Abstract
Tea plant germplasm resources are abundant in China. Effectively identifying tea varieties and evaluating the tea bud sprouting period (BSP) are critical for tea plant breeding and tea plantation management. This study focused on key tea varieties cultivated in the tea-growing regions of Shandong Province, China. UAV-based multispectral and RGB images of seven tea varieties were collected before and during the budding period in spring. Five machine learning classification algorithms, LSTM, BP, PSO-BP, GA-BP and SVM were employed to classify tea varieties and BSP. Results indicated significant spectral differences among tea varieties across six spectral bands. Spectral differences among varieties collected during budding became more pronounced than that pre-budding period. Machine learning techniques effectively distinguished different tea varieties and BSP. Models of LSTM, BP and PSO-BP established by data collected during budding enhanced the classification accuracy of tea varieties by 0.45 ~ 2.11% than that before budding. The integration of indices and texture features from both sampling periods further improved classification accuracy of tea varieties. BP model achieved the highest variety and BSP classification accuracy with a test set accuracy of 93.65% and 92.86%, respectively, followed by LSTM with the accuracy of 93.65% and 90.48%, respectively. Considering computational speed and accuracy, the BP classification model was well-suited for real-time classification needs in various application scenarios. This study provides technical support for large-scale tea variety classification and budding phenology monitoring. It also serves as a valuable reference for the rapid screening, identifying, and improving tea plant superior germplasm, thereby enhancing breeding efficiency.
