Explainable Transfer Deep Learning for Medicinal Leaf Classification, Comparative Baseline Analysis, and Experimental Validation

Abstract

Automated recognition of medicinal leaves has implications for primary healthcare support, botanical informatics and mobile decision systems. However, many published studies focus on model accuracy, with little to no reusable implementation layer, or they compare baselines without support for end-user experimentation. We introduce a research-grade Python implementation for medicinal leaf classification that includes preprocessing, augmentation, transfer learning, explainable artificial intelligence, structured artifact logging, and a step-wise graphical workbench. The implemented system is trained on a curated dataset of 1009 RGB leaf images from nine medicinal plant categories and is organized in class-specific folders. A unified preprocessing module implements resizing, denoising and grayscale conversion for segmentation support, Otsu thresholding, background masking and normalization, while the learning stage uses ImageNet-pretrained ResNet50 and MobileNetV2 backbones. The application makes each experimental stage accessible through dedicated button actions in a Streamlit interface, thus improving reproducibility and usability in comparison to script-only pipelines. Both trained models showed 100% test accuracy, weighted precision, weighted recall and F1-score on the held-out subset of a stratified 70/15/15 split. A sample inference case detected Piper betle with 93.09% confidence and interpretable Grad-CAM localization. For context, the author baseline manuscript reported around 59% for a GLCM + SVM pipeline, and 100% for a prior MobileNetV2 baseline. The results demonstrate that the current implementation is highly effective in the controlled acquisition setting, but also point to the need for larger and more diverse field-validation studies to confirm generalization beyond the current dataset.