Physically based classification and satellite mapping of biophysical characteristics in the southern boreal forest

Fundamental problems inherent to the existing land cover and biophysical characteristic algorithms are fourfold: (1) their failure to deal physically with global and interannual variations in surface reflectance arising from Sun and view angle variations, (2) decoupling of the land cover classification algorithm from the biophysical characteristic inference algorithm with no ability to control biophysical parameter estimation error arising from misclassification, (3) invalid statistical assumptions within classification algorithms used to model reflectance distribution functions, and (4) sole reliance on vegetation indices that can limit performance for several major land cover classes. To address these problems, we develop an integrated, physically based classification and biophysical characteristics estimation algorithm that utilizes canopy reflectance models to account directly for signature variations from Sun angle, topographic, and other variations. Our approach fuses into a single algorithm both land cover classification and biophysical characteristics estimation, permitting one set of physically based canopy reflectance models to be used for both. The use of canopy reflectance models eliminates the need for unrealistic assumptions, such as multivariate-normal distributions, underlying many classification algorithms. Using the algorithm, we have classified a 10,000 km2 area of the BOREAS southern study area. Our classification shows that low-productivity wetland conifer is the dominant land cover and that nearly 7% of the area is occupied by boreal fens, a major source of methane. In addition, nearly 23% of the area has been disturbed by either fire or logging in the last 20 years, suggesting an important role for disturbance to the regional carbon budget. A thorough evaluation of the physically based classifier within the southern study area shows accuracies superior to those obtained with conventional statistically based algorithms, implying even better performance when extended over multiple Landsat frames since the physically based approach can account directly for regional variations in reflectance resulting from varying illumination and viewing conditions (topography, Sun angle). The conifer biomass density estimation algorithm is based on our discovery of a convenient natural relationship between crown height and volumetric density that renders the biomass density for black spruce stands independent of tree height, and a function only of sunlit canopy fraction. This permits us to calculate directly the relationship between reflectance and biomass density. An evaluation of the algorithm using ground sites shows our algorithm can estimate black spruce biomass density with a root-mean-square error of 2.73 kg/ym2 for correctly classified sites. Our evaluation also demonstrates the importance of correct classification. Rootmean-square errors for misclassified sites were 3.96 kg/m2. Using this approach we have estimated the biomass density in the BOREAS southern study area for the dominant land cover type in the circumpolar boreal ecosystem, wetland black spruce. These results show a bimodality to the biomass density regional distribution, controlled perhaps by underlying topographic and edaphic factors.