Acquisition And Methodology Of Oilseed Rape Phenotyping Based On Low-altitude Remote Sensing Of UAV

Field-based crop phenotypic information is a visual representation of crop quality and growth information,and is an important factor affecting crop production management decisions.The real-time,efficient and accurate acquisition of phenotypic information of plants can provide a basis for accurate management of modern agricultural information,and efficient management of water and fertilizer,as well as support for rapid and accurate identification of growth information and crop breeding.The low-altitude remote sensing platform of unmanned aerial vehicles(UAVs)has become an emerging means of obtaining phenotypic information of field crops by virtue of its high spatial and temporal resolution,low cost,flexibility and adaptability to complex farmland environments.Therefore,we mainly focused on the key technologies for the acquisition of oilseed rape(Brassica napus L.)phenotypic information by UAV-based low-altitude remote sensing.As one of the most important oil crops,oilseed rape has high economic value,which is grown all around the world.Field experiments were conducted with different nitrogen gradients.Through the multi-source information from different sensors,the robust quantitative models of phenotypic information(flower number,SPAD,leaf area index,and yield of oilseed rape)were constructed.Combined with the variable selection methods and machine-learning algorithms,the accuracy and robustness of the models were improved.The results provided theoretical basis and technical support for high-throughput crop phenotyping in precision agriculture.Additionally,this study had important guiding significance for yield estimation and breeding of crops.The main contents and results were as follows:(1)With multi-source information fusion by combining vegetation indices(VIs)and image segmentation from UAV-based RGB and multispectral cameras,robust prediction models of flower number in oilseed rape were built.The results indicated that(1)a set of processes for splicing multispectral images was proposed for UAV-based multispectral imaging:vignetting correction by genetic algorithm fitting Gaussian surface-extracting features by scale invariant feature conversion(SIFT)algorithm-eliminating mismatched points by random sample consensus(RANSAC)algorithm-image fusion based on Laplacian pyramid.(2)The image segmentation of yellow flowers was realized by RGB threshold,or machine learning algorithms which including back propagation neural network(BPNN),support vector machine(SVM),K-means clustering.The best segmentation method was K-means clustering based on CIELAB color space.(3)Based on a variety of machine learning algorithms including multiple linear regression(MLR),partial least squares regression(PLSR),BPNN,least squares support vector machine(LS-SVM),and extreme learning machine(ELM),linear and nonlinear prediction models of flower number in oilseed rape were established for UAV-based RGB camera,multispectral camera and dual cameras.The regression models based on effective features(narrow-band VIs,color VIs,and flower coverage area)from the dual sensors were better than the models established by the multispectral camera or RGB camera.Additionally,all prediction models achieved good results with correlation coefficient of prediction(Rpre)above 0.89.The Rpre and root mean square error of prediction(RMSEP)of the optimal BPNN model were 0.9359 and 17.25,respectively.(4)To improve and simplify the estimation models,genetic algorithm was employed to rank the importance of variables and select the optimal variables.The simplified BPNN model had the best prediction performance(Rpre=0.9383,RMSEP=14.81).(2)The quantitative prediction models of nitrogen(SPAD)and leaf area index(LAI)of oilseed rape were built using UAV-based multispectral or UHD185 hyperspectral imaging system.The results using UAV-based 25-band multispectral camera indicated that:(1)linear(MLR and PLSR)and nonlinear(LS-SVM,BPNN,ELM and radial basis function neural network(RBFNN))prediction models were established for SPAD and LAI based on the full bands.The optimal prediction models for SPAD and LAI were ELM(Rpre=0.8593,RMSEP=0.7798)and BPNN(Rpre=0.8016,RMSEP=0.4579),respectively.(2)The optimal VIs combination,which were sensitive to SPAD and LAI prediction,were determined by the correlation coefficients contour maps.Moreover,the prediction models of SPAD and LAI were established by using different machine learning algorithms.The optimal prediction models of SPAD and LAI using the effective VIs were ELM(Rpre=0.8296,RMSEP=0.8627)and BPNN(Rpre=0.7416,RMSEP=0.5134),respectively.The results using UAV-based UHD185 hyperspectral camera indicated that(3)successive projections algorithm(SPA),genetic algorithm-partial least square(GAPLS),the combination of uninformative variable elimination(UVE)and SPA,competitive adaptive reweighted sampling(CARS),and random frog(RF)were compared and evaluated for effective wavelengths(EWs)selection.The performance of calibration and prediction models using EWs manifested better results than those using full-spectra.The CARS-BPNN model for SPAD(Rpre=0.9146,RMSEP=0.5973)showed excellent performance,whereas RF-BPNN was the optimal model for predicting LAI(Rpre=0.9152,RMSEP=0.1970).(4)Correlation analysis and one-way analysis of variance(ANOVO)were used to explore the optimal VIs for SPAD and LAI estimation.Different machine learning algorithms were developed to quantitatively predict SPAD and LAI using the VIs.The optimal prediction models of SPAD and LAI using the effective VIs were ELM(R_pre=0.8974,RMSEP=0.5596)and BPNN(R_pre=0.8699,RMSEP=0.2265),respectively.In summary,the optimal models established by the UAV-based UHD185hyperspectral camera(SPAD:Rpre=0.9146;LAI:Rpre=0.9152)were superior to those by the UAV-based 25-band multispectral camera(SPAD:Rpre=0.8593;LAI:Rpre=0.8016).It was proved that the UHD185 hyperspectral imager could obtain a large number of narrow-band and continuous hyperspectral images,which more fully presented the spectral characteristics of oilseed rape.Thus,this UAV-based hyperspectral imager has wide application potential in quantitative remote sensing of agriculture.(3)Through multi-source information fusion,the quantitative prediction models of yield of oilseed rape were built with UAV-based RGB and multispectral cameras.The results indicated that(1)the regression models(PLSR,MLR,LS-SVM,BPNN,ELM,and RBFNN)based on full-spectra from multispectral camera performed well,with R_caland R_pre both greater than 0.78.The optimal prediction model was BPNN(R_pre=0.8232,RMSEP=166.9 kg/hm²).(2)Correlation analysis and ANOVO were used to explore the optimal VIs for yield estimation.Different machine learning algorithms were developed to quantitatively predict yield.The performance of calibration and prediction models using the combination of VIs(RBFNN:Rpre=0.8143,RMSEP=171.9 kg/hm²)from multispectral and RGB cameras manifested better results than those using only narrow-band VIs(BPNN:Rpre=0.7655,RMSEP=188.3 kg/hm²)from multispectral camera.(3)Genetic algorithm was employed to select the optimal EWs and VIs.The best models for yield prediction were BPNN(Rpre=0.8114,RMSEP=172.6 kg/hm²)using optimal EWs and ELM(Rpre=0.8118,RMSEP=170.9 kg/hm²)using optimal VIs.Hence,the above results demonstrated the feasibility of using UAV-based dual-sensor platform to predict the yield of oilseed rape.Additionally,they demonstrated that a lightweight UAV with dual image-frame snapshot cameras has the great potential for high-throughput plant phenotyping and advanced breeding programs in precision agriculture.(4)Based on UAV-based UHD185 hyperspectral imager and Gaia Sky-mini hyperspectral imager,the optimal yield prediction model was obtained from the remote sensing images of oilseed rape at the pod formation stage,combined with wavelength selection algorithms,VIs optimization and machine learning algorithms.The results indicated that(1)comparing the effects of various preprocessing methods on spectral characteristics,wavelet transform was the optimal method.(2)Eight wavelength selection methods,including SPA,GAPLS,UVE,UVE-SPA,weighted regression coefficient(BW),second derivative(2-Der),CARS and RF,were compared and evaluated for EWs selection.The performance of calibration and prediction models using EWs manifested better results than those using full-spectra.Additionally,GAPLS,CARS,RF and UVE-SPA were the effective methods.The ELM models achieved excellent results.Particularly,the CARS-ELM model(Rpre=0.8122,RMSEP=170.4 kg/hm²)using UHD185 imager showed excellent performance,whereas RF-ELM(Rpre=0.8227,RMSEP=166.4 kg/hm²)was the optimal model using Gaia Sky-mini imager for predicting yield.(3)Correlation analysis and ANOVO were used to explore the optimal VIs for yield prediction.The prediction models using combined VIs were more stable and better than the models using single vegetation index.Overall,the optimal estimation model of Gaia Sky-mini imager based on combined VIs(LS-SVM:Rpre=0.8170,RMSEP=172.5 kg/hm²)was superior to the optimal model using UHD185 imager(ELM:Rpre=0.7674,RMSEP=187.6 kg/hm²).In summary,the UHD185 and Gaia Sky-mini UAV-based hyperspectral remote sensing systems achieved good prediction results for yield of oilseed rape(UHD185:Rpre=0.8122;Gaia Sky-mini:Rpre=0.8227).Moreover,the estimation models for yield with Gaia Sky-mini imager were slightly better than those of UHD185 imager.It was proved that the lightweight UAV with dual image-frame snapshot cameras(RGB and multispectral cameras),and the UAV-based hyperspectral remote sensing systems had great advantages in obtaining phenotypic information of oilseed rape.A technical route was proposed for low-altitude remote sensing of UAV:image acquisition-image stitching and processing-effective wavelengths or vegetation indices selection-linear and nonlinear prediction models using machine learning algorithms.They also provided an integrated solution for the application of UAV-based remote sensing information acquisition in precision agriculture.