Although endemic to North America and having an important role in forest ecosystem processes, the most recent outbreak of mountain pine beetle (Dendroctonus ponderosae) shows impacts on forest structure and ecosystem processes that are acute and widespread. The current outbreak may heavily impact water runoff, water quality, nutrient cycling, forest regeneration and regrowth, native and non-native plant communities, soil chemistry and wildlife habitat.
Landscape-scale outbreaks of mountain pine beetle afflict pine species, particularly lodgepole (Pinus contorta), ponderosa (Pinus ponderosa), Scotch (Pinus sylvestris) and limber pine (Pinus flexilis), and are often associated with an array of interacting disturbance variables including fire damage, drought, prolonged mild winters, interspecific and intraspecific competition, and synergistic interactions between other native and non-native insects and diseases. The beetles often spread through a “mass attack” strategy, depositing eggs which hatch inside the host tree. Additionally, the beetles transmit blue stain fungus (Grosmannia clavigera) which clogs the tree's xylem and reduces its ability to transmit water from the soil to the canopy [1-2].
In 1996, the U.S. Forest Service Rocky Mountain Region initially detected the mountain pine beetle outbreak in northern Colorado, including the Arapaho-Roosevelt, Medicine Bow-Routt and White River National Forests. By 2010 the beetle had spread across the western United States, encompassing approximately 2.5 million hectares (6.1 million acres) of pine forest. Of this total, 1.6 million
hectares (4 million acres), or approximately 66% of the spread, is located in Colorado and southern Wyoming .
The number-one priority for the region is post-outbreak mitigation. However, limited funding has hampered progress. While general aerial detection surveys including aerial sketching have effectively measured the rate of spread, they are unable to quantify the effects of beetle mortality on the forest overstory, are not replicable and do not provide spectral capabilities for analysis [4-5]. Furthermore, detailed aerial surveys including aerial digital photography, airborne videography and heli-GPS surveys are often prohibitively expensive.
The use of legislative acts to address bark beetle outbreaks and collaboration with communities of interest has become a key approach to the Forest Service’s management strategy. Effective detection and mapping of beetle mortality across multiple scales is urgently needed to assist in determining high risk areas in need of mitigation and restoration. Accurate baseline maps displaying the abundance of tree species across the landscape have become a necessary tool for addressing the array of important and often overlapping concerns of land managers, policy-makers and the public.
The DEVELOP team at Fort Collins Science Center (DEVELOP/FCSC) started with detailed data from 74 quarter-acre, fixed radius forest plots, collected by the Forest Service in 2008, which contained detailed information about the species abundance within each plot. DEVELOP/FCSC used the quarter-acre plot data in conjunction with NASA’s Earth Observing Systems (EOS), and a Boosted Regression Tree (BRT) model to create a preliminary forest cover classification model of Fraser Experimental Forest. This forest cover classification model was subsequently calibrated with field validation methods including assessment of fifty-two 7.2 meter, fixed radius forest plots and 60 roadside observations stratified across the study site that were collected in 2012. A second model was run with adjusted parameters and validated with an independent dataset comprised of a mixture of 2008 and 2012 forest plots.
The NASA EOS data used for this project included selections of Landsat 5 imagery, with four scenes selected that could be immediately downloaded (05/08/1994, 06/12/1995, 03/26/1996 and 07/03/1997). DEVELOP/FCSC interns applied the Landsat 5 calibration  to the images to estimate Top of Atmosphere Reflectance for each scene. Subsequently the team derived Normalized Difference Vegetation Index (NDVI) and Tasseled Cap (TCAP) Transformation. Field data were separated into “test” and “model” samples, the latter of which were used to generate different model runs, and the former used to verify the predictions. Twenty percent of the 74 quarter-acre plots catalogued in 2008 were withheld in addition to 20% of the fifty-two 7.2 meter validation plots collected in 2012, resulting in an independent dataset to test against the second model.
A BRT model was run, creating a mathematical relationship between the prepared predictor variables and known lodgepole and spruce-fir abundance (by percent composition) in order to predict abundance across the landscape. After running the model, the DEVELOP/FCSC team classified and interpreted the results. The 0.0 – 1.0 decimal outputs of the BRT model were generalized into classes of abundance (1 – 3) for lodgepole pine and spruce-fir. These classes were then added together to produce a single-layer output describing species dominance and abundance for each pixel on the first map (Figure 1). The decimal values of the abundance classes represent some percentage of species abundance on the landscape. Anything <= 0.15 was considered to be “not present” and thus = 0. Each color represents a specific combination of abundance values created when the rasters were summed. For example, orange areas indicate pixels with an identifier value of 300, meaning those areas consist of >55% lodgepole. Similarly, a light green area with the legend name of “Lodgepole-Spruce/Fir” indicates pixel values of 320 or 310 (lodgepole dominant with some spruce/fir). The medium green “Lodgepole\Spruce/Fir” class represents table values of 220, where either class type could be dominant. This lack of precision in the classification was one of the reasons for doing a second reclassification model. Over 100 field observations were used to validate the first model, achieving >70% accuracy using a single month of Landsat imagery (July 1997) and 2008 USDA Forest Service data. But numerous non-conifer validation points including aspen and human disturbance sites noticeably padded the model’s accuracy.
Figure 1. Map output of first BRT model run showing classification of species abundance across FEF
The threshold levels were adjusted for the second model including an increased upper bound of >90% abundance and the addition of Landsat 5 scenes from 1994, 1995 and 1996. The second model iteration produced a second map (Figure 2), with more precise classes based on new thresholds and eliminated ambiguity in class names.
Figure 2. Map output of second BRT model run showing classification of species abundance across FEF
R-squared outputs used to evaluate the forest cover classification model showed results lower than expected. Both models showed consistently low predictive capabilities based solely on the predictor variables used. Assessing the R-squared values of both models in concert with the ranking of predictor variable importance revealed the need for additional variables for the next round of model runs. None of the individual species model runs could explain more than 20% of the variance with solely the predictors used, which was inconsistent with the relatively “good” performance resulting from the visual interpretation and error matrix assessment. This necessitates bringing in important categorical datasets including soil layers and additional Landsat 5 imagery that spans all seasons of an expanded set of years. This project has carried on into the current Fall 2012 DEVELOP term. The DEVELOP/FCSC team will pursue a final model with additional predictor variables with >80% accuracy.
This project demonstrated the utility of BRT modeling in concert with Landsat imagery to create a fine resolution forest cover classification model with limited predictor data, and acts as an important baseline for subsequent model calibration by future NASA DEVELOP teams. Ultimately this methodology can be a useful decision support tool for resource managers and community members affected by mountain pine beetle. The decision support tools, including forest cover classification maps, modeling methodologies and field methods tutorials, were delivered directly to the resource managers. Furthermore, all methods and remote sensing resources will be available to interested community members through the ColoradoView website, and the team will share the project results at the annual American Geophysical Union (AGU) Fall Meeting in San Francisco, CA in early December 2012.
 Safranyik, L., and Carroll, A.L. 2007. The biology and epidemiology of the mountain pine beetle in lodgepole pine forests. In “The mountain pine beetle: A synthesis of biology, management and impacts on lodgepole pine”. Safranyik, L, and Wilson, B. (eds.) pp: 3-66.
 Amman, G.D., McGregor, M.D., and Dolph, R.E. Jr. 1990. Mountain pine beetle. Forest Insect & Disease Leaflet, 2. United States Department of Agriculture Forest Service. Retrieved from: http://www.barkbeetles.org/mountain/fidl2.htm.
 USDA Forest Service. 2010. Region 2 Forest Health Aerial Survey Highlights for 2010. Retrieved from: prdp2fs.ess.usda.gov/detail/r2/home/?cid=stelprdb5253133
 Wulder, M.A., Dymond, C.C., White, J.C., Leckie, D.G., and Carroll, A.L. 2006. Surveying mountain pine beetle damage of forests: A review of remote sensing opportunities. Forest Ecology & Management, 221: 27-41.
 USDA Forest Service FHTET. 2004. Aerial Detection Overview Surveys Futuring Committee Report. Forest Service Forest Health Technology Enterprise Team. FHTET-04-07.
 Chander, G., Markham, B.L., and Barsi, J.A. 2007. Revised Landsat 5 Thematic Mapper Radiometric Calibration. IEEE Geoscience and Remote Sensing Society, Manuscript (GRSL-00031-2007).
Editor’s Note: The DEVELOP National Program is a capacity building internship sponsored by NASA’s Applied Sciences Program that provides interns the opportunity to learn about NASA Earth Science and the practical applications of Earth observations.