Author(s): Wen-Cheng Liu; Wei-Che Huang
Linked Author(s):
Keywords: Water level measurement; Deep learning; SegNet; YOLO
Abstract: Water levels and discharge are crucial hydrological data for assessing the flow conditions within rivers. Currently, the primary methods for measuring water levels are pressure-based water level gauges and radar wave water level gauges. However, pressure-based gauges are prone to interference from sediment, and radar wave gauges are costly and must be installed on bridges, which limits their use due to topographical constraints. In recent years, with the advancement of deep learning, many studies have used semantic segmentation models (such as FCN, U-Net, and SegNet) or object detection models (such as YOLO and R-CNN) for water level measurement, but none have compared the differences between these two approaches. Therefore, this study aims to compare the differences between SegNet semantic segmentation model and YOLO object detection model in measuring water levels. This study was conducted in a drainage channel in Miaoli, Taiwan. First, a water gauge was drawn on the channel embankment for identification by the object detection model, and a virtual water gauge was established. Then, images of the channel were collected every hour for five days, with eight images collected each day, totaling 40 images as the training set for training and validating the SegNet and YOLO models. The trained SegNet and YOLO models were then applied to water level analysis of the channel for an additional four days. The results showed that the RMSE between the SegNet model and the measured water level ranged from 0.10 m to 0.21 m; the RMSE between the YOLO model and the measured water level ranged from 0.05 m to 0.18 m. This indicates that both models have similar measurement accuracy, but YOLO performed slightly better than SegNet.
Year: 2024