U.S. Department of Agriculture: Forest Service -- National Agroforestry Center


Date of this Version



Remote Sensing of Environment 124 (2012) 270–281; doi:10.1016/j.rse.2012.05.016


Mountain pine beetle outbreaks have caused widespread tree mortality in North American forests in recent decades, yet few studies have documented impacts on carbon cycling. In particular, landscape scales intermediate between stands and regions have not been well studied. Remote sensing is an effective tool for quantifying impacts of insect outbreaks on forest ecosystems at landscape scales. In this study, we developed and evaluated methodologies for quantifying aboveground carbon (AGC) stocks affected by mountain pine beetle using field observations, lidar data, and multispectral imagery. We evaluated methods at two scales, the plot level and the tree level, to ascertain the capability of each for mapping AGC impacts of bark beetle infestation across a forested landscape. In 27 plots across our 5054-ha study area in central Idaho, we measured tree locations, health, diameter, height, and other relevant attributes. We used allometric equations to estimate AGC content of individual trees and, in turn, summed tree AGC estimates to the plot level. Tree-level and plot-level AGC were then predicted from lidar metrics using separate statistical models. At the tree level, cross-validated additive models explained 50–54% of the variation in tree AGC (root mean square error (RMSE) values of 26–42 kg AGC, or 32–48%). At the plot level, a cross-validated linear model explained 84% of the variation in plot AGC (RMSE of 9.2 Mg AGC/ha, or 12%). To map beetle-caused tree mortality, we classified high-resolution digital aerial photography into green, red, and gray tree classes with an overall accuracy of 87% (kappa=0.79) compared with our field observations. We then combined the multispectral classification with lidar-derived AGC estimates to quantify the amount of AGC within beetle-killed trees at the field plots. Errors in classification, apparent tree lean caused by off-nadir aerial imagery, and a bias between percent cover and percent AGC reduced accuracy when combining multispectral and lidar products. Plot-level models estimated total plot AGC more accurately than tree-level models summed for plots as determined by RMSE (9.2 versus 21Mg AGC/ha, respectively) and mean bias error (0.52 versus −6.7 Mg AGC/ha, respectively). When considering individual tree classes (green, red, gray) summed for plots, comparisons of plot-level and tree-level methods exhibited mixed results, with some accuracy measures higher for plot-level models. Despite a lack of clear improvement in tree-level models, we suggest that tree level models should be considered for assessing situations with high spatial variability such as beetle outbreaks, especially if apparent tree lean effects can be minimized such as through the use of satellite imagery. Our methods illustrate the utility of combining lidar and multispectral imagery and can guide decisions about spatial resolution of analysis for understanding and documenting impacts of these forest disturbances.