Identification of Cyanobacteria for Harmful Algal Blooms Research Using the YOLO Framework

Cyanobacteria, an ancient type of photosynthetic microbe, inhabit most fresh and marine water on Earth. The rapid growth of cyanobacteria can lead to Harmful Algal Blooms (HABs), posing major threats to water quality and aquatic ecosystems. Rapid and accurate identification of cyanobacteria is essential for population monitoring and mitigation efforts, especially when cyanobacteria produce toxins, threatening the health of wildlife and humans. However, the diverse shapes and appearances of cyanobacteria render manual identification time-consuming and error-prone. In this study, we make multiple novel contributions to the field of microscopic cyanobacterial identification using computer vision algorithms. To begin, we utilize the YOLOv5 algorithm, known for its speed and accuracy, which has never been evaluated for its efficacy in this field. Additionally, we propose numerous methods of addressing limited dataset size and image heterogeneity. We use various image pre-processing techniques, including color-preserving CLAHE. We also construct a comprehensive dataset containing several genera of cyanobacteria by supplementing laboratory images with opensource database images for training and evaluation. To combat overfitting and avoid unrealistic model performance values, we evaluate detection performance on common microscope artifacts (detritus and water bubbles), incorporate “background images”, which contain unrelated microorganisms into the dataset, and utilize image augmentation conservatively. Finally, hyperparameter tuning was used with a genetic algorithm to optimize a specified fitness function. The final model outperformed the Faster R-CNN model used in previous literature, achieving average precision values ranging from 70% to 90% for five commonly found, toxin-producing cyanobacteria taxa in the USA, representing state-of the-art performance and great potential for usage by biologists investigating HABs.