Research On Early Detection Of Citrus Brown Spot Disease Based On Hyperspectral Imaging

Citrus is susceptible to infection by Alternaria alternata,which causes Citrus Brown Spot(CBS).Failure to control this disease in a timely manner may lead to widespread infection.Early detection of CBS is beneficial for accurate prescription decisions and targeted comprehensive prevention and control.Although traditional molecular techniques and other physical methods can accurately detect CBS,the testing process is highly specialized,requires destructive sample processing,and is costly.Hyperspectral imaging technology can accurately perceive internal quality information of citrus in a non-contact manner.Therefore,this study combines hyperspectral imaging technology to develop a machine learning-based method for detecting the degree of CBS disease in immature citrus fruits(Tribute Citru),to achieve early diagnosis of CBS disease.The main contributions are as follows:(1)High-spectral data(387-1025 nm)was collected from Emperor mandarin fruit samples that were sound and inoculated with Colletotrichum acutatum at an immature stage.The region of interest(ROI)was identified and the diseased samples were classified into early-stage(CBS-E)and late-stage(CBS-L)based on the infection time of the pathogen.The dataset comprised of 330 sound,676 CBS-E and 668 CBS-L samples.Additionally,the performance of multiple scattering correction(MSC),standard normal variate(SNV)and standardization(SL)methods were evaluated for preprocessing the high-spectral data.The results showed that MSC significantly improved the inter-class differences of local high-spectral response curves of different fruit categories compared to other preprocessing algorithms.It was more effective in describing the high-spectral response characteristics of healthy and diseased citrus fruits at different stages and degrees of infection.(2)Two-stage CBS disease early detection method based on statistical machine learning.In order to reduce the difficulty of fitting during disease detection model training,the learning problem of classifying three types of patterns is decomposed into two concatenated binary classification learning tasks through the two-stage decision-making idea.First,the first-stage binary pattern discrimination modeling is performed on the sound citrus fruit samples and the diseased citrus fruit samples(including CBS-E and CBS-L classes)to filter out healthy fruit samples.Then,the second-stage binary pattern discrimination modeling is performed on the CBS-E and CBS-L two types of diseased samples to achieve the early stage discrimination of citrus fruit infected by CBS pathogenic bacteria.Based on the two-stage decision-making idea,competitive adaptive reweighted sampling(CARS)is used to extract feature spectral bands from the high-spectral data preprocessed by MSC,and two CBS disease discrimination models are established under the two decision stages combined with support vector machine(SVM).Experimental results show that this method can obtain more accurate CBS disease detection results,and the overall average accuracy rate reaches 86.16%,which is 1.59%higher than the conventional three-class classification decision model accuracy rate.In addition,the early detection accuracy of CBS disease in this method reached 79.82%,which is 5.73% higher than the conventional three-class classification decision model accuracy rate.(3)Early detection method of CBS disease based on deep learning.In order to reduce the redundant information in hyperspectral images and the number of model parameters of deep convolutional neural networks,the principal component analysis(PCA)is first used to reduce the dimensionality of hyperspectral data,and the top three principal component images with the highest cumulative contribution rate are extracted as the input of the deep convolutional neural network model.Furthermore,the early detection performance of CBS disease is evaluated using five deep learning algorithms: Le Net,Alex Net,Micro Net,G-Ghost Net,and Mobile Net.Experimental results show that combining the dimensionality-reduced hyperspectral images helps the deep learning model focus on local areas with smaller infected areas and milder degrees,which is more conducive to achieving accurate detection of CBS early-stage diseases.Thanks to the Dropout layer reducing the degree of overfitting during model training,Alex Net achieved the best early detection effect of CBS disease,with an overall average accuracy rate of 95.6%,which is 9.44% higher than the detection method based on statistical machine learning.In addition,the accuracy rate of Alex Net in detecting CBS-E disease reached 94.17%,which is 14.35% higher than the detection method based on statistical machine learning.