Application Of Raman Spectroscopy Combined With Machine Learning In Crop Disease Detection

Crop diseases result in 20-40% of agricultural loss every year worldwide.Timely detection of crop diseases can effectively prevent the development and spread of diseases and ensure the agricultural yield.Currently,traditional detection techniques are costly,timeconsuming and destructive.Therefore,rapid,accurate and non-destructive methods are in great demand.Raman spectroscopy(RS)is a non-invasive,non-destructive and label-free analytical technique that can provide vibration and rotation related to structure of the analyte.Obtaining information of crop diseases exactly in early stage and taking timely measures are of great significance to ensure agricultural production.In this paper,a variety of apple diseases and bruises were identified,and distinctions between infection degrees of apple Rhizopus rot were explored using RS combined with cascade forest.Then,RS combined with improved Inception network was used for determination of wheat kernels infected with Fusarium head blight(FHB).The research contents were as follows:(1)RS combined with cascade forest were applied to achieve identification of healthy,bruised,Rhizopus-infected and Botrytis-infected apples,and to distinguish infection degrees of apple Rhizopus rot.First,apple spectra were preprocessed using Savitzky-Golay(SG)smoothing,normalization and principal component analysis(PCA).Then,the structure based on cascade forest was designed to compare performance with support vector machine(SVM).In the identification of apple disease species,cascade forest with full spectra obtained the best prediction accuracy of 92.80%.The prediction accuracy of PCA-cascade forest reached91.56%.Although identification result was inferior to that of full spectra,time efficiency was increased by 60%.Finally,the dataset expanded by convolutional autoencoder performed best and prediction accuracy of cascade forest reached 82.35% in detection of infection degrees of apple Rhizopus rot.The experimental results showed that RS could not only be used to detect the disease species and bruises,but also distinguish infection degrees of Rhizopus rot on the intact apple directly.Meanwhile,data augmentation could improve the ability of cascade forest to discriminate infection degrees effectively.(2)RS and improved Inception network were used for determination of FHB-infected wheat kernels.First,a hand-held Raman spectrometer was used to collect Raman spectra of healthy,mildly and severely infected kernels,and spectral changes and band attribution were analyzed.Then,the Inception network was improved by residual and channel attention modules to develop the recognition models of FHB infection.The Inception-attention network produced the best determination with accuracy in training set,validation set and prediction set of 97.13%,91.49% and 93.62% among all models.Finally,the average feature maps of the channels visualized the important information in feature extraction to clarify the decisionmaking strategy.Meanwhile,the prediction accuracy of Inception-attention network based on characteristic wavebands reached 90.42%.Overall,RS and Inception-attention network provide a rapid,accurate and noninvasive determination of FHB-infected wheat kernels and are expected to be applied to pathogens or diseases in various crops.In conclusion,RS combined with machine learning can achieve rapid,accurate and nondestructive detection of species and infection degrees of crop diseases.This study also provides useful practical basis and theoretical support for identification,assessment and prevention of other crop diseases.