In forestry applications, the development of geometric optical models of forest canopy reflectance has provided a key link between the physical and structural characteristics of forest stands and their remotely sensed spectral response. Accurate land classification is often required for local to regional scale studies which use these models. In this paper, a powerful evidential reasoning image classification algorithm has been modified and linked with geometric optical models of forest canopy reflectance within the context of forest landcover classification and spectral mixture analysis of biophysical variables such as LAI and biomass. Two issues in modifying the evidential reasoning algorithm for achieving reflectance model integration are presented and evaluated: (i) the variability of canopy reflectance model inputs to the classifier, and (ii) the method of converting these model inputs into suitable forms of evidence for use in the evidential reasoning classifier. Three approaches to model integration were tested in a BOREAS forestry application. Of these, the frequency-based method of evidential transformation applied to modelled spectral trajectories produced an overall landcover classification accuracy of 84%, a 30% increase over the other methods tested. This integrated reflectance model - evidential reasoning algorithm provides significant improvements to a physically based approach which has unified land classification and biophysical analysis. In addition to landcover output, use of a reflectance model provides sub-pixel scale mixture fractions for improved biophysical estimates compared to traditional vegetation index approaches. The power and flexibility of the approach provides a suitable framework for improved information extraction from multisource data sets (e.g. remote sensing imagery, elevation models, GIS data etc.) at scales ranging from detailed plot studies to regional and global scale multi-temporal analyses.
展开▼
