Simulation of bidirectional reflectance, modulation transfer, and spatial interaction for the probabilistic classification of Northwest forest structures using Landsat data

Thesis (Ph. D.)--University of Washington, 1998Satellite remote sensors sample the upwelling reflected radiance in the form of digital images. A hierarchical model linking several sub-models of the image acquisition process and the spatial interaction of the classes has been developed to infer forest structure classes from these samples. Forest classifications exhibit large-scale spatial patterns. A fifth-order Markov random field (MRF) is used for modelling spatial interaction consistent with forest classifications at both the local and global levels. Forests scatter irradiance anisotropically. An unobserved image of the reflected radiance field is determined by a bidirectional reflectance factor and a normalized mean reflectance value for each class. The effective system resolution is partially dependent upon the atmospheric characteristics at the time of acquisition. A point spread function models the observed image data. The joint posterior distribution for the hierarchical model is constructed using Bayes theorem to link each of these sub-models. Exact optimization and direct simulation of the resulting MRF are infeasible due to its high dimensionality. Estimates of image attributes are obtained via Markov chain Monte Carlo simulation. The marginal posterior modes estimate (MPM) minimizes the expected number of misclassifications and the posterior probability estimates provide spatially explicit information about the certainty of the MPM estimate. Parameter estimation is a formidable task. A radiosity model is used to simulate the bidirectional reflectance of each forest structure over a wide range of surface orientations. Wavelength dependent separable system point spread functions are estimated for each image classified. The class means effectively supervise the classification and several approaches for their estimation are evaluated. Prior distributions are specified for each of the covariance parameters. Multiple images are classified to assess the consistency of the solutions under varying illumination geometries. The estimated solutions are primarily sensitive to the selection of training data and the MRF parameters. The improvements achieved by detailed modelling of bidirectional reflectance remain subject to sources of variation not accounted for by this approach.The original aspects of this dissertation include the development of a hierarchical framework for classifying remotely sensed data that overcomes the assumption of class conditional independence. Classifications are based on the effective resolution of the image data. Simulating multiple slope and class dependent bidirectional reflectance distribution functions for the purpose of normalizing projected area variation in rugged terrain and between multiple images is an important and original feature of this research. The application of a fifth-order random field to forest remote sensing is also unique