Transfer Learning-Enabled Residual Networks for Fine-Grained Identification of Indian Medicinal Flora Across Leaves, Flowers, and Whole Plants

Abstract

Accurate taxonomic identification of medicinal flora is vital for safeguarding indigenous biodiversity, standardizing Ayurvedic and herbal pharmacopoeia, and intercepting commercial botanical adulteration. Conventional manual identification by taxonomists is unscalable and prone to subjective error due to subtle intra-class morphometric variations and high inter-class phenotypic similarities. In this paper, a multi-organ botanical classification framework based on deep transfer learning is developed and evaluated across three distinct modalities: whole plants, plant leaves, and flowers. A fine-tuned ResNet-18 architecture with tailored dense projection heads and differential layer-wise learning rates was trained on three comprehensive Indian medicinal flora benchmarks comprising 40 whole-plant classes, 80 leaf classes (6,900 images), and 28 flower classes (6,213 images). An online stochastic augmentation pipeline—incorporating planar rotations (), horizontal mirroring, and spatial translations—was deployed to prevent overfitting on background noise and preserve geometric invariance. Trained with the Adam optimizer over 8 epochs using stratified 70/15/15 partitions, the proposed pipeline achieved generalization test accuracies of 98.10% on whole plants, 94.79% on leaves, and 99.79% on flowers. These empirical outcomes confirm that deep residual representations generalise across distinct botanical organs, providing a computationally lightweight foundation for embedded and field-deployable botanical diagnostic tools.