Recognition of Harmful Phytoplankton from Microscopic Images using Deep Learning

TL;DR

This work tackles automated recognition of harmful phytoplankton from microscopic images by evaluating multiple CNN backbones (ResNet, ResNeXt, DenseNet, EfficientNet) under three transfer-learning strategies (linear probing, fine-tuning, and a combined approach). The best result is achieved with ResNet-50 using fine-tuning, reaching about 96.97% test accuracy across 11 genera, though discrimination among four morphologically similar genera remains challenging. Using a public, 1650-image dataset, the study demonstrates strong per-class performance overall while revealing notable confusions between Aphanizomenon, Nodularia, Oscillatoria, and Anabaena. The findings support automated, scalable HAB monitoring and point to dataset expansion as a key path toward more generalizable water-quality assessment tools.

Abstract

Monitoring plankton distribution, particularly harmful phytoplankton, is vital for preserving aquatic ecosystems, regulating the global climate, and ensuring environmental protection. Traditional methods for monitoring are often time-consuming, expensive, error-prone, and unsuitable for large-scale applications, highlighting the need for accurate and efficient automated systems. In this study, we evaluate several state-of-the-art CNN models, including ResNet, ResNeXt, DenseNet, and EfficientNet, using three transfer learning approaches: linear probing, fine-tuning, and a combined approach, to classify eleven harmful phytoplankton genera from microscopic images. The best performance was achieved by ResNet-50 using the fine-tuning approach, with an accuracy of 96.97%. The results also revealed that the models struggled to differentiate between four harmful phytoplankton types with similar morphological features.

Figures (5)

• Figure 1: Visualization of two samples per class from the toxic phytoplankton dataset.
• Figure 2: Overview of the three learning approaches employed for developing the automatic phytoplankton recognition system: (a) Learning approach using linear probing, (b) Learning approach utilizing fine-tuning, and (c) Learning approach that integrates both linear probing and fine-tuning.
• Figure 3: ResNet-50 accuracy on validation set across various batch sizes using a learning strategy that combines linear probing and fine-tuning.
• Figure 4: Evaluation of ResNet-50 on test data for different types of toxic phytoplankton, trained using fine-tuning.
• Figure 5: Confusion matrix displaying true and predicted phytoplankton class distributions on test data, generated by ResNet-50 trained with fine-tuning.