Seminar: Ladislav Hodac

Phytoplankton are key bioindicators of freshwater health, but manual species identification is time-consuming and requires expert knowledge. While deep learning (DL) offers faster, scalable solutions, it faces challenges from imaging biases—such as differences between training and test datasets—and the morphological diversity of phytoplankton species. We trained a DL model (CNN) on standardized images of laboratory cultures (algal strains) and evaluated its performance on independent environmental datasets. The model reliably classified over half of the species across these datasets, particularly those with distinctive morphologies. By analyzing the model’s extracted image representations (DL features), we confirmed that feature and domain shifts between datasets reduce classification accuracy. Comparing DL features with species shape analysis revealed that DL features effectively captured species similarity patterns, aligning well with shape-based inference. Our findings demonstrate that combining DL with expert taxonomy can enhance automated, interpretable phytoplankton monitoring systems.