MFM-5-Scale: A Physically-Based Inversion Modeling Approach for Unsupervised Cluster Labeling and Independent Landcover Classification and Description

摘要

The 5-Scale geometric-optical reflectance model has been coupled with a Multiple-Forward-Mode (MFM) approach to provide a new inversion modeling capability for both landcover classification and forest biophysical-structural parameter estimation. 5-Scale is the merging of the 4-Scale model and LIBERTY, a general-purpose radiative transfer model. Unlike regular forward mode, MFM-5-Scale does not require exact physical stand descriptors, but instead requires only input ranges and model increments which are more easily obtained. MFM software controls multiple runs of the 5-Scale model in forward mode for all forest structural input combinations at specified solar zenith and view angles and forest species endmember spectral values, with output comprised of modeled pixel reflectance and a set of physical descriptors stored in a forest look-up table (M5-LUT). Classification is achieved by matching satellite image and modeled reflectance values in the M5-LUT. All information associated with the M5-LUT entry is then used to describe that pixel (I.e. Land cover/species; sub-pixel scale fractions of sunlit canopy, background and shadow; stand density; tree height; crown radius; etc.). The LUT can optionally be stratified (generalised) according to species and/or structural class definitions (with multiple and/or hierarchical classes possible), with physically constrained, bi-level reflectance thresholds used to control allocation of pixels into mixed-forest and other foreston-forest classes. In this study, the MFM-5-Scale approach was used in two BOREAS applications to: (ⅰ) label unsupervised clusters generated through Classification by Progressive Generalization (CPG), and (ⅱ) provide an independent classification product with associated physical descriptors. Classification accuracy was evaluated against the SERM forest inventory map for the BOREAS SSA-MSA, with MFM-5-Scale accuracies ranging from 87% for conifer/deciduous classes to 71% for 12 species/density classes. We conclude that this MFM-5-Scale approach provides an inversion modeling context for sophisticated forest radiative transfer models to facilitate a higher level of information retrieval, classification and physical description, with detailed LUTs providing a rich set of forest information suitable for query, analysis, and follow-on simulation studies.