Plankton is a vital component of marine ecosystems, integral to the biogeochemical cycles and climate regulation processes. Their abundance fluctuations serve as key indicators of ocean health. Accurate identification of plankton is therefore essential. Traditional monitoring methods, however, struggle with the complex spatial and temporal dynamics of ocean environments. In response, modern computer vision techniques have increasingly been applied to enhance plankton identification in microscopic images. While existing studies have predominantly used small, controlled laboratory datasets, our research addresses this gap by employing large, unbalanced datasets, thereby aligning more closely with real-world conditions. We utilize advanced models, including Swin Transformers and DeiT 3, incorporating data augmentation and diverse loss functions to boost classification accuracy. Additionally, we perform a systematic comparison of six GradCAM algorithms to gain explainability insights into the recognition process. Our results indicate that Vision Transformers (ViTs) outperform traditional methods, showcasing their potential to revolutionize plankton image recognition.