[A-OS21] Coastal ocean circulation and material cycle

[AOS21-P10] Training data improvement for building zooplankton detection models

*Tomoki Myoken¹, Mehul Sangekar², Akiyuki Kenmochi³, Jun Nishikawa³, Dhugal Lindsay², Mitsuko Hidaka^1,2 (1.Kagoshima University , 2.Japan Marine Science and Technology Center, 3.Tokai University )

Keywords:Coastal zooplankton, deep learning, Image processing, Object detection

Understanding the distribution of zooplankton is essential to elucidate the mechanisms underlying changes in marine ecosystems. Traditional zooplankton surveys principally use planktonnet-based methods, but these methods depend on vessels, which results in low spatiotemporal resolution of the data. As a result, there is a lack of observation data to understand zooplankton communities across large areas. Image-based survey methods have the potential to be applied to existing wide-area observation networks, such as float technology or moored buoys. Therefore, various plankton sensing devices have recently been developed utilizing various optical technologies. However, for these observation techniques to become widely used, it is crucial to establish consistent methods for analyzing large volumes of plankton images. In the present study, zooplankton detection models have been developed using images captured by a shadowgraph camera as a case study to consider technical issues. The shadowgraph images were obtained through observation in Suruga Bay in Japan; the four locations from the mouth of the Fuji River to offshore were selected. The observations were conducted twice a month from June to September 2023 using a shadowgraph camera that was uniquely developed. Zooplankton in the frames were annotated, and 4949 labels were created for the model training. Using the dataset, we developed a zooplankton detection model using YOLOv8 (https://github.com/ultralytics/ultralytics.git). A total of 28 zooplankton groups were annotated, but to prevent overfitting, only the taxonomic groups with more than 50 labels given (seven groups) were used to build the model. As a result, in the three groups with a large amount of training data (around 1000 annotations)—copepods, chaetognaths, and radiolarian—the evaluation metrics, precision and recall both exceeded 70%. In contrast, sufficient precision and recall could not be achieved for the four groups with fewer labels. Therefore, data augmentation was applied to address the data shortage, and a new model was built and re-evaluated. We will present the results of the model improvement through data augmentation and an assessment of the practicality.