Classification Of Complex Mountain Vegetation Based On Multiple Classifiers Combination

The study of vegetation type is an important component in the monitoring of land resources and the environment.Using remote sensing technology to identify large-scale vegetation types is an effective approach.Mastering the distribution of vegetation types plays an important role in the protection of regional resources and the enhancement of the environment.In this paper,the vegetation of the Baishui River Natural Reserve located in the transitional area of Qinba Mountain in 2020 was classified by multiple classifiers integrations.In the study,based on the 10 bands of sentinel L2 A data and other geographical auxiliary features,five simple classifiers of Bayesian,Cart,KNN,RF,SVM were used for vegetation type recognition to explore the effectiveness of sentinel rededge bands and other features on classification.Relief F feature selection algorithm was used to decrease the dimension of feature number.On this basis,the Multiple Classifier Combination Using Weight Vote Algorithm Based Modified Weght(MCC-WVA-MW)and the D-S evidence theory(KM-DS)algorithms were used to integrate multi classifier systems,to explore the schema that had the best classification effect in different classifier combination schemas.It should have provided a scientific basis for the study of vegetation types in the Baishui River Natural Reserve.The main conclusions are as follows:(1)When five single classifiers were used to classify spectral features,the classification effect was not good,and the kappa coefficient ranged from 0.57 to 0.73.If the rededge bands were removed,the accuracy of the classification would be reduced to between 0.54 and 0.65.This showed that the rededge bands increase the accuracy of the classification.When geometric and texture features were added to the classification,the accuracy was not improved.The addition of topographic features increased the kappa coefficient to 0.66-0.83.This showed that not the more the number of features,the better the classification effect.The Relief F algorithm was used to select 51 features,and finally,the top 25 features of feature weight were retained.Among them,the weights of topographic elements and vegetation index ranked in front.When these 25 features were classified by five single classifiers,the classification effect was ideal,which effectively improved the misclassification of vegetation types.After feature selection,OA increased to 0.82-0.9,and kappa coefficient increased to 0.78-0.88.However,there were still great differences in the distribution of vegetation types.(2)The MCC-WVA-MW algorithm integrated single classifiers into a multiple classifier system.The classification results were more stable,the OA accuracy difference was only 0.04,the kappa coefficient difference was 0.05,and the average OA and kappa coefficients were 0.85 and 0.82.There was no combination with an accuracy lower than0.8,which was an ideal classification method.Combined with the diversity measure of multiple classifier systems,the best one among the 16 schemes was the integration of cart,KNN,RF,SVM.The OA of this scheme was 0.87 and the kappa coefficient was 0.85.In this method,there was no trend that the classification accuracy increased with the increase of the number of component classifiers.The accuracy of some three classifiers was higher than that of five classifiers.(3)When using the KM-DS algorithm to integrated multiple classifier systems,there were still differences between classification results,and the fluctuation of the accuracy index was obvious.Its OA difference was 0.13 and kappa coefficient difference was 0.15,which was much higher than the previous integration algorithm.Even in the sixth combination scheme(Bayesian,RF and SVM),there was a serious water misclassification.Combined with the diversity measure of multiple classifier systems,the best classifier combination scheme was the integration of KNN,RF and SVM.Although the best scheme was the integration of three component classifiers,the average accuracy of this method showed an increasing trend with the increase of the number of component classifiers.