DEEP LEARNING-BASED MULTICLASS CLASSIFICATION OF MEDICINAL PLANT LEAF DISEASES USING EFFICIENTNET-B3, CONVNEXT-TINY, SWIN TRANSFORMER AND A HYBRID ENSEMBLE ON THE CIMD DATASET

Plant diseases threaten the pharmacological quality of medicinal herbs cultivated across Central India and remain difficult to diagnose accurately without laboratory analysis. The Central India Medicinal Plant Dataset (CIMPD), released by Tomar et al. in Data in Brief (Elsevier, 2025), contains 9,130 high-resolution leaf images spanning 23 Indian medicinal plant species under both healthy and diseased conditions. No comprehensive multi-architecture deep learning benchmark has been published on this dataset beyond the ResNet18 baseline in the original paper. This work presents the first end-to-end transfer learning, ensemble, and GradCAM++ interpretability pipeline on CIMPD. Six architectures are benchmarked: a Simple CNN trained from scratch, ResNet50, EfficientNet-B3, ConvNeXt-Tiny, Swin Transformer-Tiny, and a soft-voting Hybrid Model that combines the three strongest backbones using validation-accuracy-proportional weights. Training employs MixUp augmentation (α=0.4), CoarseDropout regularisation, label smoothing (ε=0.1), AdamW optimisation, and cosine annealing with warm restarts. The proposed Hybrid Model achieves 90% test accuracy with weighted F1=0.885, Cohen's κ and MCC substantially above chance, and AUC=1.00 on all four reported class categories, outperforming the Simple CNN baseline by 15 percentage points and the prior ResNet18 result by approximately 7 points. GradCAM++ saliency maps confirm that model attention concentrates on biologically meaningful leaf regions. All experiments are reproducible through a single self-contained Jupyter notebook.