2022 Volume 78 Issue 2 Pages I_127-I_132
Understanding the number of juvenile ayu running upstream is crucial for ayu stock assessment. However, counting is currently done by visual inspection of underwater camera images, which is a time-consuming task. In this study, we used images from underwater cameras installed in a fishway at Kamogoshi weir on the Yoshii River in Okayama Prefecture to train a model for detecting juvenile ayu running upstream using YOLOv5 (i.e., a deep learning algorithm for object detection). The model's ayu detection results showed that accuracy improved as the amount of ayu data increased. When labeling an object that is easily misidentified as ayu (i.e., Clithon retropictum), accuracy decreased using small amount of data. Using our model, the number of images requiring visual counting was reduced to less than 1 % of the conventional visual confirmation. Furthermore, using our model in a two-month ayu stock survey is expected to reduce the required time by about 90 hours when compared to the conventional method.
PROCEEDINGS OF HYDRAULIC ENGINEERING
