A lightweight hybrid CNN and transformer model for medicinal leaf disease classification with explainable AI.

Abstract

Medicinal plants including Ocimum tenuiflorum L. (Tulsi), Azadirachta indica A. Juss. (Neem), and Kalanchoe pinnata (Lam.) Pers. (Patharkuchi) are essential sources of bioactive compounds, yet leaf diseases threaten their yield and phytochemical integrity. This study proposes LSeTNet, a lightweight hybrid CNN (Convolutional Neural Network) Transformer architecture with Squeeze-and-Excitation (SE) blocks, achieving 99.72% accuracy, 1.00 macro F1-score, and AUC = 1.00 across 12 disease classes (1,000 images/class post-augmentation) using only 9.38 M parameters and 2.50 GFLOPs. Five-fold cross-validation yielded 99.74% ± 0.14% accuracy, with rapid convergence and no overfitting. Explainable Artificial Intelligence (XAI) via Gradient-weighted Class Activation Mapping (Grad-CAM) (mean intensity: 0.1664-0.2702), Local Interpretable Model-agnostic Explanations (LIME), and t-distributed Stochastic Neighbor Embedding (t-SNE) (silhouette score: 0.87) confirmed biologically meaningful attention on pathological regions. External validation on the independent BD-MediLeaves dataset (8 classes, 8,000 samples) achieved 99.42% accuracy and 0.99 macro F1. With 6.98 ms/image inference latency and 35.81 MB memory, LSeTNet enables real-time, edge-based deployment. It significantly outperforms DenseNet169 (95.56%), ViT-B16 (95.61%), and LW-CNN+SE (95.39%) ([Formula: see text], paired t-tests), establishing a transparent, efficient, and generalizable benchmark for precision phytopathology and sustainable medicinal plant cultivation.

Plant phenotyping relevance

植物葉の病害状態を画像から分類するCNN・Transformer手法を開発し、交差検証、外部データセット、既存モデルとの比較で検証しており、植物フェノタイピング手法が中心である。

Code and data availability

The paper publicly releases its primary medicinal leaf image dataset (MedicinalLeaf-12) on Mendeley Data, uses a public external validation dataset (BD-MediLeaves, also on Mendeley), and provides full training/evaluation code for LSeTNet on GitHub. All three are paper-specific, public, and actionable.