This article is a part of the NASA DEVELOP’s Spring 2017 Article Session. For more articles like these, click here
This study used remote sensing and distribution modeling techniques to map aspen cover in a remote mountain range in Southeast Wyoming to support elk and mule deer habitat management.
Quaking aspen (Populus tremuloides) is one of the most widely distributed tree species in North America, and thrives in areas that experience cold winters and where precipitation is greater than evapotranspiration rates (1). In the Intermountain West, aspen dominated stands are rich in a variety of vertebrates and understory vegetation, making them ecologically valuable (2), (3). In particular, aspen stands provide critical habitat for mule deer (Odocoileus hemionus) and elk (Cervus canadensis) as both ungulate species rely on aspen stands for forage, parturition, and cover from predators (2), (4).
Current threats to aspen include climate change, fire suppression, herbivory, and natural encroachment upon aspen stands by conifers (1), (5). Landscape scale die-offs of aspen stands linked to drought conditions have been recorded throughout western North America within the past decade (5), (6). Additionally, continuous and intensive browsing by domestic and wild ungulates can degrade aspen habitat, especially with respect to aspen regeneration (4). Maintaining aspen habitat is a high priority for natural resource managers, particularly in southeastern Wyoming where ungulate species are dependent on mature aspen communities for forage and fawning habitat. However, aspen distribution and cover data can be limited and incomplete, making it difficult for wildlife managers to determine the existing extent, location, and condition of critical habitat, such as aspen dominant forest.
To address this knowledge and geospatial data gap, this project was conducted in partnership with the Wyoming Game and Fish Department (WGFD) to assist in efforts to locate and manage aspen habitat. We focused on the Laramie Peaks Unit of the Medicine Bow National Forest in southeastern Wyoming. The Laramie Mountains are remote and rugged, with elevations ranging from 6,200-10,500 feet. Dominant habitat types include native rangeland, mountain shrub steppe, and ponderosa pine (Pinus ponderosa), and lodgepole pine (Pinus contorta) forests with intermixed aspen stands (7) (Image A). We traveled to the Laramie Mountains to collect aspen presence data, and used these data in conjunction with NASA Earth observations to develop three species distribution models (SDMs) to map current aspen distribution on a local scale (30-meter resolution).
Satellite imagery from Landsat 8 Operational Land Imager and Thermal Infrared Sensor (OLI/TIRS) was acquired from the U.S. Geological Survey (USGS) Earth Resources Observation and Science (EROS) Center Science Processing Architecture (ESPA) on-demand interface (8). We selected Landsat 8 terrain corrected surface reflectance data with less than 10 percent cloud cover for path 34, row 31 for the summer and fall months of 2014 and 2015, and the summer months of 2016. Due to limitations in image quality resulting from cloud cover over the study area, two final images with minimal cloud cover captured June 21, 2016, and Oct. 22, 2014, were selected for analysis. In addition to Landsat bands 2-7, two surface reflectance derived vegetation indices were selected: NDVI (Normalized Difference Vegetation Index), and EVI (Enhanced Vegetation Index). Shuttle Radar Topography Mission (SRTM) data products were accessed through the USGS Earth Explorer portal to obtain a 30-meter, void-filled Digital Elevation Model (DEM) for the study area.
Because field data on aspen presence was largely unavailable for the study area, it was necessary to collect field data on aspen locations to complete this project. To plan efficient field sampling in the study area, we employed existing, but limited, aspen presence data (n = 36) sourced from collaborators at the Natural Resource Ecology Laboratory (NREL) and spatially fuzzed U.S. Forest Service Forest Inventory and Analysis (FIA) plots (n = 200) collected between 2000 and 2008 to fit a preliminary aspen distribution model using MaxEnt (9). We used two vegetation indices derived from Landsat 8 bands from a previously processed June 2015 image, and topographic data as predictor variables to fit the MaxEnt model with the combined aspen presence points. The MaxEnt output maps were then refined to pixels classified as greater than 41 percent probability of aspen presence (the equal training threshold determined automatically through MaxEnt) and this raster was used to stratify random field sampling point locations. Field surveys were conducted by the team from June 20-23, 2016, to collect additional aspen data (n = 21) for training species distribution models. The survey was designed as a stratified random sample and adapted from FIA protocol (10).
Field data collected from the Laramie Mountains were used to train the aspen distribution models. Aspen presence and absence points were characterized by deriving relative percent cover estimates from densiometer readings completed in the field. Plots with greater than 30 percent aspen cover were considered presence points, and plots with less than 5 percent aspen cover were considered absence points. This allowed us to maximize the quantity of sample points while considering the ability for the model to correctly identify aspen presence and absence at the 30m spatial resolution of a Landsat pixel. These data were supplemented with presence data points from a previous survey conducted by partners at NREL, Colorado State University in 2015 where all points had estimated aspen cover of 90-100 percent. This resulted in a total of 36 aspen presence points and 27 absence points (n = 63). A DEM, plus data from Landsat 8 OLI/TIRS and Landsat-based derived indices were to be used as predictors in our aspen distribution models.
The final aspen distribution models were constructed using The Software for Assisted Habitat Modeling (SAHM) open-source spatial modeling software (11). Landsat bands, vegetation indices, and topographic indices in raster format were used as predictor variables in conjunction with the 63 presence and absence points for the SDMs fit within the SAHM program. To supplement the aspen cover model that relied on a single image from June 21, 2016, an additional model was developed that included spectral data from an Oct. 22, 2014, image with the objective of determining if a spectral signature for aspen could be derived during the fall season when aspen leaves were likely yellow.
Data processing was conducted within SAHM to clip and re-project all predictor layers to match a defined template raster (i.e., the extent of the study area), and values from each predictor layer were extracted and assigned to the corresponding aspen absence or presence point. SAHM allows for pre-modeling covariate evaluation by displaying results from a univariate model fit with each predictor and a predictor correlation matrix. Predictors were retained based on ecological relevance and contribution to the univariate model, and if any two predictors had a correlation coefficient of |r| > 0.70, one of the two was eliminated as highly correlated predictors can create “unstable model fits” (11). The Random Forest (RF) classifier was selected to fit the final aspen cover models because it repeatedly outperformed other models tested based on evaluation metrics. Data were divided into cross validation folds for modeling and model evaluation (i.e., 80 percent for training and 20 percent for testing), and subsequently the validity and predictability of the models were assessed using threshold independent and threshold dependent evaluation metrics produced by SAHM.
Results and DiscussionThe preliminary MaxEnt 1 model performed reasonably well in distinguishing aspen presence, with elevation and NDVI as the most important predictors based on the Area Under the receiver operating characteristic Curve (AUC) metric (AUCCV=0.77, AUCtrain– AUCcv = 0.05) (Table 1). The final RF 1 model of aspen distribution with Tasseled Cap Greenness, elevation, Landsat blue band (June 21, 2016), and aspect as predictors best identified areas in the Laramie Mountains with significant aspen cover and areas with no aspen cover (AUCtrain-AUCCV= 0.01), PCC=79.24 (Image B). We further evaluated our models based on the percent correctly classified (PCC), which represents the percent of true presence and absence (or background) points correctly classified, sensitivity, and specificity. The RF 2 model with June and October spectral data reported a higher sensitivity (the proportion of observed presences predicted to be presence points by the model), and lower specificity (the proportion of observed absences predicted to be absences by the model or the proportion of true negatives), and a greater AUC difference (AUCtrain– AUCcv=0.04) than the RF1 model.
Areas that were classified by the preliminary aspen model to have greater than 41 percent probability of aspen presence were used to guide sampling, thus the performance of this model was evaluated during the field survey. The accuracy of the model was reasonable based on evaluation metrics. While it generally over-predicted aspen presence, this model was undeniably useful as a guide to locate aspen plots in the field across a large, remote area. This unique approach is being considered for future aspen field surveying efforts by the Wyoming Game and Fish Department.
The final aspen distribution model (RF 1) predicted aspen cover well and reported the strongest overall evaluation metrics, and after a visual inspection of probability map outputs overlaid with high-resolution imagery, it was determined to be marginally more accurate than RF 2. One potential reason why the RF2 model (utilizing June and October 2014 data) did not superiorly identify aspen cover compared to RF1 was that the trees captured in the October imagery were already in a leaf-off stage. However, the RF2 model outperformed the RF1 model in terms of the other accuracy assessment metrics given in Table 1. The visual inspection revealed some areas where RF 1 incorrectly classified absence as presence, and aspen cover was not distinguished from other deciduous vegetation such as willow (Salix spp.). This was most common in areas where moisture collects across the landscape such as riparian habitat, and was likely due to the relatively small number of absence points that we sampled during the field data collection.
The project results indicated that it was technically feasible to produce useful maps of aspen forest using Landsat data. However, more refined aspen forest maps may be obtained with additional work. Based on the results of this study, we recommend using more ideally timed multi-date images that capture variation in greenness of aspen forests throughout the growing season, as well as the “yellow” phase of aspen. Such data could be used to improve accuracy of models and highlight aspen cover across a landscape. The final map product produced by the project is being used by local natural resource managers to help estimate of the carrying capacity for mule deer and elk populations in the region, and increase available geospatial information of aspen density and distribution (Image C).
 Lindroth, R.L., & St. Clair, S.B. Adaptations of quaking aspen (Populus tremuloides Michx.) for defense against herbivores, in Forest Ecology and Management, vol. 299, 2013, pp. 14-21. doi:10.1016/j.foreco.2012.11.018
 Reynolds, H.G. Aspen grove use by deer, elk, and cattle in southwestern coniferous forests, in USDA Forest Service Rocky Mountain Forest and Range Experiment Station Research Note RM 138, 1969.
 Turchi, G.M., Kennedy, P.L., Urban, D., & Hein, D. Bird species richness in relation to isolation of aspen habitats, in The Wilson Bulletin, vol. 107, no. 3, 1995, pp. 463-474.
 Seager, S.T., Eisenberg, C., & St. Clair, S.B. Patterns and consequences of ungulate herbivory on aspen in western North America, in Forest Ecology and Management, vol. 299, 2013, pp. 81-90. doi:10.1016/j.foreco.2013.02.017
 Rehfeldt, G.E., Ferguson, D.E., & Crookston, N.L. Aspen, climate, and sudden decline in the western USA, in Forest Ecology and Management, vol. 258, 2009, pp. 2353-2364. doi:10.1016/j.foreco.2009.06.005
 Frey, B.R., Lieffiers, V.J., Hogg, E.H., & Landhausser & S.M. Predicting landscape patterns of aspen dieback: mechanisms and knowledge gaps. Canadian Journal of Forest Research, vol. 34, 2004, pp. 1379-1390.
 Wirsing, John M., and R. R. Alexander. Forest Habitat Types on the Medicine Bow National Forest, Southeastern Wyoming: Preliminary Report, 1975. Accessed 13 June 2016. http://hdl.handle.net/2027/uiug.30112104053647
 USGS. (2016). “LSDS Science Research and Development (LSRD).” US Department of the Interior,Geological Survey, May 2010. 15 June 2016. http://espa.cr.usgs.gov/
 Phillips S. J., Anderson R.P., & Schapire R.E. Maximum entropy modeling of species geographic distributions, in Ecological Modelling, vol. 190, 2006, pp. 231–259. doi:10.1016/j.ecolmodel.2005.03.026
 USFS. Forest inventory and analysis national core field guide volume I: Field data collection procedures for phase 2 plots, 2015. http://www.fia.fs.fed.us/library/fact-sheets/data collections/Phase2_3.pdf
 Morisette, J. T., Jarnevich, C. S., Holcombe, T. R., Talbert, C. B., Ignizio, D.Talbert,M.K., Young, N. E. VisTrails SAHM: Visualization and workflow management for species habitat modeling, in Ecography, vol. 36, no.2, 2015, pp. 129-135. doi:10.1111/j.1600-0587.2012.07815.x
Sarah Carroll is a recent graduate from University of California Davis working with DEVELOP at the Fort Collins node as an independent research consultant.
Katie Butler is a master’s geography graduate of San Diego State University who is working at Digital Globe in Colorado.
Aubrey Hilte is a recent graduate from Colorado State University working with DEVELOP at Wise County, Virginia, as an independent research consultant.
Megan Vashen is a master’s student of ecology from Colorado State University, graduating this spring.
Amanda West is a postdoctoral fellow at the Natural Resource Ecology Laboratory at Colorado State University, a former DEVELOP participant, and current science adviser to the Fort Collins node.
Brian Woodward is a nine-term DEVELOP participant who has managed the Fort Collins node for the past two years. He is a current graduate student in forest ecology at Colorado State University.
Paul Evangelista is a research ecologist at the Natural Resource Ecology Laboratory at Colorado State University. He serves as the lead science adviser for the Fort Collins node.