Smart Agriculture System for Real-time Plant Disease Detection Using Transfer Learning and Uncertainty-aware Deep Learning

Viswanatha V., Ramachandra A. C., Harshavardhan B. M., L. Tejas — Tue, 26 May 2026 00:00:00 +0000

Plant diseases are among the most persistent threats to agricultural productivity, responsible for an estimated 20 to 40 percent of global crop losses every year. In most farming communities, especially small-scale and rural ones, disease identification still depends on manual inspection by trained agronomists, a process that is slow, costly, and simply unavailable to the majority of farmers who need it most. By the time visible symptoms are identified and a diagnosis is made, infections have often already spread across a significant portion of the crop. This delay between onset and detection is where the largest share of yield loss occurs, making early and accurate identification not just useful, but critical. The system classifies 38 diseases and healthy states across 14 crop species from live webcam footage or uploaded leaf images, filters out non-leaf and ambiguous inputs automatically, and communicates results both through a browser-based web interface and through a physical LED indicator connected via an Arduino microcontroller. The detection model is built on MobileNetV2, a lightweight convolutional neural network architecture designed specifically for deployment on resource-constrained devices. Rather than training from scratch, the model is initialized from ImageNet-pretrained weights and fine-tuned on the PlantVillage dataset, which contains 54,306 labeled leaf images. Transfer learning in this manner dramatically reduces the training data and compute time required while preserving strong generalization capability. An entropy-based uncertainty filter is layered on top of the classifier so that inputs lacking sufficient confidence, such as non-leaf objects or blurry frames, are rejected rather than misclassified. The system is expected to achieve a validation accuracy of approximately 95.41% across all 38 classes, with per-frame inference latency of 30 to 60 milliseconds on a CPU fast enough to support smooth live detection through the webcam stream. Beyond accuracy, the work aims to demonstrate that a fully functional agricultural AI tool can be built.

Journal of Electronics and Telecommunication System Engineering

Smart Agriculture System for Real-time Plant Disease Detection Using Transfer Learning and Uncertainty-aware Deep Learning