Effects of changing landscape pattern and USGS land cover data variabilityon ecoregion discrimination across a forest-agriculture gradient

We examined the use of coarse resolution land cover data (USGS LUDA) to acc urately discriminate ecoregions and landscape-scale features important to b iodiversity monitoring and management. We used land cover composition and l andscape indices, correlation and principal components analysis, and compar ison with finer-grained Landsat TM data, to assess how well LUDA data discr iminate changing patterns across an agriculture-forest gradient in Minnesot a, U.S.A. We found LUDA data to be most accurate at general class levels of agriculture and forest dominance (Anderson Level I), but inconsistent and limited in ecotonal areas of the gradient and within forested portions of t he study region at finer classes (Anderson Level II). We expected LUDA to over-represent major (matrix) cover types and under-rep resent minor types, but this was not consistent with all classes. 1) Land c over types respond individualistically across the gradient, changing landsc ape grain as well as their spatial distribution and abundance. 2) Agricultu re is not over-represented where it is the dominant land cover type, but fo rest is over-represented where it is dominant. 3) Individual forest types a re under-represented in an open land matrix. 4) Within forested areas, mixe d deciduous-coniferous forest is over-represented by several orders of magn itude and the separate conifer and hardwood types under-represented. Across gradual, transitional agriculture-forest areas, LUDA cover class dominance changes abruptly in a stair-step fashion. In general, rare cover types tha t are discrete, such as forest in agriculture or wetlands or water in fores t, are more accurately represented than cover classes having lower contrast with the matrix. Northward across the gradient, important changes in the p roportions of conifer and deciduous forest mixtures occur at scales not dis criminated by LUDA data. Results suggest that finer-grained data are needed to map within-state ecoregions and discriminate important landscape charac teristics. LUDA data, or similar coarse resolution data sources, should be used with caution and the biases fully understood before being applied in r egional landscape management.