Mapping Mountain Hemlock and Pinpointing Ponderosa Pine: Imputation Mapping Species Distributions in Western Oregon
Event Website
http://www.nafew2009.org/
Start Date
6-22-2009 12:00 AM
End Date
6-26-2009 12:00 AM
Description
Effective forest management and conservation planning depends upon having accurate and detailed vegetation maps. Remote sensing has been an invaluable tool for mapping vegetation, but imagery alone often cannot yield detailed information on individual tree species. Because of this mismatch, there is interest in developing new statistical techniques for integrating ancillary information with remote sensing to produce a map. As part of a nationwide study, we are evaluating several variants of imputation mapping. Here, we compare the ability of two methods (gradient nearest neighbor (GNN) and random forest (RFNN)) to produce maps of tree species distributions at regional scales. These two methods make different assumptions about species-environment relationships and thus have differing strengths and weaknesses. We build the imputation models from 1778 vegetation plots from the US Forest Service’s Forest Inventory and Analysis program (annual plots), and a set of spatial data: a 2006 Landsat image mosaic, modeled climate (PRISM), and topographic variables, for a region that stretches from the Northern Oregon border into California, and encompasses the Cascade Mountains. Preliminary results show that both methods often predict species occurrence at somewhat broader geographic ranges, and a broader range of environmental conditions than they do within our original plot sample. Gradient nearest neighbor is more effective at producing range maps which represent the total area covered by a given species. Random forest predictions track species’ environmental limits more closely with the original data. However, some species are overmapped within their ranges, giving inflated estimates of total area within the mapped region. This tendency also yields different image textures. GNN produces a grainier map with higher edge densities while RFNN produces more compact patches. We conclude that model choice for mapping applications should be based upon intended use of the final map product, because each method has both strengths and weaknesses.
Mapping Mountain Hemlock and Pinpointing Ponderosa Pine: Imputation Mapping Species Distributions in Western Oregon
Effective forest management and conservation planning depends upon having accurate and detailed vegetation maps. Remote sensing has been an invaluable tool for mapping vegetation, but imagery alone often cannot yield detailed information on individual tree species. Because of this mismatch, there is interest in developing new statistical techniques for integrating ancillary information with remote sensing to produce a map. As part of a nationwide study, we are evaluating several variants of imputation mapping. Here, we compare the ability of two methods (gradient nearest neighbor (GNN) and random forest (RFNN)) to produce maps of tree species distributions at regional scales. These two methods make different assumptions about species-environment relationships and thus have differing strengths and weaknesses. We build the imputation models from 1778 vegetation plots from the US Forest Service’s Forest Inventory and Analysis program (annual plots), and a set of spatial data: a 2006 Landsat image mosaic, modeled climate (PRISM), and topographic variables, for a region that stretches from the Northern Oregon border into California, and encompasses the Cascade Mountains. Preliminary results show that both methods often predict species occurrence at somewhat broader geographic ranges, and a broader range of environmental conditions than they do within our original plot sample. Gradient nearest neighbor is more effective at producing range maps which represent the total area covered by a given species. Random forest predictions track species’ environmental limits more closely with the original data. However, some species are overmapped within their ranges, giving inflated estimates of total area within the mapped region. This tendency also yields different image textures. GNN produces a grainier map with higher edge densities while RFNN produces more compact patches. We conclude that model choice for mapping applications should be based upon intended use of the final map product, because each method has both strengths and weaknesses.
https://digitalcommons.usu.edu/nafecology/sessions/posters/14