Optical cameras are being used to identify marine species as an easy and cost-effective aquaculture method for fishermen. Deep learning-based image analysis is now widely used in many fields, including aquaculture. In this study, combining an underwater time-lapse camera and deep learning-based image analysis, we proposed a simple monitoring system for cage-cultivated small sea cucumbers on the seafloor. With the camera, we obtained numerous time-lapse images of sea cucumbers for approximately two months. For these numerous images, the application of deep learning methods is beneficial for efficient and rapid analysis. In the monitored images, however, the outline of the sea cucumbers is likely to be blurred for three reasons: their small size, the net in the background, and the relatively low resolution of the camera system. It appears challenging to automatically detect sea cucumbers with blurred outlines. First, 1,429 individual sea cucumber images were manually annotated for training and validation purposes, using the YOLOv5 model. The training model achieved a precision, recall rate, and F-measure of 0.72, 0.71, and 0.72, respectively. The validation model successfully counted sea cucumbers in a cage at an acceptable monitoring level, despite the low image resolution. Second, all collected camera images are analyzed for automated detection using the tailored model. As a result, sea cucumbers mostly gathered around the top corner of the cage, probably because of the low current. The finding is useful for integrated multi-trophic aquaculture systems, especially the design of sea cucumber cages placed on the seafloor.
Read full abstract