Benchmarking CNN and Vision Transformer Architectures for Corn Leaf Disease Classification on the Kaggle Maize Dataset

Abstract

Foliar diseases in corn pose a critical constraint on agricultural productivity, particularly in developing countries. Deep learning-based automated detection has emerged as a viable alternative to conventional manual inspection. This study presents a comparative evaluation of four contemporary deep learning architectures—EfficientNet-B3, MobileNetV3-Large, ResNet50, and Vision Transformer Small (ViT-Small)—on the publicly available Corn or Maize Leaf Disease Dataset hosted on Kaggle (4,188 image samples; four classes: Blight, Common Rust, Gray Leaf Spot, and Healthy). Class imbalance was addressed through a combination of WeightedRandomSampler and Focal Loss, while all architectures were trained via transfer learning from ImageNet pretrained weights, augmented with MixUp and CutMix. Experimental results demonstrate that ViT-Small achieved the highest classification performance, attaining 97.14% accuracy, a weighted F1-Score of 0.9716, and an AUC-ROC of 0.9961, outperforming EfficientNet-B3 (96.66%), MobileNetV3-Large (96.18%), and ResNet50 (95.71%). As an external reference, these results are also compared indicatively with the DenseNet121 accuracy (93.48%) reported by Waheed et al. (2020); it must be emphasized that this baseline was not reproduced in the present experiments, and therefore the comparison should be interpreted as indicative rather than conclusive. McNemar’s test confirmed that ViT-Small’s superiority is statistically significant (p<0.05). An ablation study verified the positive contribution of the Focal Loss and WeightedRandomSampler combination. Grad-CAM visualization corroborated that all models direct their attention to pathologically relevant lesion regions.