An Evaluation of Site-specific and Generalized Spatial Models of Aboveground Forest Biomass Based on Landsat Time-series and LiDAR Strip Samples in the Eastern U.S.A.

Aboveground forest biomass (AGB) contributes a large part of the terrestrial carbon and varies with forest types. While forest disturbance and land use changes significantly influence greenhouse gas (GHG) emissions, international conventions on climate change mitigation have identified forest conservation and management as an efficient mechanism for enhancing carbon storage in live plant biomass to offset the emissions. Regional or national scale assessment of AGB on forest lands facilitates strategic plans such as the national GHG inventory per the requirement of the United Nations Framework Convention on Climate Change (UNFCCC). Large area inventory of AGB can be efficiently carried out by using remotely sensed data and a generalized statistical model in contrast to using only field measurements and site-specific models. In this study, we have integrated specialized forest inventory plot data with spatial predictors derived from time-series Landsat imagery and LiDAR strip samples in four sites across the eastern U.S.A.- Minnesota (MN), Maine (ME), Pennsylvania-New Jersey (PANJ) and South Carolina (SC) - to formulate statistical models. Pixel-level polynomial curve fit was applied to the time-series Landsat variables to obtain projected metrics in the target year 2014. Two forms of models based on ordinary least-squares regression (OLR) and the random forest (RF) algorithm were developed for each site and with the pooled (generalized) dataset. The site-specific models were tested with the national forest inventory (NFI) data as well as specialized plot data of the other sites while the generalized models were tested with the NFI data only. We observed similar level of accuracies with the OLR and RF models at the same site. Although the RF models yielded larger values for the amount of explained variance (i.e., pseudo R2) in the site-specific models, the generalized model was also promising. The amount of variance explained with the site-site specific models were 82.2, 70.2, 70.5, and 87.4 % for ME, MN, PANJ and SC, respectively while the generalized model had 65.3%. An equivalence test of observed and predicted values of plot-level AGB revealed that the generalized model provide equivalent estimates of AGB compared to site-specific models.