Remote Sensing Estimation Of Bamboo Forest Aboveground Carbon Storage Based On Ensemble Learning

Forest aboveground carbon storage（AGC）is an important indicator to reflect the carbon sequestration capacity of forest ecosystem,and to evaluate the quality of ecosystem,forest structure,function and production potential.It is of great significance to maintain forest ecological balance.Therefore,monitoring the forest AGC is an important content of forest monitoring under the background of forestry serving the national double carbon goal.Bamboo forests are wide spread in sub-tropical areas of China,and are well-known for their rapid growth and great carbon sequestration ability.It plays an important role in the carbon balance of regional ecosystem.It is of great significance to estimate the bamboo forest AGC accurately and evaluate the role of forestry in coping with the climate change.This study taked the bamboo forest resources in Zhejiang Province as the research object and taked Landsat 8 OLI images as the data source,combined with the field survey data of bamboo forest,constructed machine learning models,spatial models and ensemble learning models respectively.At the same time,in order to achieve accurate estimation of the bamboo forest AGC,this study proposed a dynamic geographically weighted stacking regression（GWSR）model based on model coupling of different geographical locations,which fully considerd the impact of geographical location information on different models.Through research,the following three conclusions were obtained:1.In three machine learning models,the Random Forest Regression（RFR）model had high precision and low error,its prediction precision（R²）was 0.78,its error（RMSE）and its NRMSE were2.02 Mg ha^-1and 0.126;followed by the Extremely Randomized Trees（ERT）model,and the Support Vector Regression（SVR）model had the lowest precision.The R²of the RFR model was 3%and 11%higher than the ERT and SVR models,respectively.The RMSE of the RFR was same as the ERT model,35%lower than the SVR model;and their NRMSE were 23%lower than the SVR model.2.In spatial models,the R²of the Geographic Weighted Regression（GWR）model was 0.74,9%higher than the Cokriging（COK）model.The RMSE and the NRMSE of the GWR model was 2.56Mg ha^-1and 0.161,7%and 9%lower than the COK model,respectively.The optimal bandwidth of the GWR model was 156 m.The parameters of different variables in the GWR model had obvious spatial differences.The parameters B7,TM457,NDVI,and NDWI were larger in northwestern Zhejiang Province,while the parameters TM543 and W7B6VAR were larger in southern Zhejiang Province.This is the advantage of the GWR model,considering the change in parameters at different spatial positions.3.Based on the above model research,we combined RFR、ERT、SVR、GWR and Ada-CART models to build the Geographic Weighted Stacking Regression（GWSR）model.The R²of the GWSR model was 0.83,the RMSE and NRMSE of the GWSR model were 1.84 Mg ha^-1and 0.111.Compared with four single model such as RFR,ERT,SVR,GWR and so on,the maximum increase rate of the R²was 19%,the maximum decrease rate of the RMSE was 40%and the NRMSE was 32%.In addition,the R²of the GWSR model achieved better precision than other ensemble learning model(Ada Boost-CART,Ada-CART,R²=0.76,RMSE=2.127 Mg ha^-1,NRMSE=0.133).This shows that the GWSR model can realize the high-precesion estimation of the bamboo forest AGC in Zhejiang Province.The spatial pattern of the bamboo forests AGC in Zhejiang Province had a relatively dense spatial distribution in the northwest,southwest and northeast.This is in line with the actual bamboo forest AGC distribution in Zhejiang Province,indicating the potential practical value of our study.