An in-scene parameter estimation method for quantitative image analysis

In-scene parameter estimation applies image samples of a known population to find parameters in its image generation model. These parameter estimates then may be applied in models of other populations of interest. In remote-sensed imaging, the models are often complex (nonlinear, numerical, etc.), and the sample sizes are usually large. In many cases, traditional estimators are not applicable. The author presents a numerical approach to estimation that can be applied to such models. This estimator is an approximation to the maximum likelihood solution of a general non-linear model. It is based on matching simplified residuals with their known density, which is-obtained by simulation. The goodness of the match is characterized by an objective function derived from the likelihood ratio statistic. A sampling distribution, generated by replication, characterizes the estimator for a given application. The technique has been applied to estimate parameters in thermal infrared imagery, but it may be employed to estimate parameters in various single band and multi-spectral applications.