Abstract
River discharge is a fundamental variable for water resources management and flood risk assessment, yet conventional estimation methods rely heavily on extensive field bathymetric surveys and site-specific calibration, severely limiting their applicability in data-scarce small and medium-size rivers. While non-contact surface velocimetry has advanced rapidly in recent years, converting surface velocity to discharge remains bottlenecked by the requirement for cross-sectional geometry data. This study proposes a bathymetry-free CFD-driven deep learning framework that inverts river discharge using only surface-velocity profiles and channel width. A multi-scale synthetic dataset covering broad ranges of channel width, water depth, and cross-sectional mean velocity (0.1~3.0 m/s) was generated via OpenFOAM simulations under fully developed subcritical flow conditions. Three targeted strategies are adopted to improve model performance and transferability: gradient-based hydraulic feature enhancement to strengthen depth inversion accuracy, a half-profile adaptation scheme to accommodate asymmetric natural channel profiles, and automatic cross-section screening to delineate the method’s applicability domain. The framework achieved R2 values above 0.9 for both depth and mean velocity in synthetic fixed- and variable-width rectangular channels. Validation on the Swindale Beck natural river reach yielded an R2 of 0.842 for flow depth and 0.790 for mean velocity predictions. These results indicate that reliable discharge estimation is feasible using solely surface observations, with no requirement for in-situ bathymetry during deployment. The proposed framework offers a practical and transferable monitoring solution for straight, quasi-uniform subcritical small river reaches, and future work will extend on larger river systems and more complex channel geometries.

This publication has 38 references indexed in Scilit: