Plankton Classification with Deep Learning

Plankton plays a vital role in sustaining life on Earth as it forms the foundation of the aquatic food chain. Additionally, it is responsible for producing approximately half of the oxygen in the Earth's atmosphere. Understanding the distribution of plankton is crucial in studying climate change and monitoring water quality. This study utilized a dataset comprising 30,336 plankton images belonging to 121 different classes. State-of-the-art deep learning methods were employed to automatically label the plankton species, while alternative training approaches, data augmentation, and preprocessing techniques were compared to assess their impact. The study analyzed the similarities between and within plankton classes, as well as the classes that were frequently misidentified. By training models on networks from two different versions of the Kaggle dataset, the highest success rate achieved was 78% on the larger dataset with 118 classes, and 92% on the smaller dataset with 38 classes using individual models. These results represent a 1% improvement over the previously reported best-performing single model classifiers. Furthermore, by employing an averaging ensemble method, the performance was further enhanced by 2% in the first dataset and 1% in the second dataset. These findings demonstrate the efficacy of the proposed approach in accurately identifying plankton species and highlight the potential for improving classification results through ensemble techniques. Also, we address the unbalance and the varying input sizes within the Kaggle plankton dataset. We propose alternative data balancing methods and investigate three pre-processing techniques. Then we compare several state-of-the-art deep learning algorithms and show the improvement in classification rates with ensemble methods. Our results have significantly improved the best results reported in the literature over the single best models.