Pixel-class prediction for nitrogen content of tea plants based on unmanned aerial vehicle images using machine learning and deep learning
Shu-Mao Wang,Jun-Hui Ma,Zhu-Meng Zhao,Hong-Zhi-Yuan Yang,Yi-Min Xuan,Jia-Xue Ouyang,Dong-Mei Fan,Jin-Feng Yu,Xiao-Chang Wang
DOI: https://doi.org/10.1016/j.eswa.2023.120351
IF: 8.5
2023-05-13
Expert Systems with Applications
Abstract:Nitrogen management in tea ( Camellia sinensis L.) fields has deeply depended on manual work for a long term due to low-cost labor. Because of the decreasing population, costs of labor have been rising. Computer technology recently has made great success in many fields, and intelligent tea fields expert system will be updated again. Nitrogen management with computer technology instead of manual work is an inevitable attempt. In this study, UAVs were used to take photos of tea plants, meanwhile, the elder tea leaves and tea shoots were collected to determine the nitrogen content. According to the HSV color system, pixels of background, elder tea leaves, and tea shoots were screened respectively. The experimental strategy I and II then were established. In experimental strategy I, pixel-class information of elder tea leaves and tea shoots were used to calculate weighted average values on RGB and HSV color systems, respectively. The values then were respectively combined with the nitrogen content of elder tea leaves and tea shoots to build datasets. In experimental strategy II, after suppressing the background, not only the weighted average values on RGB and HSV color systems in the rest each photo but also the proportion of elder tea leaves and tea shoots were computed. The proportion values were respectively multiplied by the nitrogen content to obtain a new weighted nitrogen content, and then with weighted RGB and HSV values together formed datasets. The ordinary least squares (OLS), extreme gradient boosting (XGBoost), long short-term memory (RNN-LSTM), convolution neural network (CNN), and residual network (ResNet) models were designed and carried out on predicting nitrogen content. Generally, the performance of HSV datasets was better than that of RGB datasets. ResNet and CNN models performed best in experimental strategy I; the ResNet model had the best results in experimental strategy II. Both experimental strategies performed well when predicting nitrogen content in spring tea; as for summer tea, experimental strategy II had better performance. This method had the advantages of convenience and accuracy and could elevate intelligent tea field management. As a part of precision agriculture, it could improve intelligent tea fields expert system, and further deal with challenges caused by the labor population decrease.
computer science, artificial intelligence,engineering, electrical & electronic,operations research & management science