Comparative Analysis of Transfer Learning Models and Custom CNN for Tea Leaf Disease Classification

Abstract

Tea is one of the most consumed drinks in the world and a savior of agricultural economies especially to nations like Bangladesh, India and Sri Lanka. However, diseases of the leaves like Algal Leaf, Gray Blight, Helopeltis, Looper Infestation, and Red Spider largely impact on the tea cultivation to the extent that they lower yield and quality. Conventional techniques are time consuming, error prone, inaccurate, and could not be applied to large plantations and required automatic and accurate solutions. This paper examines how deep learning models can be applied to the problem of tea leaf disease classification by comparing InceptionV3, MobileNetV2 or a Custom Convolutional Neural Network (CNN). The data of 1,968 pictures was gathered and enriched to 7,584 samples in order to provide diversity and robustness to the six classes. To evaluate the models, accuracy, precision, recall, F1-score and confusion matrices were used. Findings indicate that InceptionV3 had the highest accuracy of 98.94, Custom CNN had 96.6, and MobileNetV2 had 83.8. InceptionV3 was much more effective, but its high calculative cost renders it inappropriate in the field of real-time use. The Custom CNN was between the performance and accuracy, whereas MobileNetV2 was applicable to the deployment on mobile or edge devices owing to its lightweight architecture. The study exemplifies how deep learning can be applied in precision agriculture, and the lessons on what model to use based on resources and application. The results may be used to inform the further development of scalable, IoT- enabled disease detection to ensure sustainable agriculture and minimize the use of excessive amounts of pesticides. Deep learning-based classification of musculoskeletal radiographs: optimizing CNN architectures with model interpretability