WWETAC Projects

Project Title:  Using lidar to improve the reliability of GNN vegetation and fuels maps for forest management and risk assessment

Principal Investigator:   Janet Ohmann, US Forest Service, PNW Research Station

johmann[at]fs.fed.us

Collaborators:  Harold Zald, Oregon State University; Robert McGaughey, PNW Research Station; Heather Roberts, Oregon State University; Mike Simpson, Central Oregon Ecology & Forest Health Protection Programs, Deschutes National Forest.

Key Issues/Problem Addressed:              

Ideally, vegetation and fuels information to assist fire and fuel managers should: 1) integrate data on fuels, species composition, and forest structure in a common format, 2) be spatially explicit (i.e. mapped), 3) accurately portray fuels and vegetation with a small enough grain size to be useful for tactical (i.e. stand-level) decision making, 4) at the same time cover a large landscape or region to be useful for more strategic planning and decision making, and 5) provide information about the type, magnitude, frequency, and location of errors in each dataset. Unfortunately, information with all of these characteristics does not currently exist. The GNN vegetation maps that currently are available for most of the Pacific Northwest are based on Landsat satellite imagery and thus are appropriate for "mid-scale" applications in forest planning and policy, but are not considered sufficiently reliable for use in stand-level management.  

Setting and Approach:  

We are mapping vegetation and fuels over the lidar acquisition area (about 1.2 million acres) of interior dry forests of the eastern Cascades of Oregon using GNN. The GNN method uses multivariate direct gradient analysis to quantify relations among FIA plot data, satellite imagery, and mapped environmental variables. Individual pixels in resulting GNN maps are associated with plots that have the most similar spectral and environmental characteristics. A suite of detailed plot variables is imputed to each pixel, allowing simultaneous and consistent predicting of a wide range of vegetation attributes. Because the plot-level variables are attached to the GIS database, a wide array of summary variables and classifications can be portrayed to meet different objectives. This study will incorporate rasters describing the forest canopy (overall height, variability in height, and cover) derived from lidar point clouds, which will be developed by Bob McGaughey. We will quantify the improvement in map accuracy achieved through the addition of lidar data (compared to models which use Landsat, climate, and other spatial predictors).

Progress to Date:

Proposed objectives are being addressed as described in the original proposal. After delays in acquiring GPS plot locations from the Forest and Landsat imagery, initial GNN models incorporating precision GPS locations of plots, lidar data, and Landsat imagery have been run. Results from initial models were presented as a poster at the ForestSat conference held on September 11-14 at Oregon State University in Corvallis, Oregon. A copy of the poster can be viewed online at: http://www.fsl.orst.edu/lemma/pubs. Initial findings suggest inclusion of lidar data and precision GPS plot locations may not improve GNN prediction accuracy using our current hybrid modeling approach, where species composition and structure occupy one response matrix used in the gradient analysis. We believe the hybrid models may have reduced prediction accuracy because they are compromises between large-scale gradients such as climate and topography that describe species composition, and small-scale gradients such as lidar vegetation metrics and Landsat-derived metrics of past disturbance that describe vegetation structure. We are now re-running all model combinations, but with separate species and structure models to improve prediction accuracy. We also found that decreasing spatial grain of field plots (from full plots to subplots) decreases prediction accuracy, contrary to expectations. This may also be the result of using the hybrid composition and structure response matrix, but could also suggest subplots have too small of a footprint to adequately relate field-measured attributes to lidar and/or Landsat imagery. Uncertainty regarding the cause of prediction accuracy declining with smaller spatial grain of field data should be resolved by the re-running of GNN models using separate composition and structure models.

Impacts/Applications:  

We had anticipated that including lidar and precise FIA plot locations would greatly improve the accuracy and utility of GNN maps for local forest management for a variety of business needs.  We still believe this is possible but with a couple caveats.  First, we expect the addition of lidar and precise plot locations will improve accuracy and utility of GNN vegetation structure maps, but not necessarily GNN species composition maps.  Second, we believe the GNN maps based on plot-level field data will be a large improvement over existing GNN products, but the finer spatial grain of sub-plot level data may not improve GNN maps.  Improved GNN maps will provide resource managers within the study area unparalleled regional data, synthesizing fuels and vegetation information in one consistent data format to inform both tactical and regional fuels management. The strength of previous GNN products for regional assessments has been tempered by their limited applicability to stand-level decision making. If successful, the improved GNN map products will be a significant step towards providing natural resource managers with tactic-level information across large landscape and regions. This study would serve as a proof-of-concept for using lidar to downscale nearest-neighbor maps for local management, and our methods could be applied anywhere lidar data are available.

WWETAC ID:      FY11NG97