Abstract:
When applied to large amounts and a variety of tea leaf variations, manually identifying and
grading tea types is laborious, subjective, and frequently unreliable. Customer trust, market
pricing, and product quality may all be impacted by this discrepancy. Furthermore, the tea
industry's ineffective data integration techniques, lack of automated grading tools, and restricted
traceability provide difficulties for growers and exporters. By developing a web-based application
that recognizes tea grade, tea type and elevation, this project seeks to close these gaps and
specifically target the Sri Lankan tea market.
The project combines cutting-edge geospatial data visualization and machine learning technology
to achieve these objectives. A deep learning-based solution was created to overcome these
obstacles by integrating OpenCV for picture preprocessing with the YOLOv8 model. Through
feature extraction from picture datasets, the model was trained to identify and categorize three
distinct tea grades across different varieties. RESTful APIs were constructed using Flask, allowing
the backend classification engine and the React-based frontend to communicate in real time. To
handle structured data, such as pictures, forecasts, and traceability metadata, MongoDB was used.
Several data science measures were used to assess the system's performance. Throughout the test
set, the model's accuracy was over 80%, and its precision and recall values showed that it
performed well in reducing false positives and false negatives. F1-scores demonstrated outcomes
that were balanced between recall and precision. For BOP (92%) and Silver Needle (89%), the
model's confidence was high; however, for OP, it was lower (71%), indicating that there may be
space for improvement in OP classification. These outcomes open the door for the system's
implementation in actual tea production and export settings by confirming its resilience in
accurately recognizing tea varieties and grades.
