Objective 2: Extend Field Data to Landscapes and Estimate Forest Variables using Satellite Data

Forest managers and decision-makers, whether from industry or government, require current forest inventory data to support operational and strategic, planning and management. Forest product companies use the data for evaluating economic potentials and limitations, while states use FIA data to evaluate impacts of industry and other development plans on the forest resource base and environment. New demands are continually placed on FIA for improving the accuracy, information content and timeliness of inventories. The most requested improvements are to increase the precision of estimates at local levels (e.g., county or township) and to provide "wall-to-wall" mapping of key forest variables. As inventory data are broken down into smaller and smaller units, sampling variability increases considerably as sample sizes decrease.

Normally multi-phase designs are used for forest inventories. In the first phase aerial photography is used to stratify forest areas into cover types. In the second phase, field data collection is used to characterize the specific conditions (e.g., volume per acre) within a cover type. The FIA collects a large amount of detailed information on forest type, growth and volume from thousands of field plots. Although this approach provides precise estimates at the state level, in reality it is a sparse (1/10,000) sample of the total area and is insufficient to produce maps of forest types, growth and volume. Conversely, satellite imagery covers the entire landscape at resolutions of 10-30 meters, or 1/4-acre for Landsat TM data. Our objective is to extend information from a sparse sampling of plots (or other known areas) to the landscape with satellite-acquired imagery such as Landsat ETM+.

Although satellite imagery are often criticized for their low spatial resolution, the real limitation is its inability to provide the detailed forest information required by foresters. Conventional methods of satellite image classification have seldom proven accurate enough for forest inventory and management. It might also be noted that conventional classifications produce thematic maps of discrete classes, while foresters are most interested in having information on continuous variables such as volume, size, age, or by species composition/cover type.

An alternative approach is to extrapolate attribute data collected in sparse samples of field plots to township, county and state levels. Such an approach, referred to as k-Nearest Neighbor (k-NN) was first proposed and developed by Tomppo (1991). The k-NN methodology using Landsat TM data has proven very effective in Finland for extending the nationwide forest inventory to provide local estimates and "wall-to-wall" mapping of a range of forest characteristics. In this part of the project we will apply, further develop and test the method in Minnesota, an area with more complex and diverse forests than found in Scandinavian countries where it has first been applied.

Link to Objective 2 Results