Detection of tea leaf diseases using deep transfer learning

Abstract

Tea leaf diseases significantly impact both the quantity and quality of tea production in Sri Lanka, a country where tea cultivation holds considerable economic importance, contributing significantly to its GDP and serving as a major export to consumer markets. Existing computer vision and machine learning methods require a large number of image samples for accurate classification, leading to a time-consuming process. To address this limitation, we propose a novel approach utilizing deep transfer learning to train classification models efficiently with limited samples, leveraging cross-domain knowledge transfer. Our method aims to detect tea leaf diseases early, thereby preserving tea quality and fostering sustainable agricultural practices. The unique contributions of this study are a) collecting a comprehensive set of tea leaf images from different tea gardens representing six tea leaf conditions, annotated manually and b) developing a pre-trained convolutional neural network (CNN) architecture, with 256, 128, and 6 fully connected layers, including Xception, DenseNet201, VGG16, InceptionV3, EfficientNetB0, and MobileNetV2, to transfer classification knowledge. Through several experimentations with various fine-tuning techniques, we achieved a notable average accuracy of 99.58% in classifying tea leaf diseases.