2:15 PM - 2:30 PM
[AOS16-03] Development of a Zooplankton detection model using a shadowgraph camera and deep learning
Keywords:Coastal zooplankton, Artificial Intelligence, Image processing , Object detection
Understanding the spatiotemporal distribution of zooplankton is essential to revealing the mechanisms of variability of the entire marine ecosystem. Much of previous research has been utilizing plankton nets towed by ships. However, field observations using ships are often hindered by bad weather conditions, and expensive operational costs limit the frequency, therefore, conducting comprehensive observations over wide areas is difficult. Consequently, data on the spatiotemporal distribution of zooplankton remains insufficiently developed and accumulated. To address this issue, the development of inexpensive observational tools utilizing a shadowgraph camera has been underway. However, data processing and analysis of zooplankton have been relied on by experts, and even with increased data collection, there will be limitations to the amount of processing. Therefore, in this study, an automatic zooplankton detection model was developed by combining images captured with shadowgraph cameras and deep learning.
To build an automated zooplankton detection model based on deep learning requires preparing a large amount of training data. Thus, field data collection was conducted at four offshore stations in Suruga Bay, two days per month from June to September 2023, to collect the training data. Zooplankton identified in the above observations were labelled to create the training data. Using the annotated data, we developed a detection model for zooplankton using YOLOv8 (https://github.com/ultralytics/ultralytics.git) and verified the performance.
For copepods, which found a large amount individuals in training data, sufficient precision and recall values were obtained. However, for other taxonomic groups, still, prediction accuracies were still low and the insufficiency of training data was suggested. For future development, it is necessary to collect more training datasets for each taxonomic group and then conduct performance verification and evaluation after the data expands.
To build an automated zooplankton detection model based on deep learning requires preparing a large amount of training data. Thus, field data collection was conducted at four offshore stations in Suruga Bay, two days per month from June to September 2023, to collect the training data. Zooplankton identified in the above observations were labelled to create the training data. Using the annotated data, we developed a detection model for zooplankton using YOLOv8 (https://github.com/ultralytics/ultralytics.git) and verified the performance.
For copepods, which found a large amount individuals in training data, sufficient precision and recall values were obtained. However, for other taxonomic groups, still, prediction accuracies were still low and the insufficiency of training data was suggested. For future development, it is necessary to collect more training datasets for each taxonomic group and then conduct performance verification and evaluation after the data expands.