A Bayesian Unmixing Algorithm for Retrieving Landcover Distributions Using Global Reflectance Data

Knowledge of land cover type and vegetation condition at continental-to-global scales is critical for quantifying the interactions between the biosphere and atmosphere, and for making predictions of climate and the global carbon cycle. For example, the response of different landcover types to their environment in many ecosystem models is typically linked to a set of parameters that are keyed to ecosystem type. Also, models which include changes in age class and structure must be initialized or constrained with information on land cover. Satellite remote sensing can provide a rich description of the land surface, and has been used extensively for mapping land cover types. Typically, these vegetation mapping methods rely upon the information contained in the reflectance of solar radiation in many wavelengths, or upon the temporal patterns of indices which combine these reflectances. Most land cover products provide an estimate of the dominant land cover class within a grid cell of a fixed size, and for a fixed number of classes. We have developed a method based on linear mixture modeling which combines low resolution (e.g., MODIS) and high resolution (e.g., ASTER) satellite data to estimate the fractional distribution of classes within grid cells of various sizes, and for classes that are regionally defined. By applying Bayes\u27 therorem, we have extended the traditional linear unmixing approach to include prior knowledge of vegetation type and condition, and to provide estimates of uncertainty in the retreived sub-pixel land cover distributions