Satellite Image Understanding Through Synthetic Images

Abstract

It is now widely recognized in the Scene Analysis community that further significant advances in computer based methods for image understanding will come about only through a solid comprehension of the physics of the image forming process and the nature of the imaged scene. To this end mathemtical models of the vision process are usually given, and various techniques are used to forecast properties of the imaged scene in terms of the model. The synthetic image, developed by Horn as a registration tool, is here proposed in such a contest for the domain of satellite imaging. Horn’s synthetic image visualizes the abstract world in which the only varying scene characteristic is terrain orientation. It has the advantage that it can be visually compared with the real image. We expand on Woodham’s idea that synthetic images be used to get insights into the imaged scene. The match between synthetic and real image is quantified, and regression analysis techniques are used to get estimates of other scene characteristics such as surface albedo and atmospheric haze. These values, in turn, are used to obtain more realistic synthetic images when the corresponding domain restriction are lifted, in a process ideally leading to the real image. This method is here motivated and explained.