Study On Classification Of Cotton Aphid Damage Level By Multispectrum Of UAV

Xinjiang,as the main cotton producing area in China,has a large cotton planting area,a variety of planting types and a long planting history,resulting in a large number of highquality raw materials.Because of the long growing period of cotton,it is easily invaded by cotton aphids,which causes huge economic losses every year.With the development of multispectral technology,the analysis and processing of UAV multispectral data can quickly and accurately identify the occurrence of cotton aphid.Provide technical support for precise pest control,variable pesticide application,increasing cotton yield and realizing modern cotton field management system.Based on this,this paper selects DJI unmanned elf P4 M as the data acquisition platform,and combines with the ground cotton aphid damage level investigation points.Extracting multi-spectral data features to establish classification and prediction models for machine learning and deep learning.The main research contents and results are as follows:（1）Classification of damage levels of cotton aphids based on multispectral data.According to the original spectral data and vegetation index,five models of K-nearest neighbor,logistic regression,Gaussian naive Bayes,decision tree and extreme gradient lifting tree after different preprocessing methods are constructed.Then,the K-nearest neighbor,logistic regression and Gaussian naive Bayes models are optimized by gradient lifting decision tree,genetic algorithm and principal component analysis.The best model is the Knearest neighbor model after convolution smoothing,multiple scattering correction and gradient lifting decision tree data preprocessing,and the accuracy of cross-validation reaches99.89%.（2）Identification of damage degree of cotton aphid based on multi-spectral data.The target label is converted into percentage of damage degree of cotton aphid,and the grade label is converted into percentage data.Gradient lifting decision tree was used to screen sensitive bands,and multiple linear regression prediction models and polynomial regression prediction models of cotton aphid severity in blue band,green band,red band,red band,near infrared band and NDVI vegetation index were established.（3）Study on segmentation of cotton aphid damage area by improving Unet network.The Unet image segmentation model based on Dice-Loss,cross-entropy loss function and FocalLoss is constructed by using the field survey results of cotton aphid damage level and fiveband multispectral image data as training characteristic variables and taking the cotton aphid damage level as the target value.The Unet model treated by Dice-Loss has the best prediction effect on the classification image of cotton aphids,reaching the overall classification accuracy of 84.04%.In this paper,aphids in cotton fields are taken as the research object,and based on the multi-spectral data of unmanned aerial vehicles,the damage classification of aphids is studied.The rapid identification and detection of cotton aphid damage in this area were realized.The research results can provide a method reference for the prevention and control of aphids in cotton fields,and provide a theoretical basis for the integration of intelligent UAV plant protection and precision agriculture remote sensing monitoring technology.