Developing Wolaita Coffee Bean Quality Grading Model Using Deep Learning

Abstract

Ethiopia produces a wide range of coffee beans. Coffee is the nation’s primary export, a key source of foreign currency, and a vital income source for farmers. However, the approach to classifying the quality of Wolaita coffee beans is an expert observation with the naked eye. This approach is unfeasible due to subjective decisions, proneness to error due to fatigue, excessive processing time, and a smaller number of experts. Therefore, it is critical to automatically classify the quality classification of the Wolaita coffee bean through a model. Thus, this research tackles the challenges of manually classifying the quality of Wolaita coffee beans by developing an effective model. To do so, the 3133 image dataset was collected from the Ethiopian Commodity Exchange (ECX) Wolaita branch, annotated with domain experts, and preprocessed to make the dataset smooth and clean. In an attempt, the study explored custom Convolutional Neural Network (CNN), VGG19, and ResNet50 algorithms. All of the work was done using Python and the supporting libraries. Consequently, an experiment result shows that the VGG19 pre-trained model achieves an accuracy of 99.2% to classify the quality of the Wolaita coffee bean into six grades: G1, G2, G3, G4, G5, and UG (ungraded). Therefore, VGG19 was taken as an effective and robust performance for the quality grading of Wolaita coffee beans. Furthermore, automatically classifying the diverse range of agricultural scenarios and refinement of the model can be a future research direction.