Inversion Model Of Chlorophyll Content For Citrus Leaves Based On Hyperspectrum

Chlorophyll is the most important pigment in photosynthesis during citrus growth.The chlorophyll content of leaves directly affects the growth of citrus fruit trees.Traditional methods of obtaining chlorophyll content of citrus leaves are time-consuming procedures with complex operations which can be harmful to citrus trees.More over,traditional methods cannot meet the demand for rapid and non-destructive monitoring of potassium content in large-scale citrus orchards.Aiming at the problem of low precision and poor generalization ability of traditional chlorophyll spectral inversion model,a chlorophyll inversion model of citrus leaves based on data fusion and deep transfer neural networks(DTNN)is proposed.In this paper,sweet orange was used as the experimental variety,and healthy leaves were collected in the germination stage,stability stage,bloom stage and picking stage of citrus.The spectral reflectance of the sample was measured by spectrometer(ASD Field Spec 3),and then the chlorophyll content and SPAD value of the corresponding sample leaves were measured by traditional spectrophotometry and TYS-B chlorophyll analyzer,respectively,during four different growth periods corresponding to germination period,stability period,bloom period and picking period.Based on the spectral data and the measured data,the chlorophyll inversion model of citrus leaves was established.The main research contents are as follows:(1)In the feature wavelength and sensitive band extraction,Pearson correlation coefficient method,independent component analysis(ICA)method,manifold learning,and autoencoder(AE)were used to extract hyperspectral data features.The partial least squares regression(PLSR)model based on sparse autoencoder(SAE)with 16-dimension features for chlorophyll content prediction of citrus leaves was the best.The determination coefficients of the calibration set and the validation set model were 0.8712 and 0.8554,respectively,and the root mean square errors(RMSE)were 0.0266 and 0.0287,respectively.(2)On modeling,citrus was established based on single vegetation spectral index,mixed vegetation index,multiple linear regression(MLR),partial least squares regression(PLSR),support vector regression(SVR),nearest neighbor regression(NNR),convolutional neural network(CNN)and deep neural network(DNN).A deep transfer network was builded to merge and transfer traditional features and deep learning features for chlorophyll content prediction.(3)In order to study the influence of different spectral forms for modeling,the first order differential,second order differential,reciprocal and logarithmic transformation were carried out on the spectral data of citrus leaves in germination period,stability period,bloom period and picking period.(4)Particle swarm optimization(PSO),cross grid(CV)search optimization,genetic algorithm(GA)optimization were used for SVR parameters optimization,and gradient operator,dropout,batch normalization,L1 and L2 regularization were used for deep learning model optimizing.The model(ICA-PSO-SVR)based on SVR with ICA features extraction and PSO optimization method has the best performance,in which the determination coefficients of the calibration set and the validation set model were 0.7467 and 0.6831,respectively.The DNN model based on dropout optimization has the best performance.The determination coefficients of the calibration set and the verification set model are 0.8409 and 0.8273,respectively,and the root mean square error is 0.0329,0.0343,respectively.(5)In the fusion of traditional features,the linear modeling based on R-I-M features fusion were optimal,and the determination coefficient model of the germination period,stability period,bloom period and picking period were 0.8617,0.8092,0.8458,0.8900,respectively.In fusion of depth features,the fusion based on AE,SAE,DAE were optimal,the determination coefficients in germination period,stability period,bloom period and picking period correction set model were 0.9147,0.8320,0.9662,0.9613,respectively.The results show that the prediction values of inversion model of chlorophyll content for citrus leaves based on multi-feature fusion deep transfer neural networks are significantly correlated with SPAD values.This model can realize rapid,non-destructive and accurate measurement of chlorophyll content for citrus leaves,which provide a basis for citrus cultivation and management,variable spraying fertilizer,agricultural machine development,and may provide a theoretical basis for nutritional diagnosis and growth monitoring of fruit trees based on hyperspectral technology.