Examining imputation techniques for time-series estimates of forest carbon stocks in the United States

Background/Question/Methods

Forest ecosystem carbon flux has been monitored by the Intergovernmental Panel on Climate Change (IPCC) since 1990 – the base year for which all subsequent IPCC reports reference. In the United States, estimates of forest carbon flux are obtained from data collected and maintained by the USDA Forest Service, Forest Inventory and Analysis (FIA) program. Over the course of the IPCC monitoring period, the FIA program transitioned from state-by-state periodic inventories – with reporting standards largely tailored to regional requirements – to nationally consistent, annual surveys tailored to large scale strategic requirements. Lack of measurements on all forest land during the periodic inventory, along with access issues, and misidentification of forest plots as non-forest due to poor aerial imagery have resulted in plot-level data gaps spread throughout the FIA database. These data gaps contribute to large differences in estimates of carbon flux between periodic and annual inventories. In this study, we compare several approaches that compensate for missing observations with respect to the accuracy and precision of stratified estimates of carbon stocks per unit area using data from the Lake States region of the U.S. Preliminary analyses were restricted to 18,702 plots where at least one accessible forest land condition was present during the annual inventory period, 1999-2010.       

Results/Conclusions

In the preliminary analysis, the proportion of missing plots was 2% in Minnesota and 3% in Michigan and Wisconsin. Nearly all missing observations (87%) were due to private landowners denying field crews access to their lands. All plot-level means from the distributions of simulated stratified estimates of live tree aboveground carbon stocks obtained with the different imputation approaches were within one standard error of the stratified mean estimates of live tree carbon calculated using observations for all plots. This suggests that there were no statistically significant differences among estimates obtained using the approaches investigated in the study. However, some approaches were computationally more efficient than others yielding similar estimates. The results from this study suggest there are several techniques for dealing with missing observations in forest inventory data. While this preliminary analysis focused on the Lake States region, the techniques described herein may be useful in aligning national carbon flux estimates from periodic and annual inventories.