Withered Branches Grade Monitoring In Malus Sieversii Forest Using Hyperspectral Data

The Malus sieversii is an ancient and rare plant species in the temperate broad-leaved forests of the Tertiary period,and numerous apple varieties around the world contain the genes of the Malus sieversii.In recent years,it has been seriously threatened by insect pests,with outbreaks of small giddyup causing widespread dieback of Malus sieversii branches,extremely slow growth and serious degradation of the wild fruit forest ecosystem to the degree of extinction.Therefore,effective disaster monitoring and prevention measures are urgently needed.To address the current problem of ineffective monitoring of pests and diseases in wild fruit forests,especially in Malus sieversii forests,this study takes the typical distribution area of Malus sieversii in Gongliu County,Xinjiang as the research area,and applies three methods,including spectral feature analysis,extraction and discrimination methods,shallow machine learning methods and deep learning,to investigate the method of monitoring the health condition of Malus sieversii forests by using hyperspectral data from unmanned aerial vehicles,and proposes a remote sensing monitoring program for the litter grade of the Malus Sieversii forest under unsupervised conditions.The main findings of the study are as follows.(1)The high-resolution CCD data of unmanned aerial vehicle(UAV)was used to segment the sample trees of different litter grades,and the high spectral information of the sample trees was extracted.The random forest(RF),support vector machine(SVM)and BP(Back Propagation)neural network classification algorithm were combined with spectral feature screening method to establish a high-precision spectral feature discriminant model for the litter grade of the Malus Sieversii forest.RF,vegetation index(VI),continuous wavelet analysis(CWA)and successive projections algorithm(SPA)were used to screen 36 important band variables,21 vegetation indexes,21 class wavelet coefficient characteristic factors and 3 characteristic bands,respectively.Using single type of characteristics,SVM,RF and BP neural network classification algorithm were used to establish the classification model of withered branches.The results show that the classification accuracy of VI-RF discriminant model is the highest,reaching 91.53 %,followed by CWA-RF discriminant model,and the classification accuracy is 90.83 %.Full feature combination and VI-CWA feature combination are performed on four types of features.The accuracy of the model constructed by three classifiers is higher than that of single type feature modelling.The classification accuracy of VI-CWA feature combination is the highest,reaching 96.19%.However,after applying the discriminant model to the hyperspectral image of the experimental area for regional withered branches grade mapping,the accuracy of withered branches grade recognition is lower than 40%,and the accurate recognition of withered branches grade based on remote sensing images cannot be carried out well.(2)Shallow machine learning image classification method was used to identify the grade of withered branches and extract the information of withered branches.RF and Soft Max(SM)classifiers were used to classify the withered branches of the whole band hyperspectral data,24 texture features combined with hyperspectral data and 24 texture features combined with the first three principal components of principal component analysis(PCA).The accuracy of each scheme was not more than 40 %.Although the classification scheme of PCA+texture features has the best effect,the accuracy is only 37.69%.The withered branches were seriously confused with objects outside the forest land.Using the same data and methods,the withered branches,healthy branches,grassland and other land types were classified first and then the withered branches rate was calculated,and then the withered branches grade was divided.The PCA + texture combination classification accuracy was the best,and the overall accuracy(OA)reached 85.32 %,the recognition accuracy of dead branches reaches 84.64%.(3)A deep learning algorithm based on a prototype network is proposed for the identification of withered branches in the Malus Sieversii forest,which realizes high precision identification of withered branches grade and stable model migration.Using the full-band hyperspectral and the first three components of PCA as data sources,the prototype network is used to identify the withered branches.The results show that the PCA window size is 23 × 23,the Dropout value is 0.5,and the convolution layer is 3 layers,which can achieve 98.78 % high-precision withered branches recognition,on this basis,91.30% of the withered branches grade are recognized with high precision.When the number of training samples accounts for 10 % of the total number of samples,more than 90 % of the withered branches recognition accuracy can still be achieved,which proves that the prototype network is suitable for the small sample dead branch recognition of ground objects in the region.When the experimental area model was applied to test area 1,the accuracy of withered branches identification was only 65.69 %.After migration learning,the accuracy of withered branches identification in test area 2 reached 95.15 %,indicating that the model had stable withered branches identification performance.