Evaluating Ecoregions for Sampling and Mapping Land-cover Patterns

Ecoregional stratification has been proposed for sampling and mapping land-cover composition and pattern over time. Using a wall-to-wall land-cover map of the United States, we evaluated geographic scales of variance for nine landscapelevel and eight forest pattern indices, and compared stratification by ecoregions, administrative units, and watersheds. Ecoregions accounted for 65 percent to 75 percent of the total variance of percent agriculture and percent forest because dominant land-cover is included in ecoregional definitions. In contrast, ecoregions explained only 13 percent to 34 percent of the variance of the other seven landscape-level pattern indices. After accounting for differences in amount of forest, ecoregions explained less than 5 percent of the variance of the eight forest pattern indices. None of the stratifications tested would be effective mapping units for land-cover pattern because within-unit variance of land-cover pattern is typically two to four times larger than between-unit variance. Introduction Tradeoffs between wall-to-wall mapping and sampling to estimate tropical deforestation rates have been the topic of recent discussion. Tucker and Townshend (2000) showed that a 10 percent sample of Landsat sensor scenes was inadequate to estimate deforestation rates for individual countries in South America, and recommended wall-to-wall coverage to capture localized land-cover dynamics. In response, Czaplewski (2003) showed that despite localized dynamics, a 10 percent sample was adequate for larger geographic areas (biomes, continents) because the sample size was larger for a fixed sampling intensity. Clearly, a sampling approach can estimate deforestation for a given geographic area, to any specified degree of precision, if the sample size is large enough. A wall-to-wall map is simply a 100 percent sample that provides measurements for all geographic areas. The discussion of sampling versus wall-to-wall mapping should consider the type of information that is expected to come from remote sensing. Whereas a focus on estimating deforestation rates leads naturally to a statistical sampling approach, a focus on managing deforestation leads to a geographical mapping approach. It is one thing to know the deforestation rate, and another to know where to act on that information. Places that are not sampled still need to be managed, and an accurate map is required to decide actions Evaluating Ecoregions for Sampling and Mapping Land-cover Patterns Kurt H. Riitters, James D. Wickham, and Timothy G. Wade for specific places. Sampling is obviated by wall-to-wall mapping. At issue is whether a map prepared from a sample is accurate for a small area, and thus useful for local land management. There are several ways to produce a wall-to-wall map from a sample of locations. For example, geostatistical methods fit spatial surface models to sample measurements, which then provide interpolated estimates for the non-sampled locations. Local accuracy depends on a low “nugget” variance, which in turn requires a sufficiently dense sample in relation to the spatial correlation of the mapped attribute. A second approach uses geographically defined stratification whereby stratum-level estimates from sampling are mapped. In this case, local accuracy depends on spatial uniformity, that is, low within-stratum variance. Neither of these approaches is likely to solve the problem of mapping spatially concentrated deforestation rates from a small sample. In the United States, the cost and difficulty of developing wall-to-wall temporal land-cover data have led to a proposal for sampling and mapping based on ecoregional stratification (Loveland et al., 2002). Stehman et al. (2003a) demonstrated that the proposed stratified one-stage cluster design is effective for estimating population and ecoregionlevel changes in land-cover composition. Stratification is effective when strata are defined with respect to the quantity that is to be estimated (Cochran, 1977), and land-cover composition is one of the components used to distinguish the Omernik (1987) ecoregions employed in the study. At the same time, substantial within-sample and within-ecoregion variance (see Figure 6 in Gallant et al., 2004) led to recommendations to consider larger sample sizes, smaller sample units, post-stratification, and regression estimators to incorporate ancillary information (Stehman et al., 2003a). In contrast to land-cover composition, ecoregional stratification may be less effective for estimating changes in land-cover spatial pattern. Spatial patterns result from both local (e.g., urban expansion, parcelization) and regional (e.g., fire suppression, abandonment of agriculture) management regimes. Even if the ecoregional environment determines if there can be a farm at all (composition), individual humans decide the size and shape (pattern) of farms. These decisions are made in a local context that also includes the history of a landscape, local traditions, and economic and regulatory criteria (Masek et al., 2000; McDonald and Urban, 2004). Thus, while ecoregional stratification is effective for estimating land-cover composition, more work is needed to prove its utility for sampling and mapping land-cover patterns. PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING J u l y 2006 781 Kurt H. Riitters is with the USDA Forest Service, 3041 Cornwallis Road, Research Triangle Park, NC 27709 ([email protected]). James D. Wickham and Timothy G. Wade are with the U.S. Environmental Protection Agency, MD 243-05, Research Triangle Park, NC 27711 ([email protected]; [email protected]). Photogrammetric Engineering & Remote Sensing Vol. 72, No. 7, July 2006, pp. 781–788. 0099-1112/06/7207–0781/$3.00/0 © 2006 American Society for Photogrammetry and Remote Sensing 05-019 6/9/06 9:58 AM Page 781