Abstract: Reliable plant species identification is essential for biodiversity conservation, agriculture, and ecological research. However, current plant species recognition systems often struggle with the rejection of unknown classes, which limits their applicability in real-world scenarios. Typically, the maximum probability score is used to reject unknown classes, relying solely on the highest output while neglecting the significance of other output scores, which may restrict the model's potential. In this research, we propose a novel scoring function named the Top-K Logit Disparity Score (TKLDS) for open-set plant species recognition using a Vision Transformer (ViT) network. We conducted extensive experiments on the VNPLANT200 dataset consisting of 200 plant species, where the ViT-L/16 model achieved the highest accuracy in closed-set recognition and the highest Area Under the Receiver Operating Characteristic curve (AUROC) between known and unknown classes compared to other state-of-the-art models, such as ResNet, ConvNeXt, Swin Transformer, and MaxViT. Our results indicate that tuning the parameter k in TKLDS consistently improved the arithmetic mean of closed-set accuracy and AUROC across all pre-trained models. Notably, larger values of k generally led to better performance, with the ViT-L/16 model yielding an arithmetic mean score of 0.975 ± 0.005 for k = 4 with 5 combinations. These findings demonstrate the potential of TKLDS as a robust scoring function for open-set recognition tasks, highlighting its effectiveness in improving performance metrics in plant species identification.