A Data Generation Method for Measuring the Surface Flow Velocity of Rivers via Deep Neural Networks

A crucial method in hydrological monitoring is radar-based flow measurement, which uses Doppler spectra of water surface echoes to extract surface flow velocity accurately. Deep neural network training is hampered by the fact that radar signals are frequently weakened by clutter and interference in real-world settings, resulting in low echo quality, large Doppler shift estimate errors, and a lack of useful data. A deep learningbased approach to producing river surface flow velocity data is suggested as a solution to this problem. Physically consistent and realistic Doppler training datasets are generated by simulating high-frequency signal transmission, reception, and echo processing based on a millimeter-wave radar model. The inputoutput feature mapping is established using a neural network framework based on a Multi-Layer Perceptron (MLP), which incorporates nonlinear activation functions to enhance model stability and convergence. Furthermore, a Multi-Scale Bitemporal Fusion Module (BFM) is presented to improve feature integration by means of channel attention mechanisms, spatial attention, and multi-scale feature extraction. According to experimental data, the synthetic Doppler spectra closely match actual measurements; over 95 % of samples have a correlation coefficient more than 0.95, and the average is 0.9784. The model performs consistently across a range of flow velocities, especially at high velocities, and reconstructs spectral structures with accuracy. High-precision modeling of the intricate nonlinear relationship between Doppler spectra and river surface velocities is made possible by this method, which also successfully alleviates the lack of real flow velocity data and facilitates the use of deep learning in hydrological monitoring.