Flow measurement in open channels using imaging techniques in conjunction with a convolutional neural network

Abstract

Data related to river velocity and discharge are important for water resource management. Non-intrusive image measurement techniques based on direct cross-correlation (DCC) algorithms, such as particle image velocimetry (PIV), are widely used to measure velocity and discharge in the field. However, environmental noise is highly complex and uncontrollable in the field, significantly reducing the accuracy of DCC-based methods. Convolutional neural networks (CNNs) are commonly used in image recognition because of their outstanding accuracy, which far exceeds that of conventional imaging methods. However, these accuracy levels cannot be directly extrapolated for flow movement estimation. Therefore, in this study, we developed an innovative sub-pixel correction technique that allows CNN-based methods to obtain stable measurements in the PIV framework. This is the first study that successfully applied the concept of CNN to measure velocity data using images. The Hamel-Oseen vortex-flow, uniform steady-flow, and plane laminar jet-flow models are established as benchmark vector fields. Non-uniform illumination and Gaussian noise with varying degrees of interference are added to the synthetic data to evaluate the performance of the CNN-based method in PIV. For noiseless images, the DCC-based and CNN-based methods achieve lower measurement errors than benchmark errors. For noisy images, the DCC suffers a fold error increase ranging from 2.77 to 31.13, whereas the CNN suffers only a fold error increase ranging from 1.25 to 1.68. A 30-m-long flume is then used in an uncontrolled environment to mimic real-world flow measurement. The dispersion of the instantaneous velocity measurements for the CNN is more concentrated than that for the DCC. The acoustic Doppler velocimetry yields an error of only 7.87% in discharge estimation using CNN. These results indicate that the CNN-based method is more robust than conventional methods and has the potential to be effectively applied to measurements in the field.