Research On Early Pest Detection Of Crops Based On Hyperspectral Imaging And Machine Learning

The production of crops is of vital importance to national economy and individual’s livelihood.As a severe threat to the crop production,insect pests may destroy the normal growth and development of crops,then affect the quality and yield of crops.Therefore,pest control is an important part of the crop production process.In order to improve the prevention and control effect of insect pests,farmers often overwhelmingly rely on pesticides and apply it to lessen pests blindly.The use of pesticides causes a series of bad results such as the residues of pesticide,the destruction of farmland soil and self-heating environments.Furthermore,pests would develop resistance due to pesticide abuse.Because of the above-mentioned factors,comprehensive pest management came into being.In this governance model,the combination of chemical,biological and other control methods are encouraged to prevent pests.And scientific and precise application of pesticides are required for improving the efficiency of pesticide control and achieving the sustainable pest control ultimately.It is of significance to early detect pests and judge the stage of their infestation in comprehensive pest management.However,pests have the habit of concealment and self-protection and it is difficult to be found by tnaked eyes.Therefore,this article explored the usage of hyperspectral imaging(HSI)technology for the early detection of borer and covert pests and investigated the employment of machine vision technique on complex pests identification.The specific research contents are as follows:(1)Visible/near infrared hyperspectral images of rice plants which infested by the rice stem borer were collected.Then continuous projection algorithm(SPA)was used to extract 17 characteristic wavelengths from hyperspectral data.Sixteen texture features were also extracted from the hyperspectral images which were transformed by principal component analysis using the gray level co-occurrence matrix.Finally,back propagation neural network model(BPNN)was established using full spectrum,characteristic wavelengths,texture features,and fusion features of spectrum and texture respectively.The results showed that the BPNN model based on the fusion features achieved the best performance with the accuracy of 95.10% in test set,which indicated that the visible / near infrared hyperspectral imaging technology could realize the early detection of rice borer infestation.(2)Visible/near-infrared hyperspectral images of pakchoi which infested by aphids were collected.Principal component analysis transformation and linear correlation analysis were used to analyze the data.The results showed that there existed certain correlation between the spectral characteristics and the stage of infringement.The four characteristic wavelength extraction methods,SPA,principal component load(PCA),variable combination cluster analysis(VCPA)and multi-cluster feature selection(MCFS),were compared.Three data analysis methods,linear discriminant analysis(LDA),non-linear BPNN,and support vector machine(SVM),were employed to construct corresponding models.The results showed that the SVM model based on characteristic wavelengths extracted by MCFS had the highest detection accuracy with the accuracy of 98.98% on test set.The results illustrated that the effectiveness of visible/near-infrared hyperspectral imaging technology on the early detection of pakchoi infested by peach aphid.(3)Visible/near-infrared hyperspectral images of tomato leaves which infested by Bemisia tabaci were collected.Four characteristic variable screening methods of SPA,VCPA,MCFS and PCA-loading are used and compared.The results showed that the SVM model based on the characteristic wavelengths extracted by SPA had the best detection accuracy of 86.21% on test set.Based on the wavelength images extracted by SPA,the texture features were extracted by GLCM and ANOVA was used to screen the most relevant entropy values.The spectral features and texture features were combined to construct the model,and the detection results showed that the performance of model was not significantly improved.This result indicated that the visible/near infrared hyperspectral imaging technique could be used to detect tomato which infected by Bemisia tabaci at an early stage,and the spectral features are more important than the texture features.(4)Pictures of cabbage worms on rape plants were collected to construct image datasets.Under the Pytorch deep learning framework,a U-Net network was used to perform pixel-level segmentation of cabbage worms on rape plants.The Dice score on the training set was 0.56.The results showed that U-Net could realize the positioning of Brassica napus.Compared with other traditional image segmentation methods,it was found that the traditional image segmentation method could not achieve the segmentation of pests under similar backgrounds.The results of this study proved that U-Net had a great advantage in segmenting pests with the similar backgrounds.The U-Net could achieve accurate positioning and preliminary identification of cabbage worms on rape plants.