Methods Of Millet Leaf Chlorophyll Content Prediction And Identification Of Varieties Based On Hyperspectral Technique

Millet,as the first of "five grains（millet,rice,wheat,soybean,sorghum）",is one of the largest crops in China.Millet is rich in vitamins,minerals and dietary fiber.Because of its high nutritional value,millet becomes a very popular miscellaneous crop.Millet is facing with many problems,such as low yield,imperfect fertilization system,adulteration in the sales process and so on.Therefore,it is necessary to realize accurate fertilization and nondestructive quality identification of millet by effective method.Hyperspectral imaging technology is widely used in crop nutrient inversion and variety quality identification of agricultural products due to its multi-band,high resolution and other characteristics.The research object of this study is millet,feature wavelength,vegetation index,trilateral parameters,peak and valley feature parameters,texture features and color features are extracted based on spectral and image information.Partial least squares regression（PLSR）and support vector machine（LS-SVM）are applied based on the correlation analysis with chlorophyll content.A hyperspectral monitoring model of chlorophyll content in millet at different growth stages was established based on the method of Attention-based convolutional neural network（Attention-CNN）.We took eight kinds of millet as the research object,the spectral and image features of single millet were extracted from hyperspectral images.The method of Attention-based convolutional neural network（Attention-CNN）and SVM are used to establish the millet atlas information and the identification model of foxtail millet varieties combined with the spectral and image information,which provided theoretical support for accurate management and on-line inspection of millet varieties.The main conclusions of the study are as follows:（1）The correlation between spectral reflectance and chlorophyll of millet leaves showed basically the same change trend in the three growth stages,and the correlation between spectrum and chlorophyll after pretreatment was improved to a certain extent through MSC.We can get conclusion as follow by correlation analysis,at jointing stage,the correlation between 526-594 nm and 697-729 nm is 0.7⁰.9;at booting stage,the correlation coefficient is 0.7⁰.9 in the range of 647-690 nm and 700-743nm;at grainfilling stage,the correlation coefficients were higher in 522-580 nm,658-682 nm and 702-741 nm.（2）The following vegetation indexes had the best universality,MTCI、NDVI、SAVI、SR713、CI_red-edge and OSAVI,were significantly related to chlorophyll content in millet leaves at the three growth stages.At jointing stage,PLSR based on vegetation index had the best model prediction effect,Rv² was 0.724,RMSEv was 1.066,RPD was 2.063.The following trilateral parameters have the best universality,Red edge position（γ_l）,kurtosis coefficient（Kur）,yellow edge amplitude（D_y）,yellow edge area（SD_y）,blue edge amplitude（D_b）and blue edge area（SDb）.The PLSR of millet jointing stage based on trilateral parameters had the best model prediction effect.The following characteristic parameters of peak and valley have the best universality,K_ge、K_gprv、A_α、A_β、K_ge/K_re、A_α/A_β、A_α/A_γ、A_β/A_γ.At booting stage,Rv² of PLSR model was 0.612,RMSEv was 1.135,RPD was 1.939,which indicated that PLSR model based on peak and valley characteristic parameters could predict chlorophyll accurately at booting stage.The following image features have the best universality,R（variance）,R-B（variance）,R-G（variance）,G-B（variance）,standard deviation and smoothness,the PLSR model of millet booting stage based on image features had the best prediction effect,Rv² was 0.504,RMSEv was 1.297,RPD was 1.697,which indicated that the PLSR model established by image features in the booting stage of millet could achieve rough inversion of chlorophyll content in leaves,but the prediction accuracy was low.（3）The characteristic wavelength was extracted based on SPA,CARS and CC-SPA to establish the PLSR prediction model of chlorophyll content in millet leaves at different growth stages was established.The results show that the CC-SPA-PLSR model has the best prediction effect for different growth stage.The characteristic wavelength and spectral characteristic parameters are combined on the basis of normalization,and the prediction models of PLS,LS-SVM and Attention-CNN are established.The results show that,for different growth stages,the reverse calculation accuracy of Attention-CNN model for chlorophyll in millet leaves was higher than the PLSR and LS-SVM models,of which,the Attention-CNN model at jointing stage has the best prediction effect.The coefficient of determination was the highest,with Rv² of 0.828;the root mean square error was the smallest,with RMSEv of 1.516;the relative analysis error was the largest,with RPD of 2.143;and the difference between Rc² and Rv² was the smallest,with a value of 0.008.This indicated that the Attention-CNN model has higher adaptability and stability to the sample than the traditional model,and has higher prediction accuracy than the traditional model in the regression predictionof chlorophyll content in millet leaves.Therefore,the Attention-CNN model is a high-performance prediction model of chlorophyll in millet leaves.（4）In the study of millet variety identification,the identification accuracy of SVM cultivar classification model based on reciprocal logarithmic spectral characteristic curve was 73.13%.The identification accuracy of SVM model using image and spectral information fusion achieved 77.50%,enhanced about 4.37% compared with the SVM model only using spectral information,and 16.25%compared with the SVM model only using image information.The minimum discrimination accuracy of millet varieties increased from 50% to 65%.The model established based on combination of image information and spectral information can improve the classification and identification of millet varieties.（5）In this study,Attention-CRNN model with attention mechanism was introduced in the identification of millet varieties.The overall accuracy of Attention-CRNN model was 87.50%,which is10% higher than SVM model.The minimum discrimination accuracy of millet varieties increased from65% to 90%.Attention-CRNN model had good feature extraction ability,improved the identification accuracy of millet varieties in general.The Attention-CRNN model shows great significance on the nondestructive identification of millet and possibly on other small grain varieties.