Selection Of Hyperspectral Features And Model Construction Of Wheat Stripe Rust At Different Scales

Hyperspectral remote sensing data has the advantages of high spectral resolution and rich information.It is widely used in the remote sensing monitoring of wheat stripe rust,but there are also the problems of data redundancy and high correlation between bands.Therefore,it is an important link to select spectral features from numerous variables.The feature selection method used in the current research is mainly based on one-time modeling,the data processing is easily affected by the number of samples,and it is difficult to achieve the best results when modeling.To solve this problem,the feature selection method based on model population analysis(MPA)in this study was used to optimize the spectral features sensitive to wheat stripe rust at leaf and canopy scales.This method makes up for the shortcomings of one-time modeling idea,can make full use of sample information,generate a large number of sub-models by random sampling for statistical analysis and extract information from the data.And then,the selected features were used to construct a remote sensing monitoring model for wheat stripe rust,and the effectiveness of the MPA feature selection method was evaluated,in order to provide a reference for the hyperspectral data to select feature variables for remote sensing monitoring of diseases and insect pests.The specific research contents and conclusions are as follows:(1)At the leaf scale,the spectral response characteristics of wheat stripe rust and the spectral characteristics suitable for disease monitoring were analyzed.The traditional correlation coefficient(CC)method,independent T-test(T-test)method and subwindow permutation analysis(SPA)and random frog(RF)feature selection algorithms based on MPA were used to select the original bands and spectral indices as the sensitive features of stripe rust on wheat leaves,then,the fisher linear discrimination analysis(FLDA)disease severity identification model was constructed and compared.The results show that the accuracy of the SPA-FLD A and RF-FLDA models constructed based on the optimized feature variables of the SPA and RF algorithms is higher than that of the CC and T-test methods,and the number of preferred feature variables is also greatly reduced and the operation efficiency is improved.Among them,the model established by the spectral reflectance band and spectral index optimized by the RF algorithm has the smallest difference in accuracy,and the discrimination accuracy are more than 80%,which realizes the purpose of distinguishing the disease degree of wheat stripe rust.It can be considered that RF algorithm is the best method to optimize the sensitive features of wheat leaf stripe rust.(2)At the canopy scale,the spectral response of wheat under different disease severity of stripe rust was analyzed.By comparing with the leaf scale,it was found that there were differences in the reflectance curves of the two in the near-infrared band,and the correlation between the spectral index features and the disease index(DI)also showed different effects.CC method and three feature selection algorithms of competitive adaptive reweighted sampling(CARS),variable combination population analysis(VCPA)and bootstrapping soft shrinkage(BOSS)based on MPA were also used to extract the spectral features sensitive to wheat stripe rust.The DI estimation model of wheat stripe rust was constructed by partial least squares regression(PLSR)algorithm.The results showed that the accuracy of the estimation model based on the feature variables selected by MPA algorithm was improved to varying degrees compared with the full-band and CC method,and the number of feature variables was greatly reduced.The prediction results based on the BOSS-PLSR model are the best,and the RPD are all greater than 2,followed by the VCPA-PLSR model.When CC method and MPA algorithm combined to select feature variables,both CC-CARS and CC-VCPA showed better joint effect and improved accuracy,among which CC-VCPA-PLSR model performed best,indicating that the joint method is an effective and feasible feature variable selection method.(3)Solar-induced chlorophyll fluorescence(SIF)contains very rich photosynthesis information,and photochemical reflectance index(PRI)can sensitively capture the changes of non-photochemical quenching(NPQ)under external stress conditions,but both are disturbed by crop biomass at the canopy scale.In order to weaken its impact,the spectral index SISP was constructed by combining the advantages of reflectance spectroscopy in biochemical parameters of crops and the advantages of canopy SIF and PRI sensitivity to photosynthetic physiology,and its effectiveness in monitoring wheat stripe rust was evaluated.The results showed that compared with the stripe rust monitoring model constructed by the traditional reflectance spectral index,the accuracy of the model constructed by the SISP index is higher.The accuracies of the PLSR,multiple linear regression(MLR)and radial basis function neural network(RBFN)models constructed with SISP and reflectance spectral index as independent variables were all higher than those constructed with reflectance spectral index only.It can be seen that the SISP index can significantly improve the monitoring accuracy of wheat stripe rust,and the RBFN model has the highest accuracy among the three modeling methods.