Classification Of Dominant Tree Species Based On Multi-source Remote Sensing Data And Vegetation Phenology Characteristics

Forest resources are an important part of terrestrial ecosystem and the basis of forestry and ecological environment construction.With the development of remote sensing technology,the types of remote sensing data are getting richer and richer,and the information contained in remote sensing data is getting richer and richer,and remote sensing technology has been widely used in disaster monitoring,ecological environment supervision,urban surface classification and so on.However,in the face of such complex ecological environment as forest,the forest information extracted by different remote sensing techniques is limited,and how to quickly and accurately obtain the information of tree species categories and distribution range in the forest is an urgent problem to be solved in modern forestry remote sensing technology,and the multisource of remote sensing data and the synergy and complementarity of multi-source data can make up for the lack of information.There is an urgent need to make full use of multi-source and multi-temporal remote sensing data to analyze the differences among features,optimize the feature structure and improve the model classification accuracy in the classification of dominant tree species.In this thesis,taking Liangshui experimental forestry field of Northeast Forestry University as the research area,we investigate the effects of different temporal features and different machine learning algorithms on the fine classification of dominant tree species through the analysis of multi-source and multitemporal satellite remote sensing data,combined with the data of forest second-class resource survey,to provide reference for modern forestry resource survey.The main research contents and conclusions of this paper are as follows:(1)Based on multi-source and multi-temporal remote sensing data,multi-dimensional feature information is extracted and analyzed,thus significantly improving the accuracy of dominant tree species classification.Based on the polarization,spectral,vegetation index and texture features extracted from Sentinel-1 and Sentinel-2 data,this paper finds that there are differences in the phenological information provided by different single-temporal features for dominant tree species based on the analysis of the phenological characteristics of dominant tree species under different temporal phases;based on the separability analysis of J-M distance,it is found that the separability of multi-temporal features is significantly higher than that of singletemporal features.The separability of multi-temporal features was found to be significantly higher than mono-temporal features.The classification accuracy of the optimized multi-temporal features using random forest classifier can reach 80.61%,which is nearly 12% higher than that based on single-temporal features.(2)To address the problem that different classification algorithms have different adaptability in classifying dominant tree species,three different classification methods,SVM,CART,and RF,are selected to classify the optimized feature combinations in this paper.The results show that CART has the best classification accuracy of 82.42%,RF has the second best classification accuracy of 80.61%,and SVM has the worst classification accuracy of 76.20%.CART has obvious advantages over RF and SVM in the application of dominant tree species classification,and is more suitable for dominant tree species classification in this study area.