A Data-Efficient Image Transformer for Multi-Crop Plant Disease Detection

Accurate plant disease diagnosis is essential for improving agricultural productivity and supporting global food security. Conventional convolutional neural networks may struggle to capture long-range spatial dependencies and subtle visual differences among plant diseases. This study investigates the use of a Data-Efficient Image Transformer for automated recognition of plant leaf diseases across multiple crops. The proposed framework employs transfer learning with pretrained weights, mixed-precision training, and classifier-head fine-tuning to enhance learning efficiency and predictive accuracy. Experiments were conducted on a publicly available dataset containing 30,945 images from eight crop species and 35 disease categories. Performance was evaluated using overall accuracy, macro-averaged precision, recall, and F1-score, together with confidence estimation and disease severity assessment. The model achieved an overall accuracy of 98.50%, with macro-precision, macro-recall, and macro-F1 scores of 0.9843, 0.9826, and 0.9833, respectively. Under identical training settings, DenseNet-121 achieved an accuracy of 97.29%. These findings demonstrate the effectiveness of the proposed framework for multi-crop plant disease classification on the evaluated benchmark dataset.