configurations (smooth, lateral composite, and longitudinal composite beds) and natural river field observations.
• Temporal Scale: Steady-state conditions for laboratory experiments and varying flow conditions for field observations.

Methodology and Data

• Models used: CNN-enhanced LSPIV and physics-informed neural networks (PINNs).
• Data sources: Laboratory flume experiments, UAV-acquired imagery, and ADCP discharge measurements.

Main Results

• The proposed framework improves the physical consistency and relative accuracy of non-contact discharge estimation compared to conventional LSPIV.
• Two PINN operational modes are designed: Velocity Field Refinement mode (PINN-VR) and Direct Water Depth Inversion mode (PINN-DDI).
• PINN-VR shows superior depth reconstruction under lateral bed configurations, while PINN-DDI performs best under longitudinal bed variations.

Contributions

• This study provides a systematic integration of CNN-enhanced LSPIV and physics-informed neural networks within an image-based discharge estimation framework.
• Two complementary inversion strategies are developed and comparatively evaluated under different bed-geometry and flow-complexity conditions.

Funding

• This research was funded by the Ministry of Science and Technology, Taiwan (MOST 109-2221-E-002-033-MY3).

Citation

@article{Lin2026Noncontact,
  author = {Lin, Yen-Cheng and Wu, C L and Ho, Hao‐Che},
  title = {Non-contact discharge estimation using physics-informed neural networks and CNN-enhanced image velocimetry},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136015},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136015}
}