Extraction And Modeling Of Physical And Chemical Properties Of Fruit Trees Flowers And Leaves Based On Spectral Analysis Technique

High-yield fruit trees play an important role in ensuring agricultural production.How to accurately obtain the information of fruit trees by modern means is becoming an urgent problem.The integration of remote sensing(RS)and geographic information system(GIS)could provide a theoretical basis for the monitoring of growth condition,health condition and information classification identification of fruit trees.With the development of RS technology,a lot of researches have been conducted on the information extraction of plant species,but due to the limitation of wave band and low resolution rate,there are few reports of researches on the RS information knowledge.Characterized by high resolution and wide wave band,the hyperspectral remote sensing enables it to extract the information of fruit trees rapidly and accurately without damage.By analyzing the spectral information of fruit trees,it is possible to identify the species of fruit trees identify diseases and pests and estimate nutrition,and the characteristic wave band obtained through the ground spectral analysis can provide a reference basis for the hyperspectral remote sensing images.In this study,with fruit trees as the research target,the data of ground spectral determination were analyzed with near infrared spectroscopy technology and data mining technology.The classification of flowering spectra of four fruit trees and the identification of three kinds of apple leaf diseases and healthy leaves were realized,and contents of P,Fe,B and Mn in apple leaves were estimated,and timing characteristics were added to the modeling of mineral nutrients in apple leaves to improve the model accuracy.The conclusions are as follows:1.In order to establish an effective model for species identification of flowering trees,the spectral data of four flower trees were collected by ASD FieldSpec 3 full band portable spectrum analyzer.On the basis of preprocessing the optimal variable,three identification models,namely partial least squares discriminant analysis(PLS2-DA),orthogonal partial least squares discriminant analysis(O-PLS-DA)and BP neural network were established,and the identification rates of the tested samples in descending order were respectively BP(96.36%)>O-PLS-DA(84.55%)>PLS2-DA(82.73%),where compared with PLS2-DA,O-PLS-DA had the modeling accuracy increased because of filtration of unrelated information in the independent variable,and as an artificial intelligence method,BP neural network could realize the complicated nonlinear mapping,with self-learning ability and the highest identification rate.2.In the identification of diseases of apple leaves,through comparison,it was found that the model in the original spectrum using SPA to extract the characteristic value had a high accuracy generally,while RF was more suitable for the extraction of characteristic value in the derivative spectrum.In the four modeling methods,the principal component analysis-Mahalanobis distance(PCA-MD),support vector machine(SVM)and BP models had higher accuracy,up to 100%.3.Using fruit leaf spectral information and through first derivative(1st-Der)and Savitzky-Golay(SG)pretreatment methods,SPA and RF2 feature extraction methods and(multiple linear regression)(MLS),partial least square(PLS),orthogonal partial least square(O-PLS),SVM,BP modeling methods,the estimation model of P,Fe,B and Mn was established,and a comparison was made on the estimation effect of the model before and after adding timing characteristic.After adding the time-sequence characteristics,for element B,the model established with SG-lst-Der-SPA-MLR had a higher accuracy,where R=0.9365,RMSE=2.8597;for element Fe,the model established with SG-1st-Der-RF-BP had a higher accuracy,where R=0.8882,RMSE=13.2240;for element Mn,the model established with SG-lst-Der-SPA-BP,where R=0.8640,RMSE=13.3883 and for element P,the model established with 1st-Der-RF-BP had a higher accuracy,where R=0.8485,RMSE=0.0669.Through the comparison of the modeling processes and results of the four elements,it can be found that:(1)Random frog method and SPA feature extraction method showed different advantages and disadvantages in elemental estimation model.SPA method was suitable for most models,and it was not random,so the characteristic value of each extraction was fixed,but the time complexity was high.The advantage of random frog was that the time complexity was low and the running speed was fast However,the number of feature variables extracted by random frog method was preset according to the number of extracted SPA,otherwise the number of extracted characteristic wavelengths was large,and each characteristic value was not fixed.If the random frog method was is used for feature extraction,it is recommended to cooperate with the SPA method and carry out more debugging as needed to select the optimal combination of wavelength.(2)Through the comparison of spectrum after smoothing treatment and the derived spectral data after smoothing,it was found that the derivate spectrum should meet a larger window size when smoothing,otherwise,the SPA feature extraction was centered at about 400nm with more noises,and the number of the characteristic variables extracted with SPA of derivative spectrum is more than original spectrum,but with higher accuracy.Therefore when using derivative spectra to establish models,the smoothing treatment of data is very important,which should be paid attention to.(3)Compared with other elements,the accuracy of the B model was higher than that of other elements.However,it was found that the correlation between B and the time was also higher(fitted with polynomial,the coefficient of correlation between the content of element and the number of weeks was 0.8596).Therefore,the addition of times-sequence characteristics had a positive effect on improving the accuracy of the model.The correlation between the Mn and time was low(fitted with polynomial,the coefficient of correlation between the content of Mn and the number of weeks was 0.6949),and in individual Mn element models,the accuracy of no time-sequence characteristic was higher than that with time sequence.Most of the past researches on fruit tree spectrum were focused on a period,with few attention to the continuous change,and in this study,ASD FieldSpec 3 full band portable spectrum analyzer was used to dynamically collect different tree species of fruit tree flowering period,and the spectrum information of the apple leaves was measured every another week from spring shoot growth period to autumn shoot stop period,and meanwhile,the contents of four elements,P,Fe,B and Mn of leaves,as well as the spectral data of three different kinds of diseases and pests of apple leaves were determined,and the result was more universal and scientific;the identification models of three kinds of diseases and pests damages of apple leaves were established with near infrared technology and data mining technology,with very high modeling accuracy,which provided reference for the prevention and control of diseases and pests in orchard in the future;while establishing the estimation model of spectrum and mineral nutrition,time sequence characteristic was added,and the result of the study showed that the addition of time sequence characteristic improved the model accuracy,provided a new thought for the research of orchard in the future,and even the estimation of mineral nutrition elements of other plants;in the past studies,usually one modeling was used to compare the advantages and disadvantages of the pretreatment algorithm,which was easy to generate partiality.In this study,different pretreatment algorithms,feature extraction methods and modeling methods were compared for their respective advantages and disadvantages in different types of spectral information,which provided a theoretical basis for the deep study in the future.