2022 Volume 3 Issue J2 Pages 479-487
In this study, we proposed a method for measuring water level by image analysis using the object detection model YOLO. In order to improve the accuracy of the analysis at night, a near-infrared camera was used to capture images. We captured images of a pole in the Seto Inland Sea near Kagawa Prefectural Fisheries Experiment Station for one week from June 8 to 14. The model which detects the pole was created by machine learning in YOLOv5. Train data is 1830 images for the first day after the start of the obserbation, and validation data is other images for 6 days. The change in height of the lower limit of the pole in the images by tide were converted to water level. As a result, the RMSE between the water level measured by YOLO and the tide lebel of Takamatsu port was 9.91 cm during the daytime and 7.14 cm during the nighttime. So the accuracy during the nighttime was higher than that during the daytime.
