A Non-Destructive Method for Identification of Tea Plant Cultivars Based on Deep Learning

Tea plant cultivar identification is normally achieved manually or by spectroscopic, chromatographic, and other methods that are time-consuming and often inaccurate. In this paper, a method for the identification of three tea cultivars with similar leaf morphology is proposed using transfer learning by five pre-trained models: EfficientNet-B0, MobileNetV2, MobileNetV3, MobileViT-S, and ShuffleNetV2. The results showed that the best test accuracy percentages for EfficientNet-B0, MobileNetV2, MobileNetV3, MobileViT-S, and ShuffleNetV2 were 98.33, 99.67, 99.33, 98.67, and 99.00%, respectively. The most lightweight model was ShuffleNetV2, and the fastest combination was ShuffleNetV2 with 112 × 112 image resolution. Considering accuracy, the number of parameters, and floating point operations (FLOPs), MobileNetV2 was not only the most accurate model, but also both lightweight and fast. The present research could benefit both farmers and consumers via identifying tea cultivars without destructive techniques, a factor that would reduce the adulteration of commodity tea.