MedicinalPlant: A Multi Class-Balanced Image Dataset for High-Fidelity Diagnostic Modeling of South Asian Medicinal Plants

This data article introduces a high-fidelity, expert-validated image dataset of medicinal plant leaf pathologies, architected to advance research in automated diagnostics and precision herbal agriculture. The repository comprises 22,000 images, representing 22 distinct health classes across seven therapeutically vital species: Aloe vera, Azadirachta indica, Centella asiatica, Hibiscus rosa-sinensis, Kalanchoe pinnata, Mikania micrantha, and Piper betle. Field acquisition was conducted between September 2025 and February 2026 across four diverse agro-ecological zones in Bangladesh (Rajbari, Ashulia, Mirpur, and Hajigonj) using a multi-device strategy with iPhone 16 Pro Max, OnePlus Nord CE 4 Lite, and OnePlus 7T sensors to incorporate significant optical and hardware variability. To ensure biological authenticity, a strict non-redundancy protocol was enforced, where each of the 1,981 original samples represents a unique leaf from a different individual plant. The diagnostic integrity was established through a two-phase audit by a senior agronomist, verifying taxonomic identity and pathological markers. The computational pipeline involved: (1) high-precision background isolation using the ISNet (U2-Net) architecture, (2) standardization to a resolution of 1440 × 1080 pixels in PNG format for pathological analysis, and (3) global min-max normalization to a [0, 1] range. To mitigate class imbalance and enhance model robustness, a stochastic enrichment pipeline using Albumentations (v2.0.8) was implemented. This included geometric transformations (random rotations ±30°, horizontal/vertical flips) and photometric adjustments (random brightness/contrast ±0.2, Gaussian noise, and sharpening filters) to generate 20,019 synthetic variants. All final outputs were resized to a training-optimized 512 × 512 resolution. Accompanied by a comprehensive metadata file containing geodetic and climatic logs (31.5°C, 78% RH), this dataset provides a robust foundation for developing device-agnostic deep learning models, Vision Transformers (ViTs), and explainable AI systems for botanical protection.