Modeling Down Dead Wood for US Forest Carbon Reporting

Abstract:

Down dead wood is an important component of forest ecosystems, affecting ecological processes and systems such as wildlife habitat, nutrient cycling, or regeneration. It also represents a variable stock, often of 10 to 25 percent of aboveground forest carbon. The importance of the role of dead wood within forests varies with stand structure, history, or location. Down wood data are systematically sampled within the national forest inventory (NFI) of the United States. Here, we develop inventory-based estimates of the larger-piece fraction of down dead wood, or coarse woody material (CWM). This forest ecosystem carbon pool of non-living woody material is intermediate between the larger, intact, standing dead trees and the smaller-piece and less differentiated forest floor.

Specifically, we develop random forests (RF) and stochastic gradient boosting (SGB) regressions trained on the sampled data and applicable to estimates of either NFI plots or spatial data pixels identified as forest lands. Site to site DWM stocks are often quite variable because many, sometimes unrelated, biotic and abiotic factors can affect rates of accumulation or loss. We explore a wide range of potential predictors in developing models. Model accuracies were generally similar between RF and SGB; predictions for the full set of NFI plots were more accurate than pixel predictions from spatial data only. The models represent improved estimates, as measured by independent test data, but clearly do not capture all CWM variability or the relatively rare extreme values. Analysis suggests additional modeling. Regression based estimates of current CWM for conus represent decreases relative to current models; changes range from -16 percent (RF, over all current NFI plots) to -22 percent (SGB, over all current NFI plots). Regional effects varied; estimates of CWM increased approximately 20 percent for Rocky Mountain forests, while Southern forests saw a decrease of 60 percent. These models expand predicted values from the subset of plots with CWM sampling to all NFI plots and in a similar way expand estimates to conus forest pixels.