Study On The Classification Of Dominant Tree Species Based On Multi Model Fusion

Although the traditional forest resources monitoring method based on field survey has reliable data sources,it also consumes a lot of manpower,material and financial resources,and the resolution of the data obtained is low,even limited by geographical conditions.Some forest field surveys are difficult to carry out in practice.With the development of remote sensing technology,the number of successful launches of high-resolution earth observation satellites in China has also continued to increase.How to promote the application of domestic high-resolution remote sensing data in forest resources information survey and dynamic monitoring has important theoretical and practical value.This paper takes Longquan city of Zhejiang Province as a research area,forest resources class as research unit,the basic research data are GF-2 satellite image,forest resources survey data and DEM data.,four dominant tree species of Phyllostachys pubescens,broad-leaved trees,Pinus massoniana and Cunninghamia lanceolata in the study area are classified.The main research contents and results are as follows:(1)Preprocessing the remote sensing image of GF-2 satellite,extracting the average spectral information of each band from multi spectral image,calculating and deriving 5 vegetation indexs,extracting 8 GLCM texture information from panchromatic image,extracting 3 terrain factor features from DEM data,using random forest(RF)to carry out quantitative analysis of the extracted features,and obtaining only based on spectrum and only based on grain.The result of classification is not ideal.The best classification scheme is to combine texture and average spectral information,and add vegetation indexs and terrain factors.(2)Four single models(XGBoost,RF,SVM,AdaBoost)and one stacking integrated model were used to classify the dominant tree species in the study area,among which the integrated model had the best classification effect,the overall classification accuracy was 82.99%,Kappa value was 0.77;XGBoost single model classification accuracy was next,the classification accuracy was 81.79%,Kappa coefficient value was 0.75;followed by RF and SVM model;The classification accuracy of AdaBoost single model is the worst,73.43%,and Kappa coefficient is 0.63.The Stacking based model enhanced learning to a certain extent,improved the classification effect of a single model,improved the classification accuracy by 1.2% and Kappa coefficient by 0.02 compared with the best single model(XGBoost),and was more suitable for tree species classification in the study area.(3)In terms of the classification of all kinds of dominant tree species,four single models and one integrated model have better classification effect on bamboo,the user accuracy is greater than 80%,and the producer accuracy is greater than 85%,especially XGBoost and stacking integrated model have the best classification effect on bamboo;but XGBoost and stacking integrated model have a larger leakage error on broad-leaved trees,about 50%;AdaBoost has the worst effect,and four tree species have the worst leakage error.The accuracy of classification is lower than that of the other four models,of which the error of misclassification and the error of omission are the largest,the error of misclassification is close to 50%,the error of omission is about 30%,and the error of omission is the highest,which is 31.34%.