A Two-Stage Space-Time image Velocimetry method based on deep learning

Abstract

Accurate and robust river flow measurements are essential under complex environmental conditions. In this study, an improvement of Deep Learning-based STIV combined with scene classification is proposed. We build datasets from real rivers and labels the Main Orientation of Texture (MOT) using scene classification and semi-automatic labeling. The STI classification model, built with EfficientNetV2 as the backbone, divides STIs into three classes, achieving an accuracy of over 97.6 % on the validation set and 91 % in generalization experiments. For detecting valid STIs, the MOT regression model employs Group Convolution and Convolutional Block Attention Module (CBAM), with a MAE of 0.49° on the validation set. The velocities corresponding to uncertain, invalid and blind areas are corrected utilizing the distribution law of section velocity. The proposed method achieves a MRE of 3.90 % in general environments and 9.48 % in extreme environments, outperforming both Gradient Tensor and Fast Fourier Transform methods.