Multi class tea leaf disease classification using feature level and output level ensemble strategies with Grad-CAM visualization

Tea is one of the most widely consumed drinks in the world and plays an important role in the economy of tea-growing regions. However, leaf diseases are a major problem for the tea industry because they lower both yield and quality, directly affecting tea growers’ livelihoods and the overall supply chain. Therefore, detecting these diseases in their early stages is important for healthy crop growth and good yield. In traditional practice, disease identification is carried out through field inspection or laboratory testing by expert farmers and plant pathologists. However, these methods are slow, require a lot of manual work and may sometimes lead to errors, making them difficult to adopt for large-scale cultivation. This study presents a novel deep learning-based methodology for eight-class tea leaf disease classification, exploring feature-level and output-level ensemble strategies. Four widely used convolutional neural networks (CNNs)-ResNet-18, VGG-16, InceptionV3 and MobileNetV2-pretrained on ImageNet were utilized. In the feature-based approach, deep features extracted from these models were compressed using principal component analysis (PCA) and classified using a Random Forest classifier, achieving an accuracy of 95.6 %. In the output-based approach, probability predictions from the above CNNs were combined, resulting in a higher accuracy of 98.3 %. Grad-CAM visualizations confirmed that the models consistently highlighted symptomatic leaf regions, improving interpretability and user trust.