Quantitative Inversion Of Maize Phenotypic Information In Breeding Plots Based On Multi-source Remote Sensing

Remote sensing technology is fast,objective,non-destructive and high-throughput acquiring information of ground and has important application in crop growth monitoring and yield prediction.At present,the lag of phenotypic information acquisition seriously hinders development of things,such as genetics analysis,genome-wide association study and so on,which becomes one of bottlenecks in the research of crop breeding.The actual situation of field breeding is that there are many breeding materials.When obtaining phenotypic information of breeding materials,a traditional manual method not only is exhausting and time-consuming but also has low efficiency and there is a phenomenon of different standards due to different subjective factors of surveyors,which exists large measurement errors.Therefore,quantitative retrieval of phenotypic information of breeding materials using remote sensing technology has become the focus of objective,rapid,non-destructive and accurate acquisition of phenotypic information.However,there is less research on the quantitative acquisition of phenotypic information of breeding materials in the field by remote sensing technology.In view of this,the paper studied the quantitative inversion of maize phenotypic information in the breeding plot of multi-source remote sensing.Experiment of the study was carried out at National Experiment Station for Precision Agriculture in town of Xiaotangshan,Changping District,Beijing City.Unmanned aerial vehicle(UAV)high-resolution digital and multi-spectral images and near earth hyperspectral data were obtained from May to September in 2017.Phenotypic data of plant height,leaf area index(LAI)and dry-biomass were obtained at the same time.The main contents and conclusions of the paper were as follows.Firstly,the digital surface model(DSM)and digital orthophoto map(DOM)of the experimental field were mosaiced using high-resolution images of UAV and ground control points(GCPs)of seeding,jointing,trumpet and anthesis silking stage.Based on DSM at different growth stages,plant height of maize breeding materials was extracted and had highly consistent with plant height of measurement values(R~2,RMSE and nRMSE were 0.93,28.69cm and 17.90%,respectively).Based on hyperspectral data of maize breeding materials canopy,partial least squares(PLS)regression analysis,support vector machine(SVM)regression analysis,and random forest(RF)regression analysis were applied to estimate plant height.The random forest model had the best estimation accuracy.The R~2,RMSE,nRMSE of model establishment and model validation of RF were 0.95,15.55cm,10.37%and 0.73,30.35cm,21.01%,respectively.By analyzing the growth rate of plant height,the average growth rate of plant height was 5.37 cm per day at the time of 45 to 57 days after planting.Secondly,based on high-resolution digital and multi-spectral images of UAV,the estimation of maize breeding materials LAI were carried out.The estimation model of stepwise regression analysis for LAI was carried out by using information extracted from high-resolution digital image of UAV.The R~2,RMSE,nRMSE of the best model establishment and validation were 0.63,0.40,26.47%and 0.68,0.38,25.51%using digital image information,respectively.When using the fusion data of plant height and digital image information,the R~2,RMSE,nRMSE of the best model establishment and validation were 0.69,0.37,24.34%and 0.73,0.35,23.49%.The way of fusion plant height can overcome the phenomenon of spectral saturation of canopy to a certain extent by comparing the two methods.Multi-spectral vegetation index were calculated by extracting different band information from multi-spectral images.PLS,SVM and RF regression analysis were used to estimate LAI.The random forest model had the best estimation accuracy.The R~2,RMSE,nRMSE of the RF model establishment and validation were 0.92,0.18,10.75%and 0.76,0.37,21.11%.Thirdly,the canopy hyperspectral data of canopy were used to estimate dry-biomass of the maize breeding materials.PLS,SVM and RF regression analysis were used to estimate dry-biomass using information by screening sensitive single band spectral reflectance data and combining hyperspectral vegetation index.At the same time,whether the canopy spectral information and plant height information were fused was compared and analyzed.The SVM model had the best estimation accuracy when canopy spectral information was used only.The R~2,RMSE,nRMSE of the SVM model establishment and validation were 0.83,1590.41kg/ha,32.07%and 0.80,1613.87kg/ha,33.97%.The PLS model had the best estimation accuracy when both canopy spectral information and plant height information were used.The R~2,RMSE,nRMSE of the PLS model establishment and validation were 0.92,1055.76kg/ha,21.71%and 0.92,1062.91%,22.37%.The comparison showed that the prediction accuracy of model could be improved by fusion of plant height information and canopy spectral information to estimate dry-biomass of maize breeding materials.In summary,remote sensing data obtained in the paper including UAV high-resolution digital and multi-spectral images and near earth hyperspectral data.Based on remote sensing data,field phenotypic parameters of plant height,LAI and dry-biomass of maize breeding material were analyzed.Quantitative inversion models of remote sensing for plant height,LAI and dry-biomass phenotypic parameters were constructed and phenotypic parameters were objectively,quickly,non-destructively and accurately estimated.The research of the paper was a useful exploration for acquisition of field crop phenotypic information by remote sensing technology.The research further improved technology of maize phenotypic information acquisition and helped to promote application of remote sensing technology in the field of phenotypic information acquisition in crops.