Regional Dominant Tree Species Identification Based On Optical And Radar Remote Sensing

As one of the important indicators of forest resource monitoring,the identification of dominant tree species is of great significance to forestry development and planning.At present,forest resource survey mainly relies on manual field survey and statistics,which costs a lot of time,manpower and material resources and has low timeliness.With the development and application of remote sensing technology in forestry,these problems are gradually alleviated.At present,most of the researches on remote sensing tree species identification are limited in the research area,and the generalization ability of models is slightly inadequate.In order to explore the identification methods of dominant tree species in a large area,this study took Chunan County as the research area,and established the identification models of seven dominant tree species using Sentinel-2optical remote sensing images,forest resource class II survey data,digital elevation model and Sentinel-1 radar remote sensing images.The model adopts a three-layer structure of gradual refinement: the first layer is the forest recognition model which is constructed by RF combined with the independent variable factors of upper optical remote sensing.The second layer is the tree species structure recognition model which based on three combination schemes with various independent variable factors,the RFE method is used to select the factors,and the RF,XGBoost and LightGBM algorithms are involving in modelling.The third layer is the dominant tree species identification model,LightGBM and Stacking is used to build the model.The identification results of the second layer and the radar independent remote sensing factor are input as independent variable factors.Finally,RFE is also used to reduce the feature dimension,reduce the model complexity,and obtain the final experimental results.The results obtained through comparative experiments show that:(1)Compared with the single-layer model,the three-layer model structure has a higher complexity,but its accuracy has been significantly improved.(2)The Stacking model constructed in this study is not as accurate as the LightGBM single model and has higher complexity.