Towards Precision Withering in Tea Factories with Non-Invasive Leaf Moisture Estimation

• Published in: 2024 16th International Conference on COMmunication Systems & NETworkS (COMSNETS) pp. 177 - 182
• Main Authors: Kumar, Abhishek, Choudhury, Swagatam Bose, Junagade, Sanket, Sarangi, Sanat, Singh, Dineshkumar, Pappula, Srinivasu
• Format: Conference Proceeding
• Language: English
• Published: IEEE 03.01.2024
• Subjects: Chemicals; Estimation; Food processing; Machine learning; Moisture; Moisture estimation; Production facilities; Spectral imaging; Support vector machines; Tea withering
• ISSN: 2155-2509
• DOI: 10.1109/COMSNETS59351.2024.10427332

The quality of manufactured tea is governed by various steps of which withering is one of the key processes. Precise control on the level of withering is foundational for producing good quality tea. Estimating the moisture content precisely is an important factor in determining the level of physical and chemical withering that the leaves have gone through. Tea factories in India have traditionally used a system of weighing the leaves before and after oven-drying to estimate the level of withering, which is a time-consuming and destructive process. With multiple withering troughs operating in tandem within a tea-factory, factory staff often just use their experience to gauge the withering level by touching the leaves to make time-efficient but subjective decisions. These challenges need to be overcome by rapid non-destructive techniques that can precisely assess the withering levels. In the withering carried out for the highly prevalent Assam tea consumed throughout the country, moisture level in leaves typically drop to about 66% before they are taken to the next stage of processing. By harnessing spectral data from a sample of leaves, we propose a method to non-invasively estimate the level of withering from the batch of leaves based on their inherent moisture levels. Multiple algorithms were used to develop and compare variants of the estimation model. We found that Support Vector Machine (SVM) model had the best overall accuracy of 85.41%.