Nondestructive Detection Method Of Potassium Content In Grape Leaves Based On Image Processing And Deep Learning

In order to improve the yield and quality of crops and provide scientific basis for field cultivation management in time,rapid non-destructive testing of crop nutrient elements has gradually become a research hotspot.In the process of grape cultivation,factors such as water application,fertilizer application and light directly affect the growth and nutrient absorption of grape plants.The potassium content of grapes is closely related to flower bud differentiation,fruit maturity and sugar content.This paper combines image processing and deep learning technology to explore the non-destructive detection method of potassium content in red earth grape leaves.The main research contents are as follows:（1）The field orthogonal experiment was carried out with Gansu Red Globe grape as the research object based on the three factors of water application,fertilizer application and shading rate.The experiment was divided into 9 plots,and the experiment was repeated for 2 rounds.The optimal combination and the order of influence of factors are determined by orthogonal experimental data processing and range analysis.The analysis of the experimental results showed that there were obvious differences in the potassium content of grape plant leaves in different treatment groups;the range analysis showed that the optimal combination of the experiment was L₃M₃N₁,that is,T9 treatment.（2）Collect field experiment grape leaf image samples,use rotation,affine and intensity transformation to expand the sample data set;use matlab to extract image color components and their combination parameters,and use Origin to analyze the correlation between image feature values and leaf potassium content,Eliminate irrelevant image feature parameters,and filter out relevant image feature values with good robustness.The experimental results showed that the leaf color characteristic values B,NGI,R/（R+G-B）of different treatments in plots T1～T9 were significantly correlated with leaf potassium content（P<0.001）.（3）Using the optimal processing group of the orthogonal experiment to screen the characteristic value of the image,the multiple stepwise regression method was used to fit the multiple linear regression equation of the leaf potassium content and the characteristic value.（4）Select the training set and validation set of image samples,test and improve the convolutional neural network model to improve the prediction accuracy,and use the improved convolutional neural network CNN-M1 to train and verify the grape leaf potassium content detection experimental model.Experimental results show that the improved model has an accuracy rate of over 90%for potassium content detection.In addition,in order to test the accuracy of different potassium content prediction models,comparative experiments were carried out on the multiple regression equation fitted by image feature values,the original convolutional neural network training model,and the improved convolutional neural network training model.The experimental results show that the average prediction accuracy of the improved CNN-M1 model is higher than the multiple linear regression fitting equation and the original CNN training model.In this paper,an experimental method combining image processing technology and deep learning is used to research and obtain a non-destructive and rapid detection method for potassium content of red earth grape leaves,which provides a new technical means for modern precision horticultural agricultural informatization.