Abstract:
Tea planters face tremendous challenges in the identification of leaf diseases and determination of
the optimal harvesting time. Manually applied inspection methods are often ineffective,
inconsistent, and subject to human mistakes, leading to reduced yield quality and economic loss.
While machine learning has been increasingly popular in agriculture, the majority of the existing
models only tackle disease detection or ripeness assessment separately. This lack of a unified
solution brings inefficiencies, resulting in farmers utilizing many systems. Additionally, high costs
of computations and inability to generalize for multiple teas further limit the utilization of such
technologies in the field.
To address these challenges, the present paper proposes TeaGuard, a Progressive Web Application
(PWA) that utilizes EfficientNetB3 and transfer learning for real-time tea leaf disease classification
and tea ripeness assessment. The system captures tea leaf images using preprocessing techniques
such as image resizing and normalization to improve model performance. High-speed
communication between the frontend and machine learning models is supported by the FastAPI
backend with the best performance in real-time inference. The model employs a lightweight but
efficient architecture with high accuracy at the cost of efficiency in deployment in low-resource
environments.
The model for classifying diseases achieved 94% accuracy, which presents very good classification
performance on diverse categories of disease. The ripeness estimation model now achieves 84%
accuracy, which is a significant advancement in predicting ideal harvest times. The system was
validated through functional, non-functional, and model tests, which confirmed the robustness,
scalability, and feasibility of the system for deployment.
