Remote Sensing Estimation Method Of Forest Aboveground Carbon Storage At District And County Level Based On Forest Types

Forests play a crucial role as a sink of carbon in terrestrial ecosystems,helping to mitigate global climate change.Accurate estimation of forest aboveground carbon stock(AGCS),a significant component of forest carbon storage,is of great significance for forest ecological function evaluation,carbon sequestration and sink enhancement potential calculation and carbon market trading.In recent years,large-scale afforestation and reforestation programs in China have resulted in forests dominated by middle-aged and young trees,with a high carbon sequestration rate and growth potential.Therefore,to achieve scientific management of China’s future forest carbon sinks and to scientifically evaluate China’s"dual carbon"targets,it is urgent to conduct research on the spatial distribution and variation of forest AGCS.Therefore,the accurate estimation and mapping of forest AGCS are crucial elements for optimizing forest carbon sink management plans as well as providing data foundations for assessing China’s future forest carbon sequestration potential.This study focuses on the forests in Huangshan District,Anhui Province,and develops a remote sensing estimation method of forest aboveground carbon stock(AGCS)at district and county level based on different forest types,by combining the airborne laser scanner(ALS)point cloud,very high spatial resolution(VHR)remote sensing images and field survey data.This study explores the applicability and accuracy of the method through three processes:sample-plot survey based on active and passive remote sensing,forest type classification based on deep learning,and construction of optimal variable library based on different forest types and model fitting.The main research contents and results are as follows:(1)This study explores the abilities of terrestrial laser scanning(TLS)and unmanned aerial vehicle(UAV)remote sensing technologies in forest sample-plot surveying from the perspectives of accuracy,efficiency,and uncertainty across different forest types,in order to provide accurate sample-plot dataset for precise estimation of forest AGCS.The results show that:(a)For the arbor and bamboo forest with high canopy cover,the combined TLS and UAV air-ground integrated remote sensing survey method was proposed to solve the problem of height extraction.Canopy height model(CHM)seed points(CSP)method proposed in this study should be preferentially utilized to extract the height of dominant trees(R~2≥0.973).Then,the location coordinates of other individual trees were obtained by the individual tree location(ITL)method,and the height of other individual trees were extracted according to CHM(R~2≥0.886).(b)For special shrubbery forest,TLS-based method was used to extract structural parameters.The height extraction accuracy was satisfactory(R~2≥0.940),while the crown width extraction accuracy was slightly poor(R~2 of tea plot was 0.903,and R~2 of shrubbery plot was 0.661).The extraction accuracy of the canopy cover was affected by the resolution of CHM,and the optimum CHM resolution was 1 cm.(c)For special shrubbery forest,an information loss rate(L)calculation method based on the volume element concept was proposed,which can directly reflect the impact of scanning mode and scanning spots on the uncertainty of TLS data collection.The single scanning mode was suitable for acquiring basic information rapidly in special shrubbery plots,while the multiple scanning mode required reasonable planning and laying of scanning spots based on the height distribution of shrubs to improve data collection efficiency in special shrubbery forest plots.In addition,the quality of the TLS point cloud data was mainly affected by the distance from the central scanning spot and the height distribution of shrubs,the farther the distance,the greater the information loss.And the height distribution of the point cloud(the appearance of higher shrubs)affected the acquisition quality of all subsequent point clouds in this scanning direction,including ground and vegetation points.Overall,the active and passive remote sensing survey technology used in this study can address the survey requirements of different forest types and complexity levels by selecting or combining appropriate active and passive remote sensing sensors,and improve survey efficiency and structural parameters extraction accuracy.(2)This study utilized the standard sample-plot size(about 0.067 hm~2)as the classification unit,combined ALS point cloud data and VHR remote sensing image data,and used three deep learning classification methods(VGG16,Goog Le Net,and Res Net50 models)based on convolutional neural network framework to solve classification problem and AGCS contribution differentiation problem of the different forest types.The results show that:(a)When using single VHR remote sensing image data for forest type classification,compared with VGG16 and Res Net50,Goog Le Net has faster convergence of loss function,more stable function curve,and higher classification accuracy(overall accuracy is 82.33%,Kappa coefficient is 0.721).(b)Combined ALS data can effectively modify the forest type classification result based on the single VHR data.The classification accuracy of the VGG16,Goog Le Net and Res Net50 models have been improved(the overall accuracy increased by 2.75%,3.5%,and 4.09%,with corresponding increases in Kappa coefficients of 0.051,0.072,and 0.069,respectively).In summary,this study comprehensively considers the spatial texture features of VHR data and the height features of ALS data,which can improve the classification accuracy of different forest types.Furthermore,using the standard plot size as the classification unit of forest type can effectively utilize the survey data of sample-plot,and provide training datasets and technical support for the study of forest type classification at district and county level.(3)In the study,spectral vegetation index variables(VHR＿SVI)and texture feature variables(VHR＿TF)were extracted based on VHR remote sensing image,height-related variables(ALS＿H),intensity-related variables(ALS＿I),density-related variables(ALS＿D)and structural attribute variables(ALS＿SA)were extracted based on ALS point cloud data.The optimal variable library suitable for AGCS estimation of different forest types was constructed through variable screening.And based on the optimal variable library,multiple linear regression(MLR),multiple nonlinear regression(MNLR),random forest(RF),support vector machine(SVM)and feedforward neural network(FNN)were compared to screen the best AGCS estimation model for the different forest types.The results show that:(a)There are significant differences in the optimal variables of different forest types.VHR＿SVI and VHR＿TF variables have only good fitting ability for shrubbery forest,while ALS＿I variables has only good fitting ability for mixed broadleaf-conifer forest,whereas ALS＿D,ALS＿H,and ALS＿SA variables have good fitting abilities for all forest types.(b)In parametric models,MNLR has the best fit in broadleaved forest and mixed broadleaf-conifer forest;MLR has the best fit in shrubbery,coniferous and bamboo forests.And shrubbery forest has the highest accuracy(R~2=0.841,RMSE=2.212 t/hm~2)and broad-leaved forest had the lowest accuracy(R~2=0.579,RMSE=2.470 t/hm~2).(c)In non-parametric models,the accuracy of FNN model is the highest(R~2≥0.907),while the accuracy of RF model is the lowest(R~2≤0.897)due to the influence of the number of samples.To sum up,this study improved the efficiency and accuracy of AGCS model estimation of different forest types through the construction of optimal variable library and the screening of fitting models,and also provided applicable model reference and variable library with mobility and verifiability for forest AGCS remote sensing estimation at district and county level.According to the results of the three processes,the optimal Goog Le Net classification model and the optimal FNN estimation model were used to map the AGCS distribution of each forest type in Huangshan region.The results showed that,the total forest AGCS in Huangshan was3,134,542.485 t,with an average of 22.605 t/hm~2.Broad-leaved forest had the highest proportion(81.15%)while shrubbery forest had the lowest(2.96%).Therefore,the AGCS remote sensing estimation method for different forest types proposed in this study can provide accurate estimation results of forest AGCS,and provide data and technical support for spatial distribution,dynamic monitoring and carbon sequestration potential assessment of forest AGCS at district and county level.