Image-Based Coffee Plant Disease Detection using Transfer Learning

ABSTRACT Agriculture stands as the cornerstone of Ethiopia's economy, accounting for a staggering 80- 85% of its GDP. Among the nation's agricultural treasures, coffee holds a special place, serving as a vital export commodity and a source of daily sustenance for countless farmers. The contribution of coffee to Ethiopia's economy is immense, generating substantial foreign exchange earnings and fueling the livelihoods of rural communities. Given the pivotal role of coffee in Ethiopia's economic landscape, ensuring its quality and safeguarding it from diseases takes paramount importance. However, the current practice of relying on manual disease identification by experts faces several limitations. This method is not only time-consuming and labor-intensive but also lacks scalability, making it impractical to cover all coffee-growing regions effectively. Timely detection of coffee diseases is crucial for minimizing damage and preventing significant economic losses. Early intervention can halt the spread of diseases, preserving the health of coffee plants and ensuring a bountiful harvest. Therefore, the development of an automated disease identification system is an urgent necessity for the Ethiopian coffee industry. This paper proposes an automatic coffee disease identification system using transfer learning and recommends treatments in Afan Oromo. The system extracts important features from coffee plants and uses them to classify the plant into five categories: coffee berry disease (CBD), coffee leaf rust (CLR), Phoma leaf spot (PLS), brown eye spot (BES), and healthy coffee. To achieve the best results during the classification of such diseases, we compared training a deep learning model from scratch and using transfer learning with a pre-trained ResNet50 model to classify coffee diseases. The ResNet50 model achieved a training accuracy of 97.6% and a test accuracy of 95%, significantly outperforming the model trained from scratch, which achieved an accuracy of 89.3%. The system was trained on a dataset of 5010 coffee plant images and achieved an accuracy of 95% with the ResNet50 architecture and 92% with the VGG19 architecture. These results suggest that deep learning algorithms have the potential to be used for automatic and accurate coffee disease identification.