Vegetation quantification in a sub-arctic salt marsh using reflectance data

Applications of remote sensing often require the derivation of a relationship between a desired variable, such as vegetation amount, and surface reflectance. Most frequently, linear regression analysis is used. For natural vegetation, relatively little attention has been paid to assessment of the statistical assumptions inherent in regression analysis, although a considerable effort has been made in developing radiometric indices. Using as an example a dataset collected at a long-term goose study site, a relationship between reflectance and aboveground biomass is established for a coastal salt marsh, a preferred foraging area for the geese. At this site, ground cover varies from a dense, short (less than 4 cm) cover of grasses and sedges to bare ground. Mosses and weedy species also occur. Reflectance data were collected for various cover types in the marsh in wavelength bands similar to the first five Landsat bands, at 0.5 m resolution using portable radiometric instruments. At each site, a subplot was sampled for biomass, leaf area index, and cover. Logistic tree regression was used to separate plots with zero vegetation. Linear regression, including data transformations and polynomial regression, was then explored. The vegetation indices tested produced broadly similar results. Band-wise regression generated equations which explained a larger proportion of variance than any index used alone, but vegetation indices with a quadratic term also provided a good fit. Model selection was based on minimizing the Akaike information criterion.