The Infl uence of Multi-season Imagery on Models of Canopy Cover: A Case Study

Quantifying tree canopy cover in a spatially explicit fashion is important for broad-scale monitoring of ecosystems and for management of natural resources. Researchers have developed empirical models of tree canopy cover to produce geospatial products. For subpixel models, percent tree canopy cover estimates (derived from fi ne-scale imagery) serve as the response variable. The explanatory variables are developed from refl ectance values and derivatives, elevation and derivatives, and other ancillary data. However, there is a lack of guidance in the literature regarding the use of leaf-on only imagery versus multi-season imagery for the explanatory variables. We compared models developed from leaf-on only Landsat imagery with models developed from multi-season imagery for a study area in Georgia. There was no statistical difference among models. We suggest that leaf-on imagery is adequate for the development of empirical models of percent tree canopy cover in the Piedmont of the Southeastern United States. Introduction Tree canopy cover is a primary component of ecosystems and is defi ned as the area covered by the vertical projection of tree crowns (Jennings, 1999). The amount and density of cover infl uences habitat suitability, fi re behavior, aesthetics, and carbon dynamics. For example, Rollins and Frame (2006) used a map of percent tree canopy cover (Homer et al., 2007) as a major component in their forest fi re behavior and fuel models. Tree canopy cover is also a critical component of forest management activities (Jennings, 1999). Additionally, both forest land use defi nitions and forest land cover defi nitions are partially based on the amount of tree canopy cover present during the time of classifi cation. For example, the defi nition of forest land cover used in the National Land Cover Database (NLCD) land cover mapping effort partially relies on identifying areas with at least 20 percent tree canopy cover (Homer et al., 2007). Likewise, the United Nations Food and Agriculture Organization (FAO) defi nition of forest land use partially relies on identifying areas with at least 10 percent tree canopy cover (FAO, 2001). Because of the importance of tree canopy cover, a national map of percent tree canopy cover, across all lands, was developed as part of the 2001 NLCD (Huang et al., 2001, Homer et al., 2007) and a 2011 version is under development (Coulston et al., 2012). The 2001 NLCD percent tree canopy cover product is a freely available 30 m dataset. Because percent tree canopy cover is not calculable from Landsat imagery directly, empirical models were developed to predict percent canopy cover US Forest Service, 4700 Old Kingston Pike, Knoxville, TN 37919 ([email protected]). Photogrammetric Engineering & Remote Sensing Vol. 79, No. 5, May 2013, pp. 469–477. 0099-1112/13/7905–469/$3.00/0 © 2013 American Society for Photogrammetry and Remote Sensing at unmeasured locations. In this case, the response variable was derived from classifying 1 m Digital Orthophoto Quarter Quadrangles (DOQQs) as either tree canopy or no tree canopy. Approximately 1 to 4 km2 per Landsat scene were purposively sampled and classifi ed using a classifi cation tree (Homer et al., 2004). The response, percent tree canopy cover, was then calculated on a 30 m pixel level for the sampled area. Multi-season (leaf-off, spring, leaf-on) Landsat-5 and -7 data and indices (e.g., tasseled cap), along with digital elevation models and derivatives (e.g., slope), and other ancillary information (e.g., 1992 NLCD) were used as the explanatory variables. Empirical models of percent tree canopy were then developed using regression trees based on the relationship between the response and explanatory variables. The current effort to produce a 2011 NLCD percent tree canopy product (Coulston et al., 2012) relies on a probabilistic sampling approach where a two stage sample is employed. The locations of the primary sampling units (PSUs) were identifi ed based on a global sampling grid (White et al., 1992), and within each PSU a systematic dot grid (105 points) covers a 90 m by 90 m area. Each point within the PSUs was classifi ed using photographic interpretation of leaf-on 1 m true color or false color imagery provided by the National Agriculture Imagery Program (NAIP). This photointerpretation technique was similar to that used by Carreiras et al. (2006). The percent canopy cover estimates of each PSU then served as the response variable for empirical model development. The explanatory variables were leaf-on Landsat-5 imagery and derivatives, elevation and derivatives, and other ancillary data such as the 2001 NLCD land cover map. There are several notable differences between the approach used to develop the 2001 NLCD percent tree canopy cover map and the approach for the 2011 NLCD percent tree canopy cover map. The scope of this research is not to compare and contrast all the differences between the 2001 and 2011 NLCD approaches. It is rather to examine whether empirical models of leaf-on percent tree canopy cover are improved by using multi-season Landsat imagery as opposed to only leaf-on Landsat imagery as explanatory variables. The available literature is comprised of examples suggesting that the use of multi-season imagery is appropriate, and others suggesting that using only single season imagery is appropriate for this type of application. For example, FrancoLopez et al. (2001) used multi-season imagery to map forest stand density, volume, and cover type in St. Louis County, Minnesota. Hansen et al. (2003) used 40-day Moderateresolution Imaging Spectroradiometer (MODIS) composites 469-477_12-024.indd 469 4/13/13 10:16 PM M a y 2 0 1 3 PHOTOGRAMMETR IC ENGINEER ING & REMOTE SENS ING 470 Land cover, based on Homer et al. (2007), in this ecoregion was 22 percent urban, 13 percent agriculture, and 56 percent forest cover. Much of the urban area was part of the Atlanta, Georgia metropolitan area. The Blue Ridge ecoregion covers 19 percent of the study area and was 6 percent urban cover, 5 percent agriculture cover, and 85 percent forest cover. A small percentage of the study area was classifi ed in the Ridge and Valley ecoregion (3 percent) and the Southeastern Plains ecoregion (1 percent). The Ridge and Valley ecoregion was 14 percent urban, 22 percent agriculture and 54 percent forest cover. The Southeastern Plains ecoregion was 5 percent urban, 21 percent agriculture, and 59 percent forest cover. Percent tree canopy cover was estimated for 4,125 sample locations (PSUs) across the study area and these estimates served as the response data. Sample locations were identifi ed based on a 4X intensifi cation of the USDA Forest Service Forest Inventory and Analysis sampling grid using the procedures described by White et al. (1992). At each PSU, a 105 point triangular-grid that fi lled a 90 m by 90 m (0.81 ha) area served as the basis for photo-interpretation (Figure 1). Each of the 105 points was manually interpreted as either “tree canopy” or “no tree canopy” using leaf-on 2009 NAIP (USDA, over the course of one year to model global tree canopy cover at 500 m. Alternatively, Carreiras et al. (2006) used leaf-on Landsat imagery to model tree canopy cover of evergreen oak woodlands on the Iberian Peninsula. Sen et al. (2011) used leaf-on imagery to quantify percent tree canopy cover of mined lands in the Appalachian Mountains in the southeastern United States. Clearly there are varying viewpoints on whether to use leaf-on or multi-season imagery for developing empirical models of percent tree canopy cover. The objective of this research is to test whether the inclusion of multi-season imagery as an explanatory variable signifi cantly improves empirical models of percent tree canopy cover and to provide some guidance on where our results are relevant. Methods The study area was approximately the size of one Landsat scene covering central and northern Georgia in the southeastern United States (Figure 1). While the area was one Landsat scene in size, it covered path-rows 19-36 and 19-37 and was specifi cally selected to capture the south to north environmental gradient. The Piedmont was the dominant (77 percent) ecoregion (USEPA, 2011) in the study area. Figure 1. (a) Georgia study area, and (b) and (c) sample design. One of the four systematic samples (PSU) is displayed in (b). For each PSU a 105 point secondary sampling unit was used for photo interpretation of