WWETAC Projects

Project Title: Mapping forest composition and structure in the Pacific Coast States with gradient nearest neighbor imputation (GNN)

Status: Completed

Principal Investigator: Janet L. Ohmann, PNW Research Station, USDA Forest Service and the LEMMA Team (Landscape Ecology, Modeling, Mapping, and Analysis)

E-mail Contact: johmann[at]fs.fed.us
Web: http://www.fsl.orst.edu/lemma

Collaborators: Western Wildland Environmental Threat Assessment Center, USFS; Forest Science Department, Oregon State University (OSU); Interagency Mapping and Assessment Project (IMAP), USFS, Oregon Dept. of Forestry, BLM; Northwest Forest Plan Effectiveness Monitoring, USFS and BLM; Remote Sensing Applications Center, USFS; Forest Inventory and Analysis (FIA), USFS; Institute of Natural Resources, OSU

Summary: Information about the distribution and characteristics of forest vegetation and land cover is fundamental to a wide array of issues related to sustainable ecosystem management. Vegetation maps are needed to assess current landscape conditions of wildlife habitats and other biodiversity indicators, timber supply, carbon budgets, and fuel conditions and fire risk, to name a few. The gradient nearest neighbor (GNN) method for predictive vegetation mapping was conceived as a way to provide highly detailed spatial data on forest vegetation that is needed to support a wide variety of applications in research, landscape analysis, and natural resource and conservation planning at mid-scales. GNN integrates plot and spatial (GIS) data, including satellite imagery, to map detailed attributes of forest composition and structure across large, multi-ownership regions. The primary goal of this project is to develop detailed maps of forest composition and structure for Oregon and Washington using GNN. In addition, we are investigating related research questions on statistical methods for spatial prediction, environmental and disturbance factors influencing landscape patterns and dynamics, and scaling and linking of vegetation maps to models of stand and landscape dynamics for regional analysis.

We are using GNN to map forest and woodland vegetation of all ownerships across Oregon and Washington. GNN is one kind of "imputation" mapping, as are KNN and MSN, which is fundamentally different from other mapping methods. The main steps are (1) developing a statistical model, using direct gradient analysis, that predicts vegetation response variables from environmental and spectral (Landsat imagery) predictor variables; (2) using the gradient model to identify the "nearest-neighbor" plot for each map unit (pixel), and assigning the plot’s vegetation data to the pixel. The end result is a map wherein each pixel has a unique plot number assigned to it, along with all of its tree and understory vegetation measurements and summary variables. For plot data, we rely mostly on the regional forest inventory (Forest Inventory and Analysis [FIA] and CVS) and Ecology programs. We use over 50 geographic information system (GIS) layers that describe climate, topography, solar radiation, parent material, disturbance history, and ownership, as well as Landsat bands and transformations. The date of any given GNN map is the year of Landsat imagery used in its development. Maps of different dates can be generated by applying the model to new imagery. GNN models can be "tuned" to address different map objectives by modifying the input data and statistical model. We currently are producing two GNN map products for each ecoregion: one emphasizing species composition and one emphasizing a combination of species and forest structure. Accuracy assessment is an integral part of GNN. The maps are evaluated using a variety of methods, including cross-validation, and at both local and regional scales. The project began in October 2005, with work being organized around ecoregions. Draft GNN maps have been released for all Oregon ecoregions as of June 2007, with final versions expected in December 2007. Given funding, western Washington is planned to be completed by December 2008. See ‘modeling regions and schedule’ at http://www.fsl.orst.edu/lemma/gnnpac/ for more information.

In general, GNN map accuracy is excellent at the regional level and good to poor at the local scale, depending on vegetation attribute and location. The maps accurately represent areal proportions of different forest conditions across large regions, as well as the regional range of variability in vegetation attributes. At the local scale, the covariance structure of all vegetation components within a map unit is maintained, avoiding unrealistic co-occurrences of species and forest structures. However, local accuracy is limited by the inherent capabilities of the Landsat sensor. Canopy measures are mapped more accurately than subcanopy features, and accuracy tends to be better in west-side than in east-side forests. Although we lack independent data for accuracy assessment at the scale of small landscapes, we think the GNN data are reliable down to the level of 5th- and 6th-field hydrologic units.

Project ID: FY05JB5