Studies On Remote Sensing Estimation Of Forest Carbon Stock For Shenzhen City

With the influence of global warming,studying carbon cycling has become a hot topic in the field of global climate change.Forests are the largest carbon pool in terrestrial ecosystems and play an important role in global carbon cycling.Urban forests,called "city lung",are an important carbon repository of urban ecosystems.Therefore,it is greatly significant to study spatial distribution characteristics of urban forest carbon stock.With the use of remote sensing technology in precision forestry,using a variety of remotely sensed data to estimate forest carbon stock has become a popular method.It is very urgent and important to develop a method that can be used to accurately map urban forest carbon stock.In this study,Shenzhen City was selected as a study area to develop and examine the proposed method using Landsat 8 and Gaofen-1(GF-1)images.Based on two kinds of images,a total of 55 spectral variables were derived respectively,including original bands,various vegetation indices,texture measures and terrain factors.A total of 168 sample plots were selected by a stratified random sampling procedure.The correlation analyses between the spectral variables and urban plot forest carbon stock were first conducted and the appropriate independent variables were selected and utilized as independent variables of models.Partial least squares regression model,Radical Basis Function(RBF)neutral network model and Random Forest model were developed to account for the relationship of the urban forest carbon stock and the selected spectral variables and to generate the spatial distributions of the urban forest carbon stock of the study area.The comparisons of the results from the models provided the potential to improve mapping the urban forest carbon stock.The main results are as follows:(1)When Landsat 8 image was used,the most accurate model for mapping the urban forest carbon stock for Shenzhen was obtained using Random Forest with root mean square error(RMSE)of 9.541(t/hm~2).However,RBF neutral network had the largest determination coefficient R~2 of 0.972.The partial least squares regression model had the lowest fitting and prediction accuracy with R~2 of 0.438 and RMSE of 12.871(t/hm~2).(2)When GF-1 images were utilized,the results were similar to those using Landsat 8 image.Random forest model had the greatest prediction accuracy with a RMSE value of 7.221(t/hm~2),while RBF neutral network model had the highest R~2 coefficient of 0.974.The partial least squares regression model led to the smallest R~2 value of 0.667 and RMSE of 10.410(t/hm~2).(3)Three models based on Landsat 8 image resulted in similar spatial distributions and patterns of urban forest carbon stock for Shenzhen.The distributions and patterns were characterized by larger estimates of urban forest carbon stock in the southeastern coastal areas,and smaller values in the urbanized and developed areas in the central and western regions.They were consistent with the distributions of the urban forests in Shenzhen.The results implied that the forest carbon stock was related to the forested areas in Shenzhen,forest carbon sink could be increased by expanding urban forests,and urban forest ecosystem carbon cycling research is of great significance.(4)The spatial distributions and patterns of Shenzhen forest carbon stock generated using three models based on GF-1 image were consistent with those of the three models based on Landsat 8 image.But,the GF-1 image based models led to the different minimum and maximum values.Overall,the partial least squares regression model had the smallest estimates with most of them falling within the interval of 0 to 60(t/hm~2)and negative values concentrating in the built-up areas and bare lands or the areas that had low vegetation cover.(5)By comparing the results of Shenzhen forest carbon stock from Landsat 8 and GF-1 image,it was found that compared with the Landsat 8 image based models,the GF-1 image based models decreased the RMSE values by 1.40(t/hm~2)to 2.320(t/hm~2).This implied that the GF-1 image was superior to the Landsat 8 image in the application of mapping Shenzhen urban forest carbon stock and can be used for quantitative monitoring of forest carbon sequestration.