Research On Remote Sensing Estimation Of Forest Aboveground Biomass Based On Improved Brouta Algorithm

Forest aboveground biomass(AGB)is one of the critical parameters to assess forest ecosystem productivity and health status.It is of great significance to the global carbon cycle and climate change.The feature variables such as band features;,vegetation indices,texture features,and terrain features related to forest AGB can be extracted by remote sensing and other means.There are often more feature variables,which affect the estimation accuracy.By improving the Boruta algorithm,the feature redundancy and the transmission of errors in the process of remote sensing estimation of forest AGB can be effectively reduced,which provides reference significance for remote sensing inversion.The study area is the Luton forest farm.Sentinel-2 images and NASA DEM were used as original data and Class II survey data in the study area were used as measured data.Texture features,single band,vegetation index,texture factor,and topographic factor were selected as the alternative independent variables,and the forest AGB was taken as a dependent variable.The optimal combination of 28 independent variables with significant correlation with forest AGB was selected by Pearson correlation analysis to construct a Support Vector Regression(SVR)model.The SVR model was then optimized with three feature selection algorithms,namely Boruta,LASSO and RF(Random Forest,RF),to compare and analyze the accuracy of the four forest AGB estimation models.Finally,the optimal model is selected to invert the AGB of the study site and produce a spatial distribution map of the AGB of the study site.The main contents and results of the study is as follows:(1)Forest AGB remote sensing features variable extraction and correlation analysis.In this study,through the combination of bands,texture information extraction,principal component analysis,and calculation of various vegetation indices of Sentinel2 remote sensing images,16 vegetation indices,10 single bands,and 11 principal component texture features were extracted.3 topographic features were extracted from NASA DEM data at 30m spatial resolution.Pearson correlation analysis was performed vegetation index was good.There was a more significant correlation with some remote sensing bands,a less significant correlation with some texture features,and a strong correlation with DEM in terrain features.(2)Forest aboveground estimation based on the SVR model.After correlation analysis of numerous variables of forest aboveground biomass,Linear kernel function,Radial Basis Function,Polynomial kernel function,and Sigmoid kernel function is adopted to establish the SVR model.The coefficient of determination(R-Squared,),Root Mean Squard Error(RMSE),and Mean Absolute Error(MAE)of the modeling and validation sets are used as the model accuracy evaluation criteria,and the SVR model parameters are sought to be optimized while comparing the modeling accuracy of different kernel functions.The study results showed that this study selected the RBF function as the kernel function of the SVR model,and the grid search method and tenfold cross-validation are used as the parameter search method to establish the SVR model for forest aboveground biomass estimation.The accuracy of the SVR model was tested using the validation data,and the results of the study showed that the validation set of the SVR model was 0.60,the RMSE was 36.46t/ha,and the MAE was 24.79t/ha.(3)Forest aboveground biomass estimation of based on improved Boruta-SVR.The improved Boruta algorithm is to create a shadow feature for the initial feature(m rows and n columns,there are m sets of samples,n initial features,m>1,n>1).Shadow features are obtained by extracting[m-p]*n sets of samples from the initial features in proportion P(0<=p<1),putting them back after random row transformation,and then mixing the initial features and shadow features to form a new feature combination.Based on multiple iterations of XGBoost,the importance of each initial feature is compared with the maximum Z-score(Zscore)of the shadow features to obtain the best feature combination,and the improved Boruta-SVR model is constructed for forest AGB estimation.The results showed that the validation set of the improved BorutaSVR model is 0.68,the RMSE is 24.31t/ha,and the MAE is 17.97t/ha,and the accuracy is greatly improved compared with that of the SVR model,which is without the improved Boruta feature selection.(4)Comparative analysis of the accuracy of forest aboveground biomass remote sensing estimation models.To reflect the advantages of the improved Boruta algorithm,LASSO and Random Forest are also introduced in the study to optimize the SVR model respectively.The accuracy of the analysis of SVR、Improved Boruta-SVR、LASSOSVR、RF-SVR models were compared.The study results showed that the SVR model based on high-dimensional remote sensing feature variables and constructed without feature optimization has the lowest accuracy,followed by the SVR model optimized by the LASSO algorithm is better than the SVR model optimized by RF,and the accuracy of the SVR model optimized by improved Boruta algorithm is the highest.The improved Boruta-SVR model was selected to invert the forest aboveground biomass in the study area.(5)In this study,the Improve Boruta-SVR model was used to invert the aboveground biomass in the research area,and the spatial distribution map of the aboveground biomass in the study area was obtained.The study results showed that the total amount of aboveground biomass in the research area is 4.53*105t,and the spatial distribution shows a north-south trend.The aboveground biomass gradually becomes larger and denser from south to north.It is mainly distributed in the northwest、northeast and central areas where the forest land is gentle.The spatial distribution of aboveground biomass at the study site is roughly consistent with the actual geological landform and forest vegetation distribution of the forest farm.