Camellia Oleifera Yield Estimation Methods Based On Unmanned Aerial Vehicle(UAV)

As one of the main oil species,Camellia oleifera plays an important role in promoting the development of agricultural economy.Rapid and accurate estimation of its yield information is of great significance for guiding the accurate management of Camellia oleifera.The quantity of fruit and the parameter information of individual tree are important indicators to reflect the yield data.The traditional method of fruit sampling and weighing is time-consuming and laborious and cannot meet the demand of large-scale yield estimation.With the development of UAV and computer technology,it is possible to estimate the yield of camellia forest by remote sensing.In this context,this paper proposed a set of yield estimation scheme of Camellia oleifera forest based on UAV and deep learning technology,which is of great significance for realizing rapid and non-destructive yield estimation of large area Camellia oleifera forest.In this study,the Camellia oleifera plantation in Shaoyang City was taken as the research area.High-precision RTK UAV was used to obtain aerial images of Camellia oleifera,and 250 sample trees were randomly selected to obtain close-range images of the tree canopy.The fruits of sample trees were picked and photographed with cameras.Based on UAV images,multi-band fusion image(RGB(red,green,and blue),CHM(canopy height model),DSM(digital surface model),EXG(excess green index)as input data,six Res-UNet models were constructed and trained to extract single wood parameter information.Based on the close-up image of UAV and the fruit data captured by the camera,the Mask-RCNN model was used to detect and count the fruit on the surface of the crown and the result of picking.Then,combined with the measured yield of per plant and the average fruit weight,a Camellia oleifera yield estimation model based on the single tree parameters and the number of fruit information on the crown surface was constructed.The accuracy of the estimated yield was compared with the measured value,and the optimal yield estimate model was selected and applied to the estimated yield of Camellia oleifera in the whole study area to verify the feasibility of the proposed method.The main research conclusions include:(1)The Res-UNet model based on RGB-CHM band combination as input image training has the highest detection accuracy for crown single wood parameters,meeting the requirements of building yield estimation model based on single wood parameters.The estimation accuracy of the crown and its projection area is affected by the input elevation data(DSM or CHM),and the estimation accuracy of the model with CHM is higher.Among the six Res-UNet models trained according to different band combinations,the model based on RGB-CHM band combination performed tree crown detection(P=88.65%,R=81.20%,F1=84.76%)and crown width estimation(R2=0.9223,RMSE=0.0918m,rRMSE=3.81%)and crown projection area estimation(R2=0.9459,RMSE=0.2759m2,rRMSE=6.34%)had the best effect.In addition,the estimated tree height based on CHM extraction has a good correlation with the measured value(R2=0.795 7,RMSE=0.311 5m,rRMSE=16.17%).(2)The identification accuracy of crown surface fruit and picked fruit based on the Mask R-CNN model is generally high,which meets the demand of constructing Camellia oleifera yield estimation model based on the quantitative characteristics of fruit in the crown.The recognition accuracy,recall rate and F1 value of fruit were 97.15%,85.58%and 91.00%respectively under natural complex conditions.The recognition accuracy of fruit was 99.37%,the recall rate was 96.20%,and the overall accuracy was 97.76%.Based on Mask R-CNN model,the recognition model of crown surface fruit and picked fruit can meet the practical application demand of fast fruit number statistics.(3)The yield estimation model based on the crown surface fruit quantity information had the highest accuracy for Camellia oleifera yield estimation.When the number of crown fruit per tree was less than 100,the yield estimation accuracy of the model was R2=0.8060,RMSE=0.2612kg,rRMSE=24.25%.When the number of crown fruit per tree was more than 100,the model estimated yield accuracy was R2=0.8434,RMSE=0.9774kg,rRMSE=22.30%.According to the yield estimation model,the yield of fresh fruit per mu is about 211.18 kg.The results showed that the method of Camellia oleifera yield estimation based on UAV remote sensing and deep learning technology was feasible.This method was time-saving and could meet the demand for rapid yield estimation of large area Camellia oleifera forest.