PlanktonFlow : hands-on deep-learning classification of plankton images for biologists

High-resolution biodiversity monitoring is crucial to understand the complex dynamics of ecosystems. High throughput image-acquisition devices now provide biologists with a wealth of data, but mainstream, automated image-classification workflows are still lagging behind, while high-performance classification models remain hardly manageable by most biologists.

To fill this gap, we developed PlanktonFlow, an end-to-end, fully automated deep-learning pipeline for plankton imagery intended for biologists having minimal technical expertise with machine learning. PlanktonFlow automates image pre-processing, model training and calculation of performance metrics for four different convolution neural network architectures (ResNet, DenseNet, EfficientNet, and YOLOv11), as well as inference on novel datasets. We highlight the potentialities of PlanktonFlow by comparing the relative performances of the four models and of the web-service EcoTaxa on classifying images from freshwater samples including 76 classes of both living and non-living objects.

EfficientNet-B5 achieved the highest macro-averaged F1 score, outperforming Eco-Taxa as well as other models. Despite class imbalance and morphological complexity, all models generalized well across frequent and rare taxa, highlighting the relevancy of the pre-processing routine. Error analyses stressed that classification challenges arose from high intra-class variability and high inter-class similarity.

To ease an appropriation and further developments by the community, PlanktonFlow is open source and has a modular structure. We foresee that future work could leverage semi-supervised learning, test the pipeline on other imaging systems, and expand its applicability to other biological systems.