Information Extraction And Identification Of Rice Blast In Cold Region Based On Multi-scale Remote Sensing Data

In recent years,with the development of spectral technology and remote sensing technology,using various types of remote sensing data for disease information extraction and monitoring has become an important and effective means for distinguishing and recognizing crop diseases and insect pests.Particularly,it as a new type of remote sensing technology for crop diseases and insect pests because of its high efficiency,flexibility,low cost and high resolution,the UAV make up for the shortcomings of satellite remote sensing data which not high precision due to the long range and the lack of real-time monitoring due to the long cycle of resetting,and provide good technical support for remote sensing monitoring at regional scale.However,how to choose the appropriate method,in the massive remote sensing data to extract useful information,for the identification of crop diseases and insect pests is an important issue.This paper comes from the National 863 Project(miniature unmanned aerial vehicle remote sensing information acquisition and crop nutrient management technology)and the Heilongjiang Province Major Science and Technology R&D Project "based on multi-scale remote sensing of agricultural comprehensive remote sensing monitoring technology research and application".This research,taking rice blast in northern China as the research object studied on the extraction and discrimination of spectral characteristics of rice blast and nitrogen-deficient rice from three scales of leaf,canopy and regional,based on indoor imaging hyperspectral,near-ground imaging hyperspectral and UAV aerial spectral data.The specific research contents are as follows:(1)At leaf level,on the basis of the comparison of the reflectance spectra of nitrogen deficiency,mild infection,severe infection and healthy rice leaves,different pretreatment methods were used to eliminate the noise.The PLS-DA models were established by partial least squares combined discriminant analysis and the PCA-SVM models were established by principal component analysis combined least squares support vector machine.The accuracy of the models was 98.4%,among them,the prediction accuracy of SNV-PLS-DA model was 100% and the classification result was the best.Then,using the successive projection algorithm to extract the characteristic wavelength combined with the discriminant analysis and the partial least squares support vector machine to establish the classification model,the prediction accuracy of the S.G-SPA-SVM model and the S.G-SPA-LDA model established with nine characteristic wavelengths were 98.4%.At the same time,in order to connect with canopy scale and regional scale,the vegetation index characteristics were constructed by using the vegetation index method.The vegetation index with significant differentiation ability of four types of rice leaves was selected by comparing the variance analysis,and the disease identification models based on vegetation index were established by stepwise discriminant analysis,support vector machine and neural network algorithm.The experimental results show that the discriminant analysis model with 12 vegetation indices achieved the best classification and discrimination effect,and the accuracy of cross validation and prediction was 96% and 96.9%,respectively.(2)Information extraction and identification of rice leaf blast at canopy level.Based on the characteristics analysis and discrimination of disease,nitrogen deficiency and healthy at leaf level,the spectral data of canopy leaf disease,nitrogen deficiency and healthy samples were extracted from rice canopy hyperspectral images obtained at jointing stage,and were processed by using Savitzky-Golay smoothing pretreatment method.After analyzing the average spectral characteristics of the three samples,the principal component analysis,vegetation index and continuous wavelet transform were used to extract sensitive features.The discriminant recognition models were established by using discriminant analysis,support vector machine and BP neural network classification algorithm and the model was cross-validated by "stay one method".In the models of the first five principal components extracted by PCA as the input vector,the cross validation results of LDA model and SVM models were 94.2%.On the basis of the obvious vegetation index of leaf level,the eight vegetation indices of canopy level were increased,and the models were established using 39 significant vegetation indices by comparison of variance analysis.Among them,the identification model of stepwise discriminant analysis algorithm only used EVI,m SR705,RVSI and PRI four vegetation index factors,and the accuracy of cross validation of canopy level health,nitrogen deficiency and rice blast was 94.2%.In addition,we also tried to extract the wavelet energy coefficient by continuous wavelet transform at the canopy level,and extracted 11 wavelet coefficients to establish the classification and recognition model.The classification accuracy of the wavelet coefficients with three algorithms is more than 94.2%.Among them,the classification result of linear discriminant analysis model was the best,and the overall cross validation classification accuracy was 97.1%.The results of the experiment showed that the CWA method was best among the three methods of extracting the characteristics for leaf blast,nitrogen deficiency and health of the canopy level,and the verification accuracy of the models of three algorithms based on wavelet features were more than 94.2%.The comparison of the three classification algorithms found that the discriminant analysis method has the best classification effect,and achieves the best classification effect in the classification model of different spectral characteristics.(3)Differentiation of neck blast at canopy level.Based on the comparison and analysis of the average spectrum of canopy neck blast samples and healthy samples after standardization and pretreatment,the spectral characteristics were extracted by PCA,CWA and vegetation index respectively.Then,combining discriminant analysis,support vector machine and neural network classification algorithm to establish the canopy level neck blast identification model: PCA method extracts the first 10 principal component based on the contribution rate of them to establish model,the LDA method and the SVM algorithm with RBF as the kernel function had the best classification result which the overall cross validation classification accuracy was 94.2%.Among the classification models based on 34 sensitive vegetation indices,the stepwise discriminant analysis model only using the two characteristic factors of GNDVI and DVI obtained the best classification result that was 94.2%,which provided the basis for the instrument development and extensive monitoring of neck blast disease.CWA method extracts 10 wavelet coefficients,and the overall classification accuracy of LDA and SVM model with linear kernel function was 94.2%.In the study of canopy level neck blast,the discriminant analysis classification algorithm(LDA or stepwise discriminant analysis)showed the best classification effect.Finally,the typical standard function of the stepwise discriminant model based on vegetation index was used to classify the canopy neck blast images.The classification accuracy and Kappa values of ISODATA classification algorithm were 90.19% and 0.7488,the classification effect was significantly higher than the use of a single significant vegetation index.(4)The differentiation of rice blast at regional scale was based on the multi-spectral data by the UAV.The regional neck blast model was established and validated using the block-level image obtained by six rotor unmanned aerial vehicle equipped with multi-spectral camera in 2016.According to the range of 6 channels covered by the multi-spectral camera,19 spectra characteristics with significant ability were selected from the vegetation index with significant difference for canopy neck blast and the reflectance spectrum of the 6 channels.Stepwise discriminant analysis method selected near red(800nm),NRI and GREEN / RED three characteristics to establish model,which was used to distinguish two fields of disease in 2015.The accuracy of sample recognition was 92.8% and the accuracy of image classification was 85.6%.Finally,the discriminant distinction model was applied to the large area image,from the whole plot,the classification accuracy rate could reach 80.8%.