Application Of UAV Phenotypic Technique In The Identification Of Yield Characters Of Winter Wheat

Wheat is one of the three major food crops in the world and the most important food crop in northern China.Affected by uncertain factors such as population growth,weather disasters,climate disasters,epidemic ecological disasters and hydrological disasters,my country’s food demand will maintain a rigid growth trend.The application of drones can non-destructively,quickly and accurately estimate wheat yield in a timely and efficient manner,which can effectively speed up the screening of high-yielding crop genotypes and assist molecular breeding,while providing agricultural production planning,food security,national economic and macro-decision-making.The machine learning method is used to fully tap the potential of multi-sensor remote sensing information of winter wheat canopy in winter wheat canopy in the estimation of wheat grain yield,and to clarify the improvement effect of multi-sensor data fusion on the estimation accuracy.The main research contents and results are as follows:（1）Based on the remote sensing information obtained by different sensors in each growth period,the visible light vegetation index and the multi-spectral vegetation index were extracted respectively,and the correlation analysis was carried out with the measured yield of the plot,and the significance test was carried out.The vegetation indices selected in the study（10 visible light vegetation indices and 13multispectral vegetation indices）were all significantly associated with the plot yield at the level of 0.01.The absolute value of the correlation coefficient between the multispectral vegetation index and the plot yield was between 0.40 and 0.84,which was greater than the absolute value of the correlation coefficient between the visible light vegetation index and the plot yield（0.37-0.66）.Comprehensively integrating the growth period of wheat,the correlation coefficient between flowering period and grain-filling period of most vegetation indices was greater than that of heading period and maturity period,and the overall characteristics of grain-filling period>flowering period>heading period>maturity period were presented.The grain filling period had better correlation with measured yield and was an ideal period for yield estimation.（2）Based on the remote sensing data of grain filling period,six algorithms of RR,SVR,RFR,GP,KNN and Cubist were used to establish a single sensor wheat yield estimation model for yield prediction.Based on the remote sensing data of the RGB sensor,the yield estimation model of the plot was constructed.The validation accuracy of the yield prediction model constructed by the six machine learning algorithms was up to 0.51,and the average RMSE and RRMSE were 380kg·hm^-2and 4.47%.The yield estimation model of the plot was constructed based on the remote sensing data of multi-spectral sensors.The validation accuracy of the yield prediction model constructed by the six machine learning algorithms was up to 0.69,and the average RMSE and RRMSE were 310 kg·hm^-2and 3.69%.Among them,the yield prediction models based on single sensor remote sensing data established by the Cubist algorithm all showed better accuracy.Overall comparison,the prediction accuracy of the multi-spectral sensor yield model is higher than that of the RGB sensor yield model,and the prediction accuracy of the RGB sensor and the multi-spectral sensor yield prediction model are all grain filling stage>flowering stage>heading stage>mature stage,which is consistent with the correlation of vegetation index.（3）The paper finds that compared with the single sensor production estimation model,the accuracy is limited by the sensor data type,multi-sensor data fusion shows great potential in UAV production estimation,and selecting a suitable algorithm is helpful to deal with multi-sensor input data The effective integration of,and then maximize the production estimation accuracy.The yield prediction model was constructed with the remote sensing data of the grain filling period as the input variable.The multi-sensor data fusion yield estimation model（R²=0.50-0.71）>multi-spectral sensor yield estimation model（R²=0.53-0.69）>RGB sensor yield estimation model（R²=0.35-0.51）.Based on the performance of each algorithm,the Cubist algorithm can better handle multi-modal fusion data,can effectively deal with multi-factor yield estimation,and perform multi-sensor data fusion.Compared with the single-sensor yield prediction model,the R²is increased from 0.48 to 0.71,and the RMSE is reduced to 290 kg·hm^-2,showing great potential.