Computer Vision-Assisted Semiautomatic Analysis of Zooplankton in a Longitudinal Study of the Ecological State of Lake Baikal

Studying zooplankton in freshwater ecosystems is crucial for ecological research, providing insight into ecosystem health, biodiversity, and water quality. This study focuses on developing a hybrid approach for studying and analyzing zooplankton communities using machine learning and human expert analysis. The goal of the study was to automate the labor-intensive process of zooplankton analysis as part of a long-term Lake Baikal monitoring program (since 1945), while maintaining continuity with traditional methods. A software and algorithmic system were developed to automate the analysis: images were processed using a two-stage pipeline (object detection using YOLO V11, classification using metric learning and visual transformers), and complex cases and new species were sent to specialists for verification. Over 240,000 images from 811 samples were processed, and models are updated using verified data to adapt to seasonal changes. An updatable database of labeled zooplankton images suitable for statistical analysis and research has been created. A comparison of manual and machine analysis revealed no significant differences in species composition, with accurate detection in 87% of images. This approach allows for scalable monitoring and the accumulation of labeled data arrays for the development of computer vision methods and the assessment of the state of Lake Baikal’s ecosystem.