The Research On The Forecasting Methods Of Rice Planthoppers Based On Image

Rice planthopper is a kind of important rice migratory pests,Nilaparvata Lugens,Sogatella furcifera and Laodelphax striatellus are the most common in rice area,which often gather on the rice plants for feeding and oviposition.The survey of population density of rice planthoppers is important to make forecasting decisions and efficient control.In most rice-growing countries,the survey methods of rice planthopper in paddy fields usually involve plant-flapping method,suction sampler,net-sweeping method and light trap method.The plant-flapping method is the most often used.The plant-flapping is affected by the density of insect,the growth period of rice and the moisture content of the inner wall of the disc and easily causes surveyor physical and visual fatigue,time-consuming work in paddy fields and low efficiency.Manual investigation only records the number of rice planthoppers types and the data of various stages and unable to confirm the validity of the data in future.Liu qing jie fully used digital image processing technology to study the effect of different features for planthoppers detection and gained good detection effects,but false detection rate is still higher.On this basis,the paper proposed a new three-layer detection,study the rice planthoppers detection rate,false detection and white-backed planthoppers classification(including long-winged adults,high-instar nymphs and low-instar nymphs)based on the different choice of image feature,and classifier model parameter,The main contributions of the paper include:(1)In the first layer detection,study different dimension of HOG features and different cascade layer number of Adaboost classifier on detection rate and false detection rate of white-backed planthopper in rice base.Firstly,we established white-backed planthoppers and non-planthoppers noise of positive and negative training sample set on the basis of 2012-2016 year collection image,Then,extracted the different dimension HOG features of the training sample,The HOG feature was used to train the different cascade Adaboost classifiers for the detection of rice planthoppers,choose the most optimal Adaboost classifier to detect 525 rice planthoppers images.then use the best Adaboost classifier to detect 525 white-backed planthopper in rice base.The results showed that the detection rate of white-backed planthopper in rice base was 90.7%,and the false detection rate was 56.2%.(2)In the second layer detection,for the existence of more mistaken noise in the first layer,we study different local image features to train SVM classifier on the non-planthopper noise recognition.These noise mainly includes water droplets,surface reflection,mud dots and rice leaves,they are quiet different from planthoppers in texture.Firstly,extract the Gabor and LBP features of the rice planthopper sample bank,use Z-score to normalize the feature,Then using Gabor,LBP and two feature fusion to train the SVM classifier,and the optimal texture feature was selected by different feature of training SVM classifier ROC curve.We find the Gabor and LBP texture feature of training SVM classifier for white-back planthoppers and non planthoppers noise recognition rate is high.Finally,use the SVM classifier to remove the non-planthopper noise from the 525first-level sub-images.The results showed that the false detection rate of the first layer was reduced from 56.2% to 10.2%.(3)In the third layer detection,we study the classification and identification of different white-backed planthopper.Using HOG feature combined with SVM classification algorithm to distinguish white-backed planthopper classification of different developmental stage.Firstly,We extracted the HOG feature of white-backed planthopper three kinds of insect long-winged adults,high-instar nymphs and low-instar nymphs in the paper,use Z-score to normalize the feature,Then use PCA and LDA method to reduce HOG feature dimension and compare different dimension reduction algorithms ability in different white-backed planthoppers(Sogatella furcifera).Finally,use the SVM classifier to classify white-backed planthoppers sub-image detected by 525 base of rice planthoppers image in the second layer.The results showed that the recognition rate of the white-backed long-winged adults(Sogatella furcifera),high-instar nymphs and low-instar nymphs was 93.2%,82.7%,86.9% respectively.According to the results of three layer test,the average recognition rate of the white-backed planthopper was 73.1%.The false detection rate was 23.3%.For the images without insects,the false detection rate was 5.6%.Thus,it is feasible to use the image processing method to forecast planthoppers on the base of rice.