They automatically identify medicinal plants in the field when faced with visually similar species and the additional constraints of privacy and limited, intermittent connectivity. In this paper, the authors present an EDGE-FOG-CLOUD architecture that uses RGB images of leaves, AS7343 optical measurements, and BME688 volatile organic compound (VOC) responses to provide a three-tier system for automated medicinal-plant identification. The authors use an independent classifier approach followed by a late fusion approach with validation-selected confidence gating to ensure that only uncertain and privacy-filtered samples are forwarded to the cloud/central server for additional review and validation. The architecture uses ECDSA with SHA-256 for authenticity and integrity, TLS 1.3 for transport protection, and AES-256-GCM encryption for signed model-update packages. Component-level benchmark results indicate that MobileNetV3-Small achieved 98.54% accuracy and 98.55% macro-F1 on a public 30-class medicinal leaf dataset using a 1.90MB float16 TFLite model, while Random Forest achieved 93.06% accuracy on CoffeePow-4 and 94.90% accuracy on Aroma-7 using public BME688 cycle data. This suggests that lightweight branch functionality can be achieved, though not necessarily full multimodal functionality. The authors conclude by defining a leakage-controlled protocol for five species of Vietnamese medicinal plants and discuss how the system was evaluated through ablation studies and a series of controlled-environment studies to assess the effectiveness and appropriate implementation of calibration, adaptive offloading, and security controls.

AIoT, edge–cloud inference, medicinal plants, multimodal fusion, privacy, sensor security