Vision Transformer with Self-Attention for Automated Plant Seedling Classification

Plant leaves are considered as valuable sources of information for accurate plant species identification. It is having significant impact in various fields such as pharmacology, ecology, medicine, and agriculture. Traditional methods of classifying plant species needed human field experts to visit the physical fields, which is a complex and extended task that often compromises the accuracy of manual identification. In recent years, artificial intelligence enabled computer vision techniques have fundamentally transformed how digital images are understood, categorized, and identified. This paper explores the application of Vision Transformer (ViT), a recent advancement in artificial intelligence, for plant seedlings classification. With its impressive performance against CNN, Vision Transformer (ViT) has become a competitive substitute for CNN in image categorization tasks. Vision transformer offers automatic feature extraction and thus eliminating the need for manual work that reducing the possibilities of errors in identification. The study compares ViT with the conventional state-of-the-art convolutional neural networks (CNNs) such as VGG16, MobileNetV2, and DenseNet169. The dataset used contains 960 plants categorized into 12 classes at different growth stages. The results exhibit that the proposed ViT model outperforms other CNN based models with an accuracy of 95.20%, precision of 95.05% and recall of 94.98%.