Research On Detection Methods Of Rice Blast Based On Hyperspectral Imaging Technology

Pests and diseases are the main disasters in agricultural production and the main bottlenecks to agricultural development.Real-time monitoring of crops and rapid detection of pests and diseases are the key to ensuring the healthy and sustainable development of agriculture,as well as the technical basis of precision agriculture and agricultural remote sensing.Traditional crop monitoring and detection methods are costly,inefficient and not timely enough to meet the needs of modern agricultural development.Spectral detection techniques,especially hyperspectral detection technology that can obtain both spectral information and image information,which have emerged in recent years have provided efficient,nondestructive and green new technical means for the rapid real-time acquisition of crop growth information.Rice is one of the most important food crops for human beings,rice blast is recognized as one of the most serious diseases affecting rice in the world.There are many different types of rice blast,depending on the time and location of the disease,but the most damaging are leaf blast and neck blast,which,if not properly controlled,can lead to reduced yields and even crop failure.Grain blast cannot be ignored either,as it not only affects yield and rice quality,but also poses a direct threat to seed safety.Therefore,it is important to achieve rapid detection of various types of rice blast disease.Chlorophyll content is an important biomass indicator of crop growth,which can effectively characterize the growth status of crops.Rapid estimation of chlorophyll content of rice leaves under the condition of rice blast is of great significance for monitoring and evaluating the impact of rice blast on rice growth.This paper studied the identification of leaf blast,neck blast and grain blast,which are the more harmful diseases of rice blast,and the estimation of leaf SPAD of rice blast using hyperspectral detection technology in a field of northern China,in order to provide theoretical support for the qualitative and quantitative detection of rice blast and to provide a research basis for accurate and intelligent rice blast control.This paper focuses on the following.(1)Early grading detection of leaf blast was studied and a hyperspectral detection model was constructed for grade 0(healthy leaves),grade 1(infected without lesions),grade 2(lesion area<10%),and grade 3(lesion area<25%)rice leaves.In order to achieve the identification of infected leaves without visible lesion at the beginning of natural infection,this study proposed to take hyperspectral data of lesion-free areas adjacent to the lesioned areas on the infected leaves as grade1 samples.Spectral data from four grades of leaves were analyzed for comparison and preprocessed in five ways.On the basis of proving the clustering effect of both the original spectra and the five pretreatment spectra by clustering analysis,each spectrum was used as input constructing an SVM detection model with the highest accuracy is the SNV-SVM model with Linear as a kernel function and prediction accuracy was 96.53%.On the basis of proving the clustering effect of both the original spectra and the five preprocessed spectra by cluster analysis,the SVM detection model was constructed with each spectrum as input,respectively.The SNV-SVM model with Linear as the kernel function had the highest accuracy,and the prediction accuracy is 96.53%.PCA was used to extract feature variables from six spectral data and construct LAD and SVM detection models.Among LAD and SVM models,SG-PCA-LDA model and SG-PCA-SVM model with linear kernel function had the highest accuracy,with the prediction accuracy of 93.75%and 95.14%respectively.The CARS algorithm was applied to select the characteristic wavelengths and construct LAD and SVM detection models for the six spectral data,respectively.Among the LAD models,MSC-CARS-LDA had the highest prediction accuracy and SG-CARS-SVM was optimal among the SVM models.SG-CARS-SVM had the highest accuracy among all models,with a prediction accuracy of 97.22%and Kappa coefficients was 0.9633.To further simplify the model,it was proposed to perform PCA dimension reduction of the SG-CARS selected characteristic bands and construct the SG-CARS-PCA-SVM model with a prediction accuracy of 96.61%,slightly lower than SG-CARS-SVM,but the input number of variables was reduced from 21 to 6,a 71.43%reduction from SG-CARS-SVM,further reducing the model’s complexity,improving the operation speed of the model.Therefore,the SG-CARS-PCA-SVM model is comprehensively evaluated as the optimal model with the accuracy at all levels as follows 97.30%,94.87%,94.29%,100.00%.The accuracy of detection of grade 1 samples was comparable to other levels,with good identification.(2)The texture feature data of hyperspectral image was extracted to realize the recognition of healthy,mildly and severely infected panicles.The dimension of hyperspectral images of all rice panicle samples was reduced by PCA,The first three PC images with clear information expression were extracted as feature images.Texture information was extracted from feature images by probabilistic statistical filtering and second-order probabilistic statistical filtering respectively,and the extracted texture information was used as input to construct LAD and SVM models.The optimal model was based on the second-order probabilistic statistical filtering of texture information(i.e.,the GLCM of the feature band image)with a kernel function of Linear’s SVM model had the highest accuracy with an overall accuracy of 93.98%.According to the PCA loadings,the feature band images corresponding to the spectral feature bands were selected,and the texture feature were extracted to build LAD and SVM models.The selected characteristic bands are 498.712 nm,546.127nm and 685.410 nm.The SVM model based on the feature band image GLCM with Linear as kernel function had the highest accuracy and the overall accuracy of the model was 92.77%.The input variables for the model were 24 texture feature parameters from three feature band images.The study proposed to further simplify the model by using the SPA algorithm to downscale the GLCM parameters of the feature band images and build SVM model.The modeling results showed that the input variables after dimensionality reduction were simplified into four characteristic parameters:average value,contrast,correlation of 498.712 nm band image,and contrast of 546.127 nm band image.The overall accuracy of GLCM-SPA-SVM model was 91.57%,the accuracy was slightly reduced,but the number of input variables was reduced from 24 to 4,a reduction of 83.3%,the model was simpler,which was more conducive to the development of portable,online and other rapid monitoring and detection equipment.(3)Machine learning combined with hyperspectral data was used to construct a grain blast detection model for the identification of four types of in-ear grains:mature healthy,immature healthy,mature infected and immature infected.The PLS-DA and SVM algorithms were used to build grain blast identification models using raw spectra and pre-processed spectra as inputs,respectively.In both the PLS-DA and SVM models,the optimal model were the ones corresponding to the SNV preprocessing spectrum,i.e.,SNV-PLS-DA and SNV-PLS-SVM models.The modeling accuracy of the two models was 95.39%,94.08%,respectively,and the prediction accuracy was94.74%.BPNN and CNN algorithms were used to construct the models with SNV pre-processing spectra as input,and the"interval rotation"method was designed to realize the CNN modeling building sample amplification.The results show that the overall recognition accuracy of the BPNN model was 94.74%as good as that of the SVM model,and the Kappa coefficient was 0.9292.The accuracy of CNN model was the highest,and the overall accuracy was 96.05%without sample amplification.After using the"interval rotation"method of sample amplification,the accuracy was further improved,and the overall accuracy of the prediction reached 98.83%,kappa coefficient was0.9843.The accuracy of the model was very high and the recognition effect was very good.The results show that grain blast detection can be achieved with good accuracy in combination with machine learning,and the accuracy of the CNN model after sample amplification can be further improved.(4)Taking the estimation of SPAD value of rice leaves as an example,a quantitative detection model of rice leaf biomass index under the condition of rice blast was established.PLSR,SVM and BPNN algorithms were used to establish the hyperspectral estimation model of leaf SPAD values under rice blast using the leaf hyperspectral feature parameters,the feature wavelengths selected by CARS and the feature variables extracted by PCA as inputs,respectively.The results show that in the nonlinear modeling approach,the accuracy of the model with hyperspectral characteristic parameters and PCA-specific feature variables as input was higher than the model with CARS-selected characteristic variables as input,while in the linear modeling approach,the model with CARS-selected characteristic bands as input had the highest accuracy.The nonlinear modeling approach,especially BPNN,was more suitable for estimating the SPAD values of rice leaves under rice blast than the linear modeling approach.The optimal model was PCA-BPNN.The R_C~2and R_p~2 of modeling set and prediction set were 0.8712 and 0.8082,respectively,the RE of prediction set was4.18%.It can be used to estimate the SPAD value of rice leaves under rice blast.