Deep Learning Models for Phytoplankton Detection for Water Quality Assessment

Abstract (EN)

In Thailand, prevalent euglenoid genera like Euglena, Phacus, Trachelomonas, Lepocinclis, and Strombomonas are typically found in organic-rich water, indicating hypereutrophic and eutrophic conditions. These euglenoids have diverse shapes, making accurate identification challenging. To address this, an object detection application was developed using deep-learning neural network models to reduce identification errors. Photographic datasets collected between June and October 2022, using both microscope and phone cameras in Khlong Mahasawat, Nakhon Pathom, Thailand, covered five Euglenoid genera. These datasets were manually labeled and used to train four deep-learning neural network models: Detectron2, YOLOv5, YOLOv7, and YOLOv8. Precision and recall of the models were improved through image augmentation, mimicking variations in image quality. The model which proved best for phytoplankton identification was YOLOv5l, which yielded precision and recall of 0.839 and 0.873, respectively. This model exhibited high performance in terms of the accuracy, with a low rate of misclassification.