Information about the spatial distribution of vegetation is one of the most fundamental needs in land management planning. RCR personnel are actively engaged in all aspects of vegetation mapping, from developing innovative ways to analyze geospatial data to managing large vegetation mapping projects. In addition, RCR projects have varied widely in terms of the scale of mapping ranging from nationwide maps of forest types to local maps of invasive plants along small riparian corridors. The following projects are a sampling of the recent vegetation resource mapping performed by RCR.
Humboldt-Toiyabe National Forest, NV Vegetation Map
RCR and RSAC personnel used satellite imagery and other data in an innovative effort to map vegetation type canopy cover, and size class for six districts on over 6 million acres of the Humboldt-Toiyabe National Forest to assist in their Forest Plan Revision. The process used image segments, derived from eCognition™ software, to delineate stands of vegetation. The segments were then labeled using regression tree software (See5™). The See5™ regression tree process used 55 continuous spatial layers along with nearly 2000 field and photo sample sites to predict classification values for un-sampled areas. This project cost the Forest approximately $0.11 per acre and was completed in a 13 month time frame.
Bridger-Teton National Forest, WY Vegetation Map
RCR and RSAC personnel used satellite imagery and other data to map vegetation type canopy cover, and size class across nearly 3.5 million acres. In a manner similar to the Humboldt-Toiyabe project, the process used image segmentation, 115 different geospatial layers, and an improved See5™ regression tree process. This project cost the Forest approximately $0.16 per acre and was completed in a 21 month time frame.
Comparison of the GNN imputation of forest inventory variables
using both 30-m and 250-m resolution imagery
Traditional mapping methods create separate models(maps) for each land cover variable of interest, such as life form, dominance type, cover and height class. In contrast, imputation methods assign to each pixel in the map a tree list that has actually been measured in the field and therefore retains the covariance structure that naturally exists on the landscape. Imputation is attractive because it produces a data surface that allows for the investigation of vegetation pattern and process relationships through time and space using simulation modeling. This project is collaborative with USFS Pacific Northwest Research Station scientists and compares the results from Gradient Nearest Neighbor (GNN) imputation where the source data comes from mid-resolution Landsat TM (30-m) and moderate resolution MODIS imagery (250-m).
Forest Land Cover / Land Use Modeling using Dasymetric Modeling Techniques
MODIS image-based maps of forest biomass and forest type/group rely on an underlying mask of forest/non-forest. Because MODIS pixels of 250-m or 500-m spatial resolution usually contain mixtures of land cover/use, a binary thematic forest/non-forest mask of this resolution may produce anomalous effects on other geospatial datasets that are masked to exclude non-forest land. Using existing free remote sensing and GIS layers (e.g., protected areas database and census data), this project is creating a continuous forest land use / forest land cover product at 250-m resolution for the continuous U.S. using dasymetric modeling techniques to subdivide the 250-m pixels. As a result of this modeling, each pixel will be assigned continuous values from 0 to 100 percent for each of several attributes, rather than being assigned a binary presence / absence value.
Umatilla National Forest, OR. Forest Mapping Using Most Similar
Neighbor (MSN) Modeling Procedures
Landsat ETM satellite imagery and digital elevation models (DEMs) were used as inputs to statistical evaluation methods for extrapolating detailed stand data to areas for which detailed data are lacking. The process, called Most Similar Neighbor, was tested for several watersheds on the Umatilla National Forest to demonstrate how nationally available data (ETM satellite imagery, DEMs, and stand data) might be used to map detailed forest characteristics. Methods for developing wall-to-wall vegetation characteristics using commonly available data sets such these, have been sought after for many years. Comparisons with alternate methods, such K-Nearest Neighbor (KNN) were tested by RedCastle Resources personnel. This project is being done to support national forest planning efforts.
Tongass National Forest (Forest Structure Mapping with Texture)
Maps of forest structure provide important insights for wildlife modeling, forest ecology, and timber management. Obtaining forest structure information from traditional remotely sensed data has been problematic. RedCastle Resources staff completed a pilot project on the Tongass National Forest to determine how to best map forest structure using commonly available imagery. The project demonstrated how digital orthophotos (DOQs) and Landsat TM satellite imagery could be used together to increase the accuracy of forest structure estimates. The project compared three combinations of texture – derived from DOQs – and Landsat TM satellite imagery. Results showed that including textural information in an image classification can improve the mapping accuracy of forest size class/structure by almost ten percent. Two reports and customized software for merging DOQs and Landsat TM were delivered to the Forest Service. Other deliverables included attributed maps of tree crown closure, species, and size/structure.
