Contextual simulation of landscape based on remotely sensed data

An observed image can be considered as a single sample of a stochastic process under an assumed model. It is often desirable to generate a multitude of scenes which have the same stochastic properties as the original scene as a means of evaluating and validating proposed models. A class of stochastic models has been developed to characterize landscape processes represented in multispectral imagery and then to simulate these processes. In particular, models are derived from remotely sensed imagery and utilized to develop initial conditions for temporal simulations of vegetation in an ecology study. Landscape observed in remotely sensed imagery often exhibits characteristic patch mosaic structures at the large scale and class dependent variability within each region at the detailed scale. A Markov random field (MRF) model is employed to model the region process as a large scale characteristic and generate a spatially aggregated class map. Boundary variation between adjacent regions is represented using a fuzzy approach implemented within a multiresolution data structure. Class dependent variability and noise are superimposed on the resultant regions