Explainable and Energy-Efficient Lightweight Deep Learning on Embedded Edge-Vision Hardware for Intelligent Medicinal Plant Recognition and Herbal Health Applications

Abstract

Accurate identification of medicinal plants is not only the basis of Ayurveda but also of biodiversity conservation and rural healthcare, and it has been hampered by the scarcity of experts, morphological similarity between species, and the reliance of most deep-learning approaches on cloud connectivity and black-box predictions. This paper introduces an energy-efficient, edge-vision system that combines a light-weight deep neural network with embedded electronic sensing to recognise medicinal plants and perform herbal-health mapping on the edge, with the main goal of maintaining the energy efficiency and interpretability of the system. The backbone networks used include MobileNetV3-Large, EfficientNetV2-Small, DenseNet121 and a Tiny Vision Transformer (ViT-Ti), which are trained using the same transfer-learning protocol and evaluated based on accuracy, precision, recall and F1-score. The models are down-scaled in three steps — magnitude pruning, INT8 post-training quantisation and knowledge distillation — which allows them to be deployed on battery-powered edge devices, and an additional dual explainability layer (Grad-CAM and SHAP) ensures that predictions are based on biologically relevant parts of the leaves. The computational-complexity, latency and per-inference energy trade-off is revealed on an embedded system-on-chip. When tested on a 10-class dataset reported for medicinal plants, the compressed models achieve a top-1 accuracy of 96–99% with a significantly smaller memory requirement, typically 3.9× smaller, and a per-inference energy cost of 46 mJ, making the system suitable for sub-2 W offline operation. The framework clearly outlines contributions across four areas (AI, electrical/electronic sensing, agriculture and medicinal-plant health) and establishes a deployable blueprint for trusted decision support at the edge for herbal health.