Linking Hyperspectral Imagery and Forest Inventories for Forest Assessment in the Central Appalachians

—Hyperspectral imagery from EO-1 Hyperion and AVIRIS were used in conjunction with continuous forest inventory (CFI) data to map detailed forest composition in the state forests of Western Maryland. We developed a hierarchical vegetation classification that conformed to the National Vegetation Classification Standard (NVCS) at the Alliance level and mapped these forest types as a function of hyperspectral reflectance using decision trees. Overall classification accuracy for vegetation at the Alliance level was very high (60-80 percent), with field validation indicating accuracies ranging from 65-70 percent. In an area dominated by oaks, the hyperspectral imagery was able to accurately distinguish plots dominated by individual red oaks with acceptable success (>60 percent). Hyperspectral imagery also differentiated between conifers more than 70 percent of the time. Overall, the accuracies were improved over similar analyses conducted using multi-date Landsat data. Our research demonstrates the capacity for hyperspectral imagery to remotely monitor, map and model forest systems in the Central Appalachians. The resulting forest composition maps can inform forest management decisions with a level of information content not previously available. Mapping forest types is one of the primary applications of remote sensing data to forest management. In the past, standard satellite imagery available from Landsat era sensors has been limited in its ability to differentiate between specific hardwood forest types such as those described at the Alliance level of the National Vegetation Classification Standard (NVCS) (http://biology.usgs.gov/npsveg/nvcs.html). Forest maps with species level detail continue to be desired by forest managers and scientists alike, yet remain difficult to produce from satellite imagery without complex methods, extra data sources, and intense time and effort. Hyperspectral imagery improves upon the limits of Landsat data by measuring the reflectance of light in more than two hundred narrow bands, compared to Landsat TM’s seven wide bands. Specific portions of the hyperspectral spectrum have been linked to forest indicators such as forest stress measured by canopy chlorophyll content (Sampson et al. 2003) and the biophysical content of leaves (Townsend et al. 2003, Smith et al. 2003). As such, the high data content found in hyperspectral data promises to greatly enhance forest mapping capabilities when used in conjunction with typical forest inventory data. We tested the ability of hyperspectral imagery from two sensors combined with Continuous Forest Inventory (CFI) data to map forest composition to the Alliance level and compared the results to a similar classification created from multi-season Landsat TM data.