Estimation Of Anthocyanin Content Of Apple Leaves And Identification Of Mosaic Disease Based On Hyperspectral Data

Apples are rich in various minerals and vitamins,and are known as the"king of the world’s four major fruits."Shaanxi Province,due to its superior natural conditions,has become a natural choice for high-quality fruit production,and is the world’s largest concentrated and contiguous planting base of high-quality apples and the only national wholesale market for apples in China.However,Shaanxi’s orchards are plagued by mosaic disease.Mosaic disease is a viral disease that causes leaf withering and yield reduction,and has a long infection cycle and significant impact.The prevention and control of mosaic disease can only be achieved by early detection and treatment.Anthocyanin(Anth)is an indicator for determining the severity of apple mosaic disease and can be used to monitor the health of apples.Estimating anthocyanin content through remote sensing technology can achieve rapid and large-scale identification and monitoring of apple mosaic disease.In this paper,non-imaging hyperspectral data and imaging hyperspectral data of apple leaves during the high incidence period of mosaic disease were obtained,and the content of anthocyanin in the leaves was measured.The spectral characteristics and anthocyanin content changes of leaves with different degrees of mosaic disease were analyzed.Through hyperspectral dimensionality reduction,feature selection bands were obtained,and hyperspectral indices highly sensitive to anthocyanin content were constructed to build an estimation model of anthocyanin content.Based on the estimation results of anthocyanin content,the identification and image inversion of apple mosaic disease were realized.In addition,to improve the simplicity of the model,this paper integrated variable importance in projection(VIP),partial least squares regression(PLSR),and Akaike information criterion(AIC)to select the optimal PLSR model,obtain the optimal independent variables of the model,and proposed a sparrow search algorithm-based random forest algorithm(SSA-RF)to improve the accuracy of the estimation model.This paper constructed an apple mosaic disease anthocyanin content estimation model at the leaf scale,which can provide a reference for research at other scales and provide a theoretical basis for achieving large-scale,efficient,and accurate estimation of anthocyanin and monitoring of apple mosaic disease through remote sensing technology.Results showed that:(1)With the severity of leaf disease increasing,the content of anthocyanins gradually increased,and the reflectance of leaf spectrum in the visible light band also significantly increased,with the slope of the red edge significantly increased.Spectral transformation processing highlighted the spectral features of diseased leaves:the first-order differential spectrum highlights the"red edge shift"phenomenon that occured in diseased leaves,the second-order differential spectrum enlarged the difference interval between healthy leaves and diseased leaves,and the reciprocal logarithmic transformation enhanced the relative reflectance value of the spectrum,further highlighting the spectral features of diseased leaves.(2)The original data of non-imaging hyperspectral data and imaging hyperspectral data,as well as three types of spectral transformation data,were significantly correlated with anthocyanin content.The non-imaging hyperspectral data had a strong correlation with the content of anthocyanin.The highest correlation coefficient between non-imaging hyperspectral data and anthocyanin content was 0.831,and the highest correlation coefficient between imaging hyperspectral data and anthocyanin content was 0.623.Spectral transformation improved the correlation between spectral data and anthocyanin content,which was beneficial for improving the accuracy of estimation models.The partial least squares model of non-imaging hyperspectral data had the most significant improvement,and the second-order differential spectrum model reduced the RMSE of the original spectral model by 5.17%and increased the R~2 by 4.28%.(3)The characteristic bands and optimal spectral indices showed a certain potential in estimating anthocyanin content.The successive projections algorithm based on partial least squares algorithm had a good dimensionality reduction effect.The dimensionality reduction ratio was as high as 98.34%for non-imaging hyperspectral data and 98.67%for imaging hyperspectral data.The characteristic bands were mostly located at the characteristic positions of the spectral curve and preserved the spectral feature information.Spectral index screening effectively reduced the number of parameters,17 optimal spectral indices were selected from 23 spectral indices for non-imaging hyperspectral data,and 16 were selected for imaging hyperspectral data.The estimation of anthocyanin content based on characteristic bands and optimal spectral indices both obtained good estimation results,with the best estimations of RMSE being 0.022 and 0.023,and R~2 being 0.955 and 0.0951.(4)The sparrow search algorithm had an improving effect on the random forest model.In the modeling of non-imaging hyperspectral data,the model constructed based on the optimal spectral indices had the most significant improvement.Compared with the traditional RF model,SSA-RF reduced the RMSE of the modeling set by 31.25%and increased R~2 by 5.18%;in the modeling of imaging hyperspectral data,the model constructed based on the optimal spectral indices also improved significantly,with the RMSE of the modeling set reduced by 22.12%and the R~2 increased by 3.56%.(5)Estimation of apple mosaic disease based on anthocyanin content achieved high recognition accuracy.The SSA-RF model constructed based on the optimal spectral index obtained the highest recognition accuracy in both non-imaging and imaging hyperspectral data,with overall accuracies of 96.369%and 90.246%,and Kappa coefficients of 0.926 and0.845,respectively.Differential processing and improvements to the SSA algorithm for the RF model helped improve the recognition accuracy of apple mosaic disease.Moreover,inverting apple mosaic disease images based on anthocyanin content is feasible.The inverted images not only revealed the size of the affected area,but also indicated the degree of the disease.This provides the possibility for large-scale remote sensing monitoring of crop pests and diseases in the future.