Document: JAN-3-40-28

Predictive mapping of forest composition and structure using direct gradient analysis and nearest neighbor imputation.

OHMANN, J.L.* ¹ and M.J.GREGORY ²

USDA Forest Service, Corvallis, OR 97331 USA ¹
Oregon State University, Corvallis, OR 97331 USA ²

Abstract:
Spatially explicit information on the species composition and structure of forest vegetation is needed at broad spatial scales for ecological research, biodiversity assessment, and policy analysis. Satellite remote sensing has been successfully used to map coarse attributes of vegetation, but elucidating fine-resolution attributes such as species identities or understory canopies is a more difficult problem. We developed and tested a GIS-based, predictive modeling approach that integrates vegetation measurements from regional networks of field plots, mapped environmental data, and Landsat TM data to characterize forest vegetation across a region. The method applies direct gradient analysis and nearest neighbor imputation to ascribe detailed ground attributes of vegetation to each patch in a regional landscape. When applied to the coastal physiographic province of Oregon, USA, the mapped predictions maintained the covariance structure among multiple response variables (relative abundances of tree species and sizes), maintained the range of variability present in the plot data, and portrayed spatial heterogeneity in an ecologically realistic way. Gradients in species composition were most strongly associated with regional climate and geographic location, whereas forest structure was best predicted by Landsat TM data. Elevation, topography, and geology also explained significant amounts of variation. Model accuracy was excellent at the regional scale, and only moderate at the local scale, as expected. Correlations between predicted and observed values for specific sites were 0.49 for total stand basal area and 0.59 for stand quadratic mean diameter. Prediction accuracy for species occurrence varied widely, depending on the rarity of the species and how strongly its distribution is controlled by the available explanatory measures. Model predictions are now being used to initialize landscape conditions for simulation modeling of alternative scenarios for land management policy, and for assessing ramifications for ecological and economic response variables.

Keywords: predictive vegetation mapping, gradient analysis, imputation, landscape ecology, Landsat TM, vegetation composition and structure, biodiversity, Oregon

This abstract is being presented at: 10:30 AM in session:
Poster Session #5: Landscape Ecology.