A lightweight deep learning framework for tea cultivar identification based on leaf images

Accurate tea cultivar identification is a long-standing challenge in tea plant propagation and cultivation. Traditional cultivar identification based on morphological observation relies heavily on expert experience, leading to strong subjectivity; while DNA fingerprinting-based approaches are time-consuming and costly, restricting their large-scale application. This study attempts to construct a deep learning- based on leaf morphological and visual characteristics. We collected an original dataset comprising 11,404 leaf images from 10 clonal tea cultivars for model development and evaluation. Comparative experiments were conducted using five models: ResNet50, EfficientNet-B3, MobileNetV3-large, ViT-B16, and ConvNeXt-Tiny. ResNet50 achieved the highest classification accuracy, whereas MobileNetV3-Large provided a better balance between accuracy and computational efficiency and was therefore selected as the baseline model for lightweight optimization. Channel pruning and PTDQ were further applied to compress and optimize MobileNetV3-Large, resulting in a lightweight and efficient tea cultivar identification model. The optimized model achieved a test accuracy of 97.15% on the independent testing set, while reducing the model size from 34.40 MB to 7.67 MB and decreasing inference latency from 42.35 ms to 27.67 ms. Future work will focus on integration into Android-based field applications. This study provides an efficient lightweight solution for practical tea cultivar identification and mobile deployment.