Apple Brown Spot Disease Detection Model Based On Convolutional Neural Network And Imaging Hyperspectrum

Apple Marssonina Blotch(AMB)is a major disease that affects the leaves of apple trees,leading to a reduction in fruit yield and quality,which has a detrimental effect on the growth and development of apple trees.With the development of deep learning technology,8U neu ral networks can now be used to identify and classify crop diseases,and combined with imag ing hyperspectral technology,the problem that the training images do not fully reflect the ch aracteristics of the disease can be solved,thus achieving accurate and rapid detection of the e xtent of apple leaf brown spot disease.This study uses apple leaf imaging hyperspectral as the data source and leaf brown spot d isease detection as the research objective.Firstly,we obtain the region of interest and spectra l information of apple leaves,and analyse the spectral response characteristics of healthy and brown spot-infested leaves;after that,we extract the spectral characteristics of leaves using photochemical vegetation index and quadratic principal component analysis methods,and fo rm three types of data sets,namely RGB images,photochemical vegetation index images an d quadratic principal component images,using the synthetic images of RGB main band of i maging spectra as the control;the above three types of images are Alex Net,VGG16 and Mo bile Net V3 convolutional neural network models were constructed as training input data.The best model for detecting the degree of apple leaf brown spot disease was obtained by analyz ing the training results of different disease spectral feature extraction methods combined wit h different convolutional neural networks.The main findings are as follows:(1)Spectral information of brown spot disease leaves and regions of interest were obtain edThe spectral reflectance of the leaf and the background was calculated,760 nm was selec ted as the threshold value,the background was removed by binarisation mask,the spectral i mage of a single apple leaf was obtained,the ROI of the disease area was extracted,and the apple brown spot disease level was classified.The analysis of the raw spectra of healthy and brown spot diseased leaves showed that the trends of the spectral curves of the healthy and b rown spot diseased samples were generally similar,and the spectral reflectance of the brown spot diseased leaves was always higher than that of the healthy leaves after 633 nm,and the difference between the two fluctuated more obviously.This is due to the fact that the reflecta nce in this region is obtained by the internal tissues and cells of the leaf reflecting many time s,and the internal cell tissues and structures of the leaf are damaged after the leaf is infested with the disease,which leads to an increase in the spectral reflectance.(2)Extraction of spectral characteristics of apple leaf brown spot diseaseOn the one hand,the various vegetation indices of the samples were calculated according to the spectral reflectance,and it was found that the photochemical vegetation indices had th e highest correlation with the spectral data,with R~2reaching 0.859 on average,which could r eflect the disease information of the leaves to a certain extent;on the other hand,the seconda ry principal components were extracted from the spectral data of the leaves for which the pri ncipal component features had been extracted,and the cumulative variance contribution of t he first three types of secondary principal components could finally reach On the other hand,the secondary principal components extracted from the leaf spectral data with extracted prin cipal component features,the cumulative variance contribution of the first three types of sec ondary principal components can finally reach 100%,which can completely contain the mai n spectral information of the leaves.The RGB primary band synthetic images were used as c ontrols,together with photochemical vegetation index(PRI)images and secondary principal component analysis(PCA-PCA)images as datasets,and brought into the convolutional neur al network model for subsequent training,respectively.(3)A convolutional neural network-based model for the diagnosis of apple leaf brown sp ot disease extent was constructedThe RGB primary band composite images,photochemical vegetation index images and s econdary principal component images were used as training data and input into Alex Net,VG G16 and Mobile Net V3 convolutional neural networks to build a convolutional neural networ k model to identify the degree of brown spot disease on apple leaves by learning and extracti ng features through deep learning models.The training results of the models based on Mobil e Net V3 were the best,with the highest accuracy of 92.66%for PCA-PCA-Mobile Net V3;am ong the training results of the models based on VGG16 networks,PCA-PCA-VGG16 based on secondary principal component images was also the best,with an average accuracy of 92.41%;the training results of the models based on Alex Net networks were the worst,with the average accuracy of 92.41%for RGB-PCA-VGG16.The average accuracy of the training ef fect of the test set of Alex Net was only 74.24%.By comparing the model detection results u nder different combinations of spectral feature extraction methods and convolutional neural networks,it was concluded that the PCA-PCA-Mobile Net V3 model combining secondary pr incipal component images and Mobile Net V3 had the best ability to identify and classify the degree of brown spot disease on apple leaves.