Restoration and maintenance of aspen communities is a BLM priority in the Little Mountain Ecosystem, and the USGS has been working with the BLM and the WGFD to monitor aspen stands in that area as part of its WLCI Effectiveness Monitoring work. LANDFIRE and ReGAP maps are considered the best spatial products for representing aspen distribution at regional and landscape scales; however, these products were not designed to support decisions at localized scales, such as that of the Little Mountain Ecosystem. In 2010, this study filled a critical information gap with production of a model (fine-scale map) that delineates aspen distribution for the Little Mountain Ecosystem. To accomplish this, we used classification and regression tree (CART) analysis applied to uncompressed National Agriculture Imagery Program (NAIP) color-infrared imagery. (Imagery is compressed to reduce the image file sizes, but this can introduce error and a loss of data. Furthermore, uncompressed imagery is necessary for machine analysis used by remote-sensing software.) Although our effort was successful, we identified several shortfalls in the method that would make it untenable for mapping aspen stands across large areas, such as the entire WLCI study area. The two primary limitations of the CART method developed in 2010 were (1) the amount of noise associated with the fine-scale resolution (1 m) of the imagery, and (2) the large sizes of input files, which limits the transferability of this method at the landscape scale. To address these limitations, we modified the methodology developed in 2010 and applied it to SPOT 5 imagery (10-m resolution).
The modified method entailed using photographs and stand-evaluation data collected in 2010 (July) and 2011 (June and August) in the Little Mountain Ecosystem to inform the development of data we used to “train” the model. Two cloud-free SPOT 5 scenes of the Little Mountain area, preprocessed to level 1T (terrain-corrected data), were obtained from the USGS EarthExplorer archive. We then conducted a temporal analysis of summer imagery (aspen leaf-on) and autumn imagery (aspen leaf-off) to improve delineation of deciduous aspen trees from coniferous species, such as subalpine fir (Abies lasiocarpa) and Douglas-fir (Pseudotsuga menziesii). The leaf-on and leaf-off scenes had been acquired on September 7, 2010 and October 19, 2010, respectively. The imagery was georectified (root mean square error less than 0.5 pixel) and converted to Top-of-Atmosphere reflectance. This ensured proper radiometric calibration between the images, which is necessary for creating the high-quality data series needed for change-detection analysis (Chander and others, 2009).
To mask out non-woodland areas from the analysis, a woodland mask was developed from Landsat-derived NDVI representing the peak of the growing season (July 2010). For the training dataset, more than 250 points were randomly generated in the mapped woodland area, with an enforced minimum distance of 50 m between each point. Each point was manually classified as aspen, conifer, or non-woodland based on field data and aerial photos. For each training point, a series of covariates was generated from the satellite imagery and other biophysical features. CART was employed using the TREE package in the R Statistical Package (R Development Core Team, 2011). The fitted tree contained 15 nodes and had an overall accuracy of 89.9 percent. However, CART tends to generate complex and over-fit models (Brown and others, 2006), so a cross-validation procedure (prune tree function in R) was used to select a model that maximized the deviance explained while minimizing the misclassification error rate.
The final model had only 4 nodes but an overall accuracy of 86 percent. A smoothing algorithm (3 × 3 window) was applied to the data to remove isolated pixels and improve the interpretability of the map. Visual inspection of the classified map revealed that small, irrigated meadows were misclassified as aspen (seasonal reflectance values for aspen and irrigated meadows are similar); thus, aspen cover likely would be overestimated (false positive) if this methodology is applied in other areas. A new model was developed by using image texture as an additional covariate; however, this did not improve the classification, so we used expert knowledge of the study area to reclassify the irrigated meadows appropriately in the final map. The output map was converted from a raster dataset to the vector environment to ease usability for WLCI partners. Several attributes were calculated for each contiguous woodland patch type, including patch area and perimeter. We have distributed the dataset to our partners and will work with them to ensure proper interpretation and usability of the product.
Chander, G., Markham, B.L., and Helder, D.L., 2009, Summary of current radiometric calibration coefficients for Landsat MSS, TM, ETM+, and EO-1 ALI sensors: Remote Sensing of Environment, v. 113, no. 5, p. 893-903.
R Development Core Team, 2011, R: A language and environment for statistical computing: Vienna, Austria, R Foundation for Statistical Computing, at http://www.r-project.org/.
Brown, K., Hansen, A.J., Keane, R.E., and Graumlich, L.J., 2006, Complex interactions shaping aspen dynamics in the Greater Yellowstone Ecosystem: Landscape Ecology, v. 21, no. 6, p. 933-951.
Products Completed in FY2011
- Final map of aspen and conifer stands in Little Mountain Ecosystem to provide baseline information on their size, location, and so on.
- Map of Sierra Madre area of interest with contrast stretch applied to aerial photography to improve interpretation and delineation of aspen and conifer forest.
- Map used to support other tasks, such as informing sampling design for sample site selection.
- Methodology and covariates completed.
- Final map used to inform BLM fire-treatment map.
Products Completed in FY2010
- Draft classification map of aspen vegetation on Little Mountain to provide baseline information on locations of aspen stands.
- Map used to support other tasks, such as informing sampling design for selecting sampling sites.
- Photo geodatabase, including location and direction of photos, created as part of groundtruthing effort.
- Methodology and covariates completed.
- Used to inform BLM fire-treatment map.
- Draft classification map of aspen vegetation on Little Mountain to provide baseline information on locations of aspen stands.
- Map used to support other tasks, such as informing sampling design for selecting sampling sites.
- Photo geodatabase, including location and direction of photos, created as part of groundtruthing effort.
- Methodology and covariates completed.
- Used to inform BLM fire-treatment map.
Products Completed in FY2009
- Geospatial map of study area with WLCI-funded aspen treatments shown.
- Geospatial database and documentation of Landsat data and ancillary GIS data.
- Ground-truthed data set of aspen and conifer stands from aerial photography and field investigation.
