| Crop growth monitoring and nutrition diagnosis based on remote sensing and spectral characteristics of UAVs is an important means to realize process management and digital decision-making in modern agricultural production.It is also a hotspot in modern remote sensing technology and smart agriculture research in recent years.This study aimed at the problems of single digital information acquisition,weak information collection infrastructure and inaccurate field dynamic management in the process of growth and development of drip irrigation in Ningxia.The use of drones to obtain digital image information in the field of maize seedlings,using digital image processing technology.The emergence status of maize was extracted,and the prediction model of maize emergence rate based on UAV remote sensing was established.The spectral information of functional leaves of various growth stages of maize was obtained by hyperspectral imaging system.The nitrogen content of maize canopy leaves was measured simultaneously,and maize was established by machine learning.A hyperspectral inversion model of nitrogen nutrition;after maize harvesting,the moldy maize kernels were identified using a self-developed machine vision recognition system to screen for moldy maize kernels.Providing technical support for high yield and accurate assessment of maize is of great significance for optimizing the precise planting technology of water and fertilizer integration in Ningxia maize.The specific research contents are as follows:(1)The drone is equipped with a digital camera to estimate the maize emergence rate.By obtaining high-definition images of maize seedlings,the OBLAB ORB algorithm and distance-weighted fusion algorithm are used to synthesize the image of the drone,through binarization,corrosion expansion and other depths.The optimized processing technique was used to obtain the outline of the maize seedling image,and then the MATLAB eight-bit connected domain and the ARCMAP 10.3 calculation method were used to automatically plan the route and calculate the number of maize emergence.Based on the field artificial survey data,a linear regression analysis method was established to establish a linear relationship between the number of maize seedlings and the number of maize seedlings.The results show that the linear regression model has a coefficient of determination of R2,RMSE and nRMSE of 0.895、4.359 and 2.436%.(2)Based on the 400 nm-1000 nm full-band hyperspectral maize nitrogen nutrition diagnosis,the average nitrogen content of 180 maize leaves fluctuated around 3.788%,and the maximum leaf nitrogen content was 4.984%,the minimum was 2.247%,and the standard deviation was is 0.541.Comparisons in the PLS model show that the PLS-De-trending pretreatment model is superior to several other pretreatment models and is superior to the original spectra to establish mathematical models.According to the conclusions of the modeling set and the prediction set,the coefficient of decision is changed from 0.527 to 0.519,and the root mean square error RMSE is changed from 0.372 to 0.374.On the contrary,the coefficient of determination of the prediction set increases from R2 from 0.578.By 0.621,the root mean square error RMSE of the prediction set decreased from 0.383 to 0.362.The characteristic bands are obtained by the SPA algorithm,which are 420.551 nm,425.352 nm,444.558 nm,468.566 nm,506.978 nm,511.780 nm,843.086 nm,895.903 nm,and 919.911 nm;for the nine bands,the fitting equation isy=1.18-0.627x not only effectively improves the accuracy and calculation speed,but also reduces the number of wavelengths in the full spectrum by 92.8%.While reducing the computational workload during the band extraction process,it also effectively improves modeling efficiency and stability.(3)Regression model built by back-propagation neural network(BP),compared with the modeling of partial least squares(PLS)model,the accuracy is obviously improved.The model set model with seven pre-processed full-spectral bands and the original full-spectral band is constructed by BP neural network.The coefficient R2 can reach 0.778 or more,and the average coefficient R2 is increased from 0.499 to 0.902 in the partial least squares model.The accuracy of the six preprocessing modeling models of SNV,MSC,De-trending,Baseline,1-Der and 2-Der all reached 0.905.The prediction set models RAW,SG,MSC,De-trending and Baseline determination coefficient R2 were all achieved.0.543 or more.(4)The regression model built by the support vector machine(SVM)has a significant improvement compared with the modeling precision of the partial least squares(PLS)model.The support model vector(SVM)is used to construct the model set model with seven pre-processed full-spectrum bands and the original full-spectrum band.The coefficient R2 is more than 0.529,and the average decision coefficient R2 is from the partial least squares(PLS)model.The 0.499 is raised to 0.613.The precision R2 of the three kinds of pre-processing modeling models including Baseline,1-Der and 2-Der all reached 0.606,and the prediction set model 1-Der and 2-Der determination coefficient R2 all reached 0.637.(5)Random forest(RF)algorithm combined with different pretreatments to establish a nitrogen nutrition diagnostic estimation model for the whole spectrum of 180 maize leaves.The minimum mean square error RMSE of the predicted set in the 7 pretreatments was 3.524.The effect of the goodness of fit in the table is similar to R2 in the regression analysis,and the maximum goodness of fit is 59.63%.Among them,the baseline algorithm(Baseline)preprocessing has the best modeling effect,and the mean square error RMSE of the prediction set model ranges from 3.524 to 6.019.Compared with the prediction accuracy,the best hyperspectral modeling method is the SVM model combined with the second derivative(2-Der)preprocessing model(predictive set R2=0.710,RMSE=O.363)with high precision and high reliability.Optimal silage maize nitrogen nutrition spectrum prediction model. |
PDF Full Text Request