Self-Supervised Plankton Classification via DINO and Gradient-Based Loss Re-weighting

Manual identification of plankton species is labor intensive and impractical for large-scale ecological analysis. Deep learning offers a promising solution but suffers when trained on small and imbalanced datasets, common in microscopic image domains. In this paper, we present a comparative evaluation of several convolutional neural networks (CNNs), including VGG16 and DenseNet201, and propose a novel approach integrating a DINO vision transformer for feature extraction. Additionally, we introduce a gradient-based loss weighting mechanism to estimate the effective number of samples per class, improving training stability under imbalance. Experimental results on a 19-class plankton dataset show that our DINO-based model achieves a 97% test accuracy, significantly outperforming all CNN baselines. These findings highlight the effectiveness of self- supervised transformers and tailored loss strategies for robust plankton classification.