Research On The Distribution Of Chlorophyll Concentration Of Longan Leaves Based On Hyperspectrum

Photosynthesis is an important physiological index to measure the function of plant synthesis while chlorophyll is the material basis of plant photosynthesis,its content can often objectively reflect the plant's health status.Traditional chlorophyll content measurement methods include spectrophotometry,high performance liquid chromatography,etc.These chemical methods are not only time-consuming and laborious but also to blade grinding,damaging blade's structure,and the technical requirements of the operator are high.The traditional rapid detection method also includes the commonly used SPAD instrument,but the instrument cannot predict the distribution of chlorophyll content,and the hyperspectral feature"pattern integration"can exactly meet the disadvantages of"point detection".Therefore,it is essential for developing variable spraying equipment in precise agriculture and real-time monitoring of crops to measure the distribution of chlorophyll content in a rapid,non-destructive,comprehensive and accurate method instead of in accordance with a farmer's experience.Artificial intelligence is a popular research field,and deep learning is its core technology.In this research the performance of deep learning model in the prediction of chlorophyll content was discussed.In this paper,inversion models of chlorophyll content of Longan leaves in different maturity were established based on hyperspectral imaging system(HyperSIS,Zolix,China).The real value was obtained by spectrophotometry to improve the accuracy of prediction.The main contents were as follows:(1)In the correlation analysis of chlorophyll content based on spectral information and image information,Pearson's correlation coefficient was used to analyze the correlation between reflectivity and image texture features of each band and chlorophyll content.The numbers of bands at strong correlation with chlorophyll content for young leaves,mature leaves,more mature leaves and the leaves at 3 stages were 194,149,95 and 182,decreasing with maturity.Contrast,correlation,energy,and homogeneity were used to characterize the texture features of feature band images at 550nm,677nm and 703nm.It was found that the correlation between the contrast and chlorophyll content was negatively correlated in the three leaf growth stages and the whole growth period,but the correlation was weak;The correlation,energy and homogeneity and chlorophyll content were positively correlated in the three leaf growth stages and the whole growth period,and overall showed an increase in correlation with leaf growth.Energy and homogeneity showed a strong correlation with chlorophyll content(the correlation coefficient was higher than 0.6).(2)In the prediction of Longan chlorophyll content and its distribution based on SVR models,penalty vector c and the radial basis kernel function kernel parameters?of the SVM model were optimized by the particle swarm optimization algorithm,genetic algorithm and cross validation method.In the validation set,the PCA-PSO-SVR model had the highest determination coefficient(_cR~2)value of 0.8820,the lowest root mean square error(RMSE_c)value of 0.1793,the SVR penalty parameter c was 299.9535,and the radial basis kernel function parameter?was 0.0032.In the prediction of chlorophyll content based on the spectral feature parameters,the regression model established by the combination of the minimum reflectivity position of the absorption bands and the total reflectivity of the absorption bands performed best(_cR~2=0.8568,RMSE_c=0.2195,_vR~2=0.7712,RMSE_v=0.2862).(3)In the prediction of Longan chlorophyll content and its distribution based on 3 deep learning models,convolutional neural network,sparse autoencoder and deep belief network,the sparse automatic encoder model based on the fusion of spectroscopy and texture features had the best prediction effect.When the leaf samples from three growth stages were used as validation sets,the standard deviation of_vR~2 was 0.0125,which were 66.22%and 68.83%lower than the standard deviation of_vR~2 of convolutional neural network and deep belief network,respectively.This showed that the sparse automatic encoder had high prediction accuracy and stability.The prediction results of the optimal model GRNN in the traditional model were:_cR~2=0.9397,RMSE_c=0.2201,_vR~2=0.8962,RMSE_v=0.3339.The best model for deep learning SAE was:_cR~2=0.9212,RMSE_c=0.1918,_vR~2=0.9012,RMSE_v=0.2115.