Multi-Model Estimation Of Forest Canopy Closure By Remote Sensing Based On Sentinel-2 Images

Forest canopy closure is an essential parameter in forest resource surveys and plays a significant role in forest ecosystem management.In remote sensing,red-edge bands are important indicators for monitoring vegetation growth.In order to explore the application potential of red-edge band in forest canopy closure estimation,this paper takes the forest land area of Chifeng City,Kalaqin banner,Ningcheng county and Aohan banner in Inner Mongolia as the research area,and constructs empirical statistical model,machine learning model and physical model based on sentinel-2 data,SRTM1 data with ground measured data.The main research contents are as follows:(1)The image feature factors of the study area were extracted from Sentinel-2 data and SRTM1 data.The importance of the feature factors was ranked according to the correlation between the feature and canopy closure and the contribution of the feature variables to the canopy closure.The relatively independent and highly important preferred features were selected as the independent variables to build the model.(2)The empirical statistical model(multiple stepwise regression model)and machine learning model(BP neural network model)were constructed based on the selected preferred feature factors and 70% of the ground measured data.Moreover,the remaining 30% of the measured data were used to verify the results.At the same time,the physical model was selected to estimate the forest canopy closure in the study area.The fully constrained mixed pixel decomposition model and Li-Strahler geometric-optical model were used to obtain the inversion results which were verified the accuracy with all the measured data.(3)This paper compared the modeling accuracy of the multiple stepwise regression model and the BP neural network model.Meanwhile,verifiesd the estimation accuracy of the three models and analyzed the error.Compared with the existing models of forest canopy closure estimation based on multispectral images at home and abroad,the application potential of red-edge bands in forest canopy closure estimation was analyzed.The main conclusions are as follows:(1)Combined with the correlation and response level between each characteristic variable and canopy closure,it was found that spectral features have the largest contribution to canopy closure,followed by red-edge vegetation index features.Texture features also have a certain contribution to canopy closure,while terrain factors and vegetation indices have the smallest contribution.The spectral feature variables: Blue,Red,SWIR2,red-edge feature variables: SRre2,NDVIre3,texture feature variables:Mea2,Cor7,SM1,Var3 and terrain feature variable: DEM were selected as the optimal feature variables to build the model.(2)Among the three models of canopy closure inversion constructed in this paper,the fitting accuracy of the canopy closure inversion results based on BP neural network model was higher,with the estimation accuracy was 84.7%.The error volatility of the model was small,and the model stability was good.However,it had some underestimation phenomenon of high canopy closure value and low canopy closure value,and the generalization ability of the model was weak.To some extent,the stepwise regression model could quickly estimate the forest canopy closure at the regional scale,with the estimation accuracy was 76.81%.There were obvious underestimation of high canopy closure areas and overestimation of low canopy closure areas in the estimation results,and the model was lack of stability.The canopy closure results obtained from the inversion based on the fully constrained least square mixed pixel decomposition model and Li-Strahler geometric-optical model.The accuracy of the model was 71.38%.There were large errors in the estimation results,which may be due to the low accuracy of the input parameters of the model,and the actual crown shape did not conform to the assumptions of the model.(3)Red-edge bands have great application potential in building high-precision forest canopy closure estimation model.The red-edge vegetation indices extracted in this paper are highly correlated with canopy closure.In the process of constructing the regression model,the optimal regression equation retains the red-edge indices,which indicates that the red-edge bands have a great contribution to the inversion of forest canopy closure.In addition,the multiple stepwise regression model constructed by the red-edge indices and the BP neural network model both obtain higher accuracy results,which are better than those of similar models at home and abroad shows that the red-edge band has research significance in forest canopy closure estimation.This study provides a new method for forest monitoring and evaluation and a scientific basis for remote sensing model selection.