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

High throughput image-acquisition devices tremendously increase our capacity to observe biodiversity. However, for many biologists, the high-performance deep learning models that are needed to make biological sense out of very-large image sets remain difficult to implement. To fill this gap in biologists' toolkit, we developed PlanktonFlow, a Python pipeline that streamlines the automation of plankton-image taxonomic assignment. PlanktonFlow makes it easy for inexperienced users to run a whole sequence of (i) automated image pre-processing and augmentation of rare classes, (ii) training up to four different high-performance convolution neural networks (CNNs: ResNet, DenseNet, EfficientNet, and YOLO), (iii) computing model classification-performance metrics so as to choose the best-performing model, and (iv) running inference on novel image sets. PlanktonFlow further includes routines to easily fine tune model hyper-parameters and optimize model's performances. Using a tutorial style, we demonstrate the usage of PlanktonFlow to analyse freshwater-plankton images produced with the FlowCAM, comparing the relative classification performances of the four optimized CNN architectures. For a baseline comparison with a reference tool used by plankton biologists, we further assessed the classification performances of the EcoTaxa web-service when used without any eye validation in a pure-prediction mode. In line with a previous study on a benchmark plankton dataset, we found that EfficientNet-B5 achieved the highest macro-averaged F1 Score, outperforming other CNN models, which all surpassed EcoTaxa. Hyper-parameter optimization was key to improving model performances. To ease an appropriation and further developments by the community, PlanktonFlow is open source, comes with a detailed documentation, and has a modular structure. We foresee that future work could integrate new deep-learning architectures (e.g., vision transformers, semi-supervised learning), and test the pipeline on images produced by other devices or from other taxonomic groups.