| China is the world’s largest producer of pears,but it is threatened by various diseases during the growth of pears.According to incomplete statistics,there already have been more than 80 kinds of pear diseases,the most typical of which are black spots disease,anthrax,rust,etc.It will not only affect the quality of pear fruit,but also the large-scale outbreak of diseases will cause unexpected economic losses.The traditional disease detection technology is slow and could cause large error.The physiological structure of the pear is damaged.Hyperspectral technology can quickly,accurately and non-destructively detect the pear disease types,so that targeted disease prevention measures can be taken in time to reduce economic losses.In this paper,taking the leaves of Dangshan Pear as the research object,the hyperspectral images of normal leaves,anthracnose leaves and black spot leaves of Dangshansu pear were collected by hyperspectral imaging system.Extract average spectral reflectance of the images and perform principal component analysis on images with ENVI 4.7 software.Build a classification recognition model based on software such as MATLAB R2015.The main results of this study are:First,application of hyperspectral image preprocessing methods.In order to improve the signal-to-noise ratio of spectral data,multipcative scattering correction(MSC),S-G(Savitzky-Golay)and standard normalized variate(SNV)are used to preprocess the original spectral datas.The MSC can eliminate baseline shifts and offsets that occur between samples.S-G convolution smoothing could remove high-frequency noise in spectral datas and preserve useful low-frequency information.SNV can eliminate spectral errors due to scattering in the sample.Second,the selection of characteristic bands.There are hundreds of spectral variables in the original hyperspectral datas.And there is a strong correlation between the bands and the bands.A lot of information is repetitive and useless.The use of the original spectrum to establish a classification recognition model has a large amount of datas and could be computationally complex.In this experiment,principal component analysis(PCA),succesive projection algorithm(SPA),uninformation variable elimination(UVE),competitive adaptive reweighed sampling(CARS)and random frog(SFLA)were used to extract characteristic bands.The original 339 bands could be compressed into tens or even dozens of bands,which greatly reduced the input of data modeling and improve the efficiency of the model.Third,the establishment of disease identification model.The classification model was established by SVM and BP neural network in supervised classification.The accuracy of the PCA-SVM model set is 93.17%,the accuracy of the test set is 90.83%,the mean square error is 0.2051 and the correlation coefficient is 0.8254.The accuracy of the CARS-SVM model set is 91.43%.,the accuracy of test set is 84.17%,the mean square error is 0.1583 and the correlation coefficient is 0.796.The accuracy of the SAP-SVM model set is 94.80%.,the accuracy of test set is 93.25%,the mean square error is 0.1547 and the correlation coefficient is 0.8847.The accuracy of the UVE-SVM model set is 87.14%,the accuracy of the test set is 86.23%,the maen squared error is 0.1750 and the correlation coefficient is 0.7960.The accuracy of the SFLA-SVM model set is 85.17%.The accuracy of test set is 84.26%,the mean square error is 0.2417 and the correlation coefficient is0.6794.The accuracy of the full-band model set is 87.42% and the accuracy of the test set is 87.37%.The accuracy of classification recognition model established by BP neural network is 84.16%.The results prove that hyperspectral techniques can identify melasma and anthracnose in Dangshansu Pear. |
