Modeling Of Near Infrared Nondestructive Testing For Citrus Huanglongbing

The citrus Huanglongbing is a lethality disease which is easy to spread.It seriously impacts the development of citrus industry in the world.While the precondition for effective prevention and control of Huanglongbing is fast and accurate diagnostics for Huanglongbing.Near infrared spectroscopy analysis technology is a kind of fast,simple and effective chemical analysis technology,and has been preliminarily researched in Huanglongbing detection.But,existing methods is mainly based on traditional chemical analysis technologies which is too simple and not suitable for different spectrometers and citrus varieties.To improve the robustness of algorithm on different spectrometers and citrus varieties,this paper reseaches the problem as follow:(1)Existing methods ignore the impact of different processing methods on multispectral point,this paper compare comprehensive evaluation,mean,center and random point respectly,and comprehensive evaluation is found better.Meanwhile,existing methods ignore the impact of data selecting with different spectral resolution,this paper explore the ability difference of spectral features decomposing organics based on spectral data with four spectral resolution.(2)To detect citrus Huanglongbing fast and accurately,this paper proposes Gabor-based Extreme Learning Machine with kernel(GKELM)to detect Huanglongbing.It combines the advantage of Gabor on feature processing with the fast learning speed and excellent generalization ability of extreme learning machine with kernel.Firstly,standard normal transform and principal component analysis are used to process spectral.Then,Gabor is used to further highlight key features.Extreme learning machine with kernel is used to classify.The results shows that GKELM can get high accuracy with low sensitive,and is robust for different spectral collection.(3)To improve the weak robustness due to citrus varieties,this paper proposes a Huanglongbing detection algorithm which merges multi-feature extraction and ensemble multi-classifier(KP-TEPS)based on stacking strategy of ensemble learing.This algorithm ensemble multi-classifier on feature layer and decision layer,and the diversity of single classifiers is used to enhance the generalization of model.The results shows that different metrics of ensemble classification model are outstanding.It means that KP-TEPS can utilize the diversity of ensemble learning to get batter detection results than single feature and single classifier.And it is also more robust to different citrus varieties.