Plankton plays a crucial role in the marine ecosystem, contributing to the biogeochemical cycle and climate regulation. Traditional plankton monitoring methods struggle with the complex dynamics of ocean ecosystems, leading to a growing interest in computer vision techniques for plankton identification in microscopic images. However, most studies have used small, controlled datasets. Our research is the first to employ large, unbalanced datasets to achieve state-of-the-art plankton image recognition, bridging the gap between laboratory and real-world conditions. We addressed the challenge of training large models in a heterogeneous GPU environment, utilizing advanced models like Swin Transformers and DeiT 3, along with data augmentation and varied loss functions. Our results demonstrate that Vision Transformers (ViTs) excel in plankton image recognition, offering significant potential to improve climate change mitigation strategies.