SCREENING DESIGN FOR MODEL SENSITIVITY STUDIES

This paper describes a different approach to sensitivity studies for e nvironmental, including atmospheric, physics models. The sensitivity s tudies were performed on a multispectral environmental data and scene generation capability The capability includes environmental physics mo dels that are used to generate data and scenes for simulation of envir onmental materials, features, and conditions, such as trees, clouds, s oils, and snow. These studies were performed because it is difficult t o obtain input data for many of the environmental variables. The probl em to solve is to determine which of the 100 or so input variables, us ed by the generation capability, are the most important. These sensiti vity studies focused on the generation capabilities needed to predict and evaluate the performance of sensor systems operating in the infrar ed portions of the electromagnetic spectrum. The sensitivity study app roach described uses a screening design. Screening designs are analyti cal techniques that use a fraction of all of the combinations of the p otential input variables and conditions to determine which are the mos t important. Specifically a 20-run Plackett-Burman screening design wa s used to study the sensitivity of eight data and scene generation cap ability computed response variables to 11 selected input variables. Th is is a two-level design, meaning that the range of conditions is repr esented by two different values for each of the 11 selected variables. The eight response variables used were maximum, minimum, range, and m ode of the model-generated temperature and radiance values. The result is that six of the 11 input variables (soil moisture, solar loading, roughness length, relative humidity, surface albedo, and surface emiss ivity) had a statistically significant effect on the response variable s.