A Comparative Study of DETR and YOLOv8n for Plankton Detection and Classification in the Makassar River

This study compares two advanced deep learning models that are currently considered the best in their field, YOLOv8n and DETR, for the automated detection and classification of phytoplankton species from images collected in the Makassar River. Plankton are crucial bioindicators of environmental health, but their traditional identification through manual microscopic analysis is a time consuming and error prone process. This research addresses this bottleneck by leveraging deep learning. The methodology involved preprocessing and augmenting an initial dataset of 78 raw images from the Makassar River to create a final dataset of 333 images with annotations for 20 representative species. The models were evaluated on key performance metrics: detection accuracy, inference speed, and performance with small and dense objects. The results show a clear trade off between the models. The DETR model achieved a higher average accuracy of 84.5% compared to YOLOv8n's 81.7%, demonstrating superior performance in handling complex morphological features. However, YOLOv8n was significantly faster, with an inference speed of approximately 180 ms versus DETR's 450 ms. In conclusion, the choice between these models depends on the application's priorities. DETR is more suitable for tasks requiring high precision and accuracy, while YOLOv8n is the preferred choice for real time monitoring where speed is the primary constraint. This research contributes to the development of scalable technology for aquatic ecosystem monitoring and provides a foundation for future studies on hybrid or ensemble model approaches.