Abstract:
Accurate and rapid acquisition of nitrogen content of winter wheat in critical growth period plays an important role in decision-making of nitrogen fertilization in field. Using unmanned aerial vehicle (UAV) with digital camera, the growth information of winter wheat can be obtained in a short time, and the dynamic monitoring of nitrogen content of winter wheat can be realized. In this study, three thresholding segmentation methods were used to separate the field plant crops from the soil background based on the digital image of winter wheat UAV in Xiaotangshan, Beijing, in 2015. By comparing the timeliness and accuracy of image segmentation methods, the visible-band difference vegetation index (VDVI) was finally determined to extract vegetation information. According to the requirements of the experiment scheme, winter wheat was divided into 48 material plots by three repeated experiments under different nitrogen and water stress management. In order to increase the difference of crop nitrogen content in each experimental plot, two different varieties of wheat were planted, and different water and nitrogen supply were added at the same time. Each treatment scheme was repeated three times. A total of 48 experimental plots were designed. The planting area of each plot was 48 m². Draw lessons from construction method of hyperspectral vegetation Index, 25 vegetation indices were constructed according to the average DN (digital number) values of red, green and blue channels extracted from the plot boundary. The correlation analysis was used to screen the digital image variables between the constructed vegetation index and the nitrogen content of different components of winter wheat in each material plot. Because of the high coupling degree between vegetation indices, principal component analysis was used to reduce the dimension of original data and extract feature vectors to participate in modeling. Various factors affecting the selection of modeling parameters and modeling were discussed. The nitrogen retrieval model was established by multiple linear regression analysis, and the best model was selected by determining coefficient (R2), root mean square error (RMSE) and normalized root mean square error (nRMSE) to explore the sensitivity of nitrogen content and digital variables. Using the model and UAV digital image, the retrieval image of winter wheat nitrogen was drawn, which visually display the spatial distribution of winter wheat nitrogen content. The results showed that the estimated value of winter wheat nitrogen content retrieved from UAV digital image had high fitting accuracy with the measured data. In terms of the accuracy of the inversion model, the three data processing results were integral segmentation > partial segmentation > segmentation by VDVI. The inversion effect of nitrogen content in different organs of winter wheat was different. Taking the flag-flying period of winter wheat as an example, the R2 and nRMSE of the verification model of integral segmentation was 0.85 which was 0.14 and 0.43 higher than that of the partial segmentation and VDVI segmentation, RMSE and nRMSE of the verification model of integral segmentation was 0.235 and 6.1% respectively, which was 0.068 and 1.77 percentage points lower than those of the partial segmentation, 0.141 and 3.67 percentage points lower than those of VDVI segmentation, respectively. The results can provide reference for decision-making and management of water and fertilizer in winter wheat field.