327 / 2024-02-29 16:07:25

Prediction of river cross section morphology under the water using deep learning

mountain river,cross section morphology,deep learning

River boundaries are crucial for understanding riverine geomorphological evolution, calculating material flux in rivers, and simulating hydrology and sedimentation processes. The simulation of flood processes under extreme rainfall conditions depends on three-dimensional river geometry information. Currently, data on river cross-sectional morphology are primarily obtained through in-situ measurements, including pin, RTK (Real-Time Kinematic), and ADCP (Acoustic Doppler Current Profiler). However, these methods are limited in data-scarce mountainous regions. Cross-sectional morphology above the minimum water level can be acquired through existing high-precision DEM databases, UAV remote sensing, and 3D laser scanning. Nonetheless, measuring cross-sectional morphology below the minimum water level presents challenges.

This study focuses on the mountain rivers in the Yellow River Basin. It uses more than 500 in-situ measured cross sections within and outside the Yellow River Basin as the training set. High-precision DEMs above the water surface were utilized to apply deep learning technology for generalizing underwater cross-sectional morphology. Mountain rivers in the source region and the Loess Plateau region of the Yellow River Basin served as the validation set. Employing Google Earth imagery, river reaches were classified into single-thread and multi-thread patterns. The results show that the R² and the root mean square errors (RMSE) of the two river patterns are 0.67 and 0.87, respectively. A three-dimensional river network of the upper Yellow River was constructed, with future work to analyze its spatial variation and topological structural characteristics. The results will extend the in-situ measured cross-sectional morphology to the entire river network and provide boundary conditions for hydrological modeling.