Abstract:
Tea cultivation plays a vital role in the agro economy of countries like Bangladesh, contributing significantly to both economic output and rural livelihoods. But various diseases of tea leave such as Algal Leaf Spot, Red Rust, Brown Blight, Gray Blight, Helopeltis often disrupt the productivity of quality tea crops. These conditions not only reduce yield but also affect leaf quality, thereby diminishing commercial value. Timely and precise disease detection is critical for mitigating such losses, yet traditional methods remain manual, time-consuming, and prone to human error. In this study, I propose a deep learning based automated detection framework utilizing transfer learning and fine tuning strategies to classify six categories of tea leaf health conditions, including healthy leaves. Augmentation methods helped to increase an original dataset of 1,885 background removed image data to over 7,000, hence correcting class imbalance and enhancing generalization. Performance was assessed using several pre-trained convolutional neural network (CNN) models: Custom CNN, DenseNet121, InceptionV3, ResNet50, VGG19, Xception and a hybrid InceptionV3-LSTM. Among them, InceptionV3 yielded the highest test accuracy of 97.35%, followed by Xception with 94.69%, and DenseNet121 with ~88%. The custom CNN, although basic, reached 71.29% test accuracy. The hybrid InceptionV3-LSTM model achieved 80.66%, indicating the potential of sequence modeling in visual classification. Meanwhile, ResNet50 showed the lowest performance with 67.55% accuracy, possibly due to underfitting on augmented data. The results show how well modern CNN architectures work in precisely detecting tea leaf diseases and offer a solid foundation for the development of real time agricultural monitoring systems. The proposed system can empower tea growers in Bangladesh with an intelligent, scalable, and accessible tool for early disease detection, contributing to sustainable crop management and economic resilience. Keywords: Tea Leaf Disease, Deep
