Forest inventory data still represent the most direct, accurate and reliable source of information over a long period. Given the substantial labor, material and finances required, we need a reasonable sampling density (SD) to reduce the workload of field investigation. SD is a key issue on the accuracy and cost of estimation. A minimum number of sample plots given certain accuracy requirements is the most economical solution for spatial estimation of forest carbon. At present, most research on forest carbon storage is only valid for a particular SD. Studies of estimation accuracy at different sampling densities are rare. Generally, the greater the SD, the smaller the error. However, blindly increasing SD is not desirable and does not continuously reduce total error.
This study is based on Forest Management Inventory-measured aboveground carbon storage data of Xianju County for Taizhou (Zhejiang Province), and Landsat Thematic Mapper imagery of 30 m×30 m resolution. Using sequential Gaussian co-simulation (SGCS) and sequential Gaussian block co-simulation (SGBCS) algorithms, a geostatistical image-based conditional simulation technique was used to map and analyze uncertainty of natural resources and environmental systems during recent years. We explored the effect of SD on forest carbon and its spatial distribution estimates, uncertainties, and spatial variability at four SD levels, i.e., SD₁=100%, SD₂=80%, SD₃=60%, and SD₄=40% of total plots. Because the international forest carbon market needs various scales of spatial distribution, we designed two scale levels:1) the impact of different SD_s on the spatial distribution of carbon estimation at regional scale, using the SGCS algorithm with spatial resolution 30 m×30 m; and 2) the impact of different SD_s on upscaling for regional forest carbon estimation, using the SGBCS algorithm with spatial resolution 900 m×900 m. This study is an attempt to reduce the investigation workload and provides a reference for implementation of a forest resource inventory.
The results show the following. 1) Under different SD_s, SGCS and SGBCS had the same distribution trends in estimation of forest carbon density. SGCS estimation was able to meet accuracy requirements when for SD₂, carbon density was 0-67.485 Mg/hm² with mean 15.425 Mg/hm², consistent with the measurement. SGBCS carbon density estimation was less influenced by SD, all SD_s could meet the accuracy requirements, and a smaller SD had no substantial impact on upscaling. 2) Uncertainty of the SGCS and SGBCS estimation had overall rising trends, and the increase rate was smallest for SD₂. For SD₁, uncertainty of SGCS and SGBCS estimation increased by 1.08% and decreased by -1.71%, respectively. Uncertainty of carbon density estimation by SGBCS was less influenced by SD. When SD was changed from SD₂ to SD₃, it reduced the plot number, resulting in the greatest impact on uncertainty of SGBCS estimation. SD^had less contribution to estimation of the spatial variability. 3) Estimation of forest carbon storage and its distribution with the SGCS/SGCBS algorithms could reduce the requirement of SD appropriately. Not only were we able to obtain reliable estimation information, but we could also reduce the workload of the forest survey by at least 20% for SD at the SD₂ level (about 0.010% of total regional area).